<!--
doc/src/sgml/ref/select.sgml
PostgreSQL documentation
-->
<refentry id="SQL-SELECT">
<refmeta>
<refentrytitle>SELECT</refentrytitle>
<manvolnum>7</manvolnum>
<refmiscinfo>SQL - Language Statements</refmiscinfo>
</refmeta>
<refnamediv>
<refname>SELECT</refname>
<refname>TABLE</refname>
<refname>WITH</refname>
<refpurpose>retrieve rows from a table or view</refpurpose>
</refnamediv>
<indexterm zone="sql-select">
<primary>SELECT</primary>
</indexterm>
<indexterm zone="sql-select">
<primary>TABLE command</primary>
</indexterm>
<indexterm zone="sql-select">
<primary>WITH</primary>
<secondary>in SELECT</secondary>
</indexterm>
<refsynopsisdiv>
<synopsis>
[ WITH [ RECURSIVE ] <replaceable class="parameter">with_query</replaceable> [, ...] ]
SELECT [ ALL | DISTINCT [ ON ( <replaceable class="parameter">expression</replaceable> [, ...] ) ] ]
[ * | <replaceable class="parameter">expression</replaceable> [ [ AS ] <replaceable class="parameter">output_name</replaceable> ] [, ...] ]
[ FROM <replaceable class="parameter">from_item</replaceable> [, ...] ]
[ WHERE <replaceable class="parameter">condition</replaceable> ]
[ GROUP BY <replaceable class="parameter">expression</replaceable> [, ...] ]
[ HAVING <replaceable class="parameter">condition</replaceable> [, ...] ]
[ WINDOW <replaceable class="parameter">window_name</replaceable> AS ( <replaceable class="parameter">window_definition</replaceable> ) [, ...] ]
[ { UNION | INTERSECT | EXCEPT } [ ALL | DISTINCT ] <replaceable class="parameter">select</replaceable> ]
[ ORDER BY <replaceable class="parameter">expression</replaceable> [ ASC | DESC | USING <replaceable class="parameter">operator</replaceable> ] [ NULLS { FIRST | LAST } ] [, ...] ]
[ LIMIT { <replaceable class="parameter">count</replaceable> | ALL } ]
[ OFFSET <replaceable class="parameter">start</replaceable> [ ROW | ROWS ] ]
[ FETCH { FIRST | NEXT } [ <replaceable class="parameter">count</replaceable> ] { ROW | ROWS } ONLY ]
Improve concurrency of foreign key locking
This patch introduces two additional lock modes for tuples: "SELECT FOR
KEY SHARE" and "SELECT FOR NO KEY UPDATE". These don't block each
other, in contrast with already existing "SELECT FOR SHARE" and "SELECT
FOR UPDATE". UPDATE commands that do not modify the values stored in
the columns that are part of the key of the tuple now grab a SELECT FOR
NO KEY UPDATE lock on the tuple, allowing them to proceed concurrently
with tuple locks of the FOR KEY SHARE variety.
Foreign key triggers now use FOR KEY SHARE instead of FOR SHARE; this
means the concurrency improvement applies to them, which is the whole
point of this patch.
The added tuple lock semantics require some rejiggering of the multixact
module, so that the locking level that each transaction is holding can
be stored alongside its Xid. Also, multixacts now need to persist
across server restarts and crashes, because they can now represent not
only tuple locks, but also tuple updates. This means we need more
careful tracking of lifetime of pg_multixact SLRU files; since they now
persist longer, we require more infrastructure to figure out when they
can be removed. pg_upgrade also needs to be careful to copy
pg_multixact files over from the old server to the new, or at least part
of multixact.c state, depending on the versions of the old and new
servers.
Tuple time qualification rules (HeapTupleSatisfies routines) need to be
careful not to consider tuples with the "is multi" infomask bit set as
being only locked; they might need to look up MultiXact values (i.e.
possibly do pg_multixact I/O) to find out the Xid that updated a tuple,
whereas they previously were assured to only use information readily
available from the tuple header. This is considered acceptable, because
the extra I/O would involve cases that would previously cause some
commands to block waiting for concurrent transactions to finish.
Another important change is the fact that locking tuples that have
previously been updated causes the future versions to be marked as
locked, too; this is essential for correctness of foreign key checks.
This causes additional WAL-logging, also (there was previously a single
WAL record for a locked tuple; now there are as many as updated copies
of the tuple there exist.)
With all this in place, contention related to tuples being checked by
foreign key rules should be much reduced.
As a bonus, the old behavior that a subtransaction grabbing a stronger
tuple lock than the parent (sub)transaction held on a given tuple and
later aborting caused the weaker lock to be lost, has been fixed.
Many new spec files were added for isolation tester framework, to ensure
overall behavior is sane. There's probably room for several more tests.
There were several reviewers of this patch; in particular, Noah Misch
and Andres Freund spent considerable time in it. Original idea for the
patch came from Simon Riggs, after a problem report by Joel Jacobson.
Most code is from me, with contributions from Marti Raudsepp, Alexander
Shulgin, Noah Misch and Andres Freund.
This patch was discussed in several pgsql-hackers threads; the most
important start at the following message-ids:
AANLkTimo9XVcEzfiBR-ut3KVNDkjm2Vxh+t8kAmWjPuv@mail.gmail.com
1290721684-sup-3951@alvh.no-ip.org
1294953201-sup-2099@alvh.no-ip.org
1320343602-sup-2290@alvh.no-ip.org
1339690386-sup-8927@alvh.no-ip.org
4FE5FF020200002500048A3D@gw.wicourts.gov
4FEAB90A0200002500048B7D@gw.wicourts.gov
13 years ago
[ FOR { UPDATE | NO KEY UPDATE | SHARE | KEY SHARE } [ OF <replaceable class="parameter">table_name</replaceable> [, ...] ] [ NOWAIT ] [...] ]
<phrase>where <replaceable class="parameter">from_item</replaceable> can be one of:</phrase>
[ ONLY ] <replaceable class="parameter">table_name</replaceable> [ * ] [ [ AS ] <replaceable class="parameter">alias</replaceable> [ ( <replaceable class="parameter">column_alias</replaceable> [, ...] ) ] ]
[ LATERAL ] ( <replaceable class="parameter">select</replaceable> ) [ AS ] <replaceable class="parameter">alias</replaceable> [ ( <replaceable class="parameter">column_alias</replaceable> [, ...] ) ]
<replaceable class="parameter">with_query_name</replaceable> [ [ AS ] <replaceable class="parameter">alias</replaceable> [ ( <replaceable class="parameter">column_alias</replaceable> [, ...] ) ] ]
Support multi-argument UNNEST(), and TABLE() syntax for multiple functions.
This patch adds the ability to write TABLE( function1(), function2(), ...)
as a single FROM-clause entry. The result is the concatenation of the
first row from each function, followed by the second row from each
function, etc; with NULLs inserted if any function produces fewer rows than
others. This is believed to be a much more useful behavior than what
Postgres currently does with multiple SRFs in a SELECT list.
This syntax also provides a reasonable way to combine use of column
definition lists with WITH ORDINALITY: put the column definition list
inside TABLE(), where it's clear that it doesn't control the ordinality
column as well.
Also implement SQL-compliant multiple-argument UNNEST(), by turning
UNNEST(a,b,c) into TABLE(unnest(a), unnest(b), unnest(c)).
The SQL standard specifies TABLE() with only a single function, not
multiple functions, and it seems to require an implicit UNNEST() which is
not what this patch does. There may be something wrong with that reading
of the spec, though, because if it's right then the spec's TABLE() is just
a pointless alternative spelling of UNNEST(). After further review of
that, we might choose to adopt a different syntax for what this patch does,
but in any case this functionality seems clearly worthwhile.
Andrew Gierth, reviewed by Zoltán Böszörményi and Heikki Linnakangas, and
significantly revised by me
12 years ago
[ LATERAL ] <replaceable class="parameter">function_name</replaceable> ( [ <replaceable class="parameter">argument</replaceable> [, ...] ] )
[ WITH ORDINALITY ] [ [ AS ] <replaceable class="parameter">alias</replaceable> [ ( <replaceable class="parameter">column_alias</replaceable> [, ...] ) ] ]
[ LATERAL ] <replaceable class="parameter">function_name</replaceable> ( [ <replaceable class="parameter">argument</replaceable> [, ...] ] ) [ AS ] <replaceable class="parameter">alias</replaceable> ( <replaceable class="parameter">column_definition</replaceable> [, ...] )
[ LATERAL ] <replaceable class="parameter">function_name</replaceable> ( [ <replaceable class="parameter">argument</replaceable> [, ...] ] ) AS ( <replaceable class="parameter">column_definition</replaceable> [, ...] )
[ LATERAL ] ROWS FROM( <replaceable class="parameter">function_name</replaceable> ( [ <replaceable class="parameter">argument</replaceable> [, ...] ] ) [ AS ( <replaceable class="parameter">column_definition</replaceable> [, ...] ) ] [, ...] )
Support multi-argument UNNEST(), and TABLE() syntax for multiple functions.
This patch adds the ability to write TABLE( function1(), function2(), ...)
as a single FROM-clause entry. The result is the concatenation of the
first row from each function, followed by the second row from each
function, etc; with NULLs inserted if any function produces fewer rows than
others. This is believed to be a much more useful behavior than what
Postgres currently does with multiple SRFs in a SELECT list.
This syntax also provides a reasonable way to combine use of column
definition lists with WITH ORDINALITY: put the column definition list
inside TABLE(), where it's clear that it doesn't control the ordinality
column as well.
Also implement SQL-compliant multiple-argument UNNEST(), by turning
UNNEST(a,b,c) into TABLE(unnest(a), unnest(b), unnest(c)).
The SQL standard specifies TABLE() with only a single function, not
multiple functions, and it seems to require an implicit UNNEST() which is
not what this patch does. There may be something wrong with that reading
of the spec, though, because if it's right then the spec's TABLE() is just
a pointless alternative spelling of UNNEST(). After further review of
that, we might choose to adopt a different syntax for what this patch does,
but in any case this functionality seems clearly worthwhile.
Andrew Gierth, reviewed by Zoltán Böszörményi and Heikki Linnakangas, and
significantly revised by me
12 years ago
[ WITH ORDINALITY ] [ [ AS ] <replaceable class="parameter">alias</replaceable> [ ( <replaceable class="parameter">column_alias</replaceable> [, ...] ) ] ]
<replaceable class="parameter">from_item</replaceable> [ NATURAL ] <replaceable class="parameter">join_type</replaceable> <replaceable class="parameter">from_item</replaceable> [ ON <replaceable class="parameter">join_condition</replaceable> | USING ( <replaceable class="parameter">join_column</replaceable> [, ...] ) ]
<phrase>and <replaceable class="parameter">with_query</replaceable> is:</phrase>
<replaceable class="parameter">with_query_name</replaceable> [ ( <replaceable class="parameter">column_name</replaceable> [, ...] ) ] AS ( <replaceable class="parameter">select</replaceable> | <replaceable class="parameter">values</replaceable> | <replaceable class="parameter">insert</replaceable> | <replaceable class="parameter">update</replaceable> | <replaceable class="parameter">delete</replaceable> )
TABLE [ ONLY ] <replaceable class="parameter">table_name</replaceable> [ * ]
</synopsis>
</refsynopsisdiv>
<refsect1>
<title>Description</title>
<para>
<command>SELECT</command> retrieves rows from zero or more tables.
The general processing of <command>SELECT</command> is as follows:
<orderedlist>
<listitem>
<para>
All queries in the <literal>WITH</literal> list are computed.
These effectively serve as temporary tables that can be referenced
in the <literal>FROM</literal> list. A <literal>WITH</literal> query
that is referenced more than once in <literal>FROM</literal> is
computed only once.
(See <xref linkend="sql-with" endterm="sql-with-title"> below.)
</para>
</listitem>
<listitem>
<para>
All elements in the <literal>FROM</literal> list are computed.
(Each element in the <literal>FROM</literal> list is a real or
virtual table.) If more than one element is specified in the
<literal>FROM</literal> list, they are cross-joined together.
(See <xref linkend="sql-from" endterm="sql-from-title"> below.)
</para>
</listitem>
<listitem>
<para>
If the <literal>WHERE</literal> clause is specified, all rows
that do not satisfy the condition are eliminated from the
output. (See <xref linkend="sql-where"
endterm="sql-where-title"> below.)
</para>
</listitem>
<listitem>
<para>
If the <literal>GROUP BY</literal> clause is specified, the
output is combined into groups of rows that match on one or more
values. If the <literal>HAVING</literal> clause is present, it
eliminates groups that do not satisfy the given condition. (See
<xref linkend="sql-groupby" endterm="sql-groupby-title"> and
<xref linkend="sql-having" endterm="sql-having-title"> below.)
</para>
</listitem>
<listitem>
<para>
The actual output rows are computed using the
<command>SELECT</command> output expressions for each selected
row or row group. (See
<xref linkend="sql-select-list" endterm="sql-select-list-title">
below.)
</para>
</listitem>
<listitem>
<para><literal>SELECT DISTINCT</literal> eliminates duplicate rows from the
result. <literal>SELECT DISTINCT ON</literal> eliminates rows that
match on all the specified expressions. <literal>SELECT ALL</literal>
(the default) will return all candidate rows, including
duplicates. (See <xref linkend="sql-distinct"
endterm="sql-distinct-title"> below.)
</para>
</listitem>
<listitem>
<para>
Using the operators <literal>UNION</literal>,
<literal>INTERSECT</literal>, and <literal>EXCEPT</literal>, the
output of more than one <command>SELECT</command> statement can
be combined to form a single result set. The
<literal>UNION</literal> operator returns all rows that are in
one or both of the result sets. The
<literal>INTERSECT</literal> operator returns all rows that are
strictly in both result sets. The <literal>EXCEPT</literal>
operator returns the rows that are in the first result set but
not in the second. In all three cases, duplicate rows are
eliminated unless <literal>ALL</literal> is specified. The noise
word <literal>DISTINCT</> can be added to explicitly specify
eliminating duplicate rows. Notice that <literal>DISTINCT</> is
the default behavior here, even though <literal>ALL</literal> is
the default for <command>SELECT</> itself. (See
<xref linkend="sql-union" endterm="sql-union-title">, <xref
linkend="sql-intersect" endterm="sql-intersect-title">, and
<xref linkend="sql-except" endterm="sql-except-title"> below.)
</para>
</listitem>
<listitem>
<para>
If the <literal>ORDER BY</literal> clause is specified, the
returned rows are sorted in the specified order. If
<literal>ORDER BY</literal> is not given, the rows are returned
in whatever order the system finds fastest to produce. (See
<xref linkend="sql-orderby" endterm="sql-orderby-title"> below.)
</para>
</listitem>
<listitem>
<para>
If the <literal>LIMIT</literal> (or <literal>FETCH FIRST</literal>) or <literal>OFFSET</literal>
clause is specified, the <command>SELECT</command> statement
only returns a subset of the result rows. (See <xref
linkend="sql-limit" endterm="sql-limit-title"> below.)
</para>
</listitem>
<listitem>
<para>
Improve concurrency of foreign key locking
This patch introduces two additional lock modes for tuples: "SELECT FOR
KEY SHARE" and "SELECT FOR NO KEY UPDATE". These don't block each
other, in contrast with already existing "SELECT FOR SHARE" and "SELECT
FOR UPDATE". UPDATE commands that do not modify the values stored in
the columns that are part of the key of the tuple now grab a SELECT FOR
NO KEY UPDATE lock on the tuple, allowing them to proceed concurrently
with tuple locks of the FOR KEY SHARE variety.
Foreign key triggers now use FOR KEY SHARE instead of FOR SHARE; this
means the concurrency improvement applies to them, which is the whole
point of this patch.
The added tuple lock semantics require some rejiggering of the multixact
module, so that the locking level that each transaction is holding can
be stored alongside its Xid. Also, multixacts now need to persist
across server restarts and crashes, because they can now represent not
only tuple locks, but also tuple updates. This means we need more
careful tracking of lifetime of pg_multixact SLRU files; since they now
persist longer, we require more infrastructure to figure out when they
can be removed. pg_upgrade also needs to be careful to copy
pg_multixact files over from the old server to the new, or at least part
of multixact.c state, depending on the versions of the old and new
servers.
Tuple time qualification rules (HeapTupleSatisfies routines) need to be
careful not to consider tuples with the "is multi" infomask bit set as
being only locked; they might need to look up MultiXact values (i.e.
possibly do pg_multixact I/O) to find out the Xid that updated a tuple,
whereas they previously were assured to only use information readily
available from the tuple header. This is considered acceptable, because
the extra I/O would involve cases that would previously cause some
commands to block waiting for concurrent transactions to finish.
Another important change is the fact that locking tuples that have
previously been updated causes the future versions to be marked as
locked, too; this is essential for correctness of foreign key checks.
This causes additional WAL-logging, also (there was previously a single
WAL record for a locked tuple; now there are as many as updated copies
of the tuple there exist.)
With all this in place, contention related to tuples being checked by
foreign key rules should be much reduced.
As a bonus, the old behavior that a subtransaction grabbing a stronger
tuple lock than the parent (sub)transaction held on a given tuple and
later aborting caused the weaker lock to be lost, has been fixed.
Many new spec files were added for isolation tester framework, to ensure
overall behavior is sane. There's probably room for several more tests.
There were several reviewers of this patch; in particular, Noah Misch
and Andres Freund spent considerable time in it. Original idea for the
patch came from Simon Riggs, after a problem report by Joel Jacobson.
Most code is from me, with contributions from Marti Raudsepp, Alexander
Shulgin, Noah Misch and Andres Freund.
This patch was discussed in several pgsql-hackers threads; the most
important start at the following message-ids:
AANLkTimo9XVcEzfiBR-ut3KVNDkjm2Vxh+t8kAmWjPuv@mail.gmail.com
1290721684-sup-3951@alvh.no-ip.org
1294953201-sup-2099@alvh.no-ip.org
1320343602-sup-2290@alvh.no-ip.org
1339690386-sup-8927@alvh.no-ip.org
4FE5FF020200002500048A3D@gw.wicourts.gov
4FEAB90A0200002500048B7D@gw.wicourts.gov
13 years ago
If <literal>FOR UPDATE</>, <literal>FOR NO KEY UPDATE</literal>, <literal>FOR SHARE</literal>
or <literal>FOR KEY SHARE</literal>
is specified, the
<command>SELECT</command> statement locks the selected rows
against concurrent updates. (See <xref linkend="sql-for-update-share"
endterm="sql-for-update-share-title"> below.)
</para>
</listitem>
</orderedlist>
</para>
<para>
You must have <literal>SELECT</literal> privilege on each column used
Improve concurrency of foreign key locking
This patch introduces two additional lock modes for tuples: "SELECT FOR
KEY SHARE" and "SELECT FOR NO KEY UPDATE". These don't block each
other, in contrast with already existing "SELECT FOR SHARE" and "SELECT
FOR UPDATE". UPDATE commands that do not modify the values stored in
the columns that are part of the key of the tuple now grab a SELECT FOR
NO KEY UPDATE lock on the tuple, allowing them to proceed concurrently
with tuple locks of the FOR KEY SHARE variety.
Foreign key triggers now use FOR KEY SHARE instead of FOR SHARE; this
means the concurrency improvement applies to them, which is the whole
point of this patch.
The added tuple lock semantics require some rejiggering of the multixact
module, so that the locking level that each transaction is holding can
be stored alongside its Xid. Also, multixacts now need to persist
across server restarts and crashes, because they can now represent not
only tuple locks, but also tuple updates. This means we need more
careful tracking of lifetime of pg_multixact SLRU files; since they now
persist longer, we require more infrastructure to figure out when they
can be removed. pg_upgrade also needs to be careful to copy
pg_multixact files over from the old server to the new, or at least part
of multixact.c state, depending on the versions of the old and new
servers.
Tuple time qualification rules (HeapTupleSatisfies routines) need to be
careful not to consider tuples with the "is multi" infomask bit set as
being only locked; they might need to look up MultiXact values (i.e.
possibly do pg_multixact I/O) to find out the Xid that updated a tuple,
whereas they previously were assured to only use information readily
available from the tuple header. This is considered acceptable, because
the extra I/O would involve cases that would previously cause some
commands to block waiting for concurrent transactions to finish.
Another important change is the fact that locking tuples that have
previously been updated causes the future versions to be marked as
locked, too; this is essential for correctness of foreign key checks.
This causes additional WAL-logging, also (there was previously a single
WAL record for a locked tuple; now there are as many as updated copies
of the tuple there exist.)
With all this in place, contention related to tuples being checked by
foreign key rules should be much reduced.
As a bonus, the old behavior that a subtransaction grabbing a stronger
tuple lock than the parent (sub)transaction held on a given tuple and
later aborting caused the weaker lock to be lost, has been fixed.
Many new spec files were added for isolation tester framework, to ensure
overall behavior is sane. There's probably room for several more tests.
There were several reviewers of this patch; in particular, Noah Misch
and Andres Freund spent considerable time in it. Original idea for the
patch came from Simon Riggs, after a problem report by Joel Jacobson.
Most code is from me, with contributions from Marti Raudsepp, Alexander
Shulgin, Noah Misch and Andres Freund.
This patch was discussed in several pgsql-hackers threads; the most
important start at the following message-ids:
AANLkTimo9XVcEzfiBR-ut3KVNDkjm2Vxh+t8kAmWjPuv@mail.gmail.com
1290721684-sup-3951@alvh.no-ip.org
1294953201-sup-2099@alvh.no-ip.org
1320343602-sup-2290@alvh.no-ip.org
1339690386-sup-8927@alvh.no-ip.org
4FE5FF020200002500048A3D@gw.wicourts.gov
4FEAB90A0200002500048B7D@gw.wicourts.gov
13 years ago
in a <command>SELECT</> command. The use of <literal>FOR NO KEY UPDATE</>,
<literal>FOR UPDATE</literal>,
<literal>FOR SHARE</literal> or <literal>FOR KEY SHARE</literal> requires
<literal>UPDATE</literal> privilege as well (for at least one column
of each table so selected).
</para>
</refsect1>
<refsect1>
<title>Parameters</title>
<refsect2 id="SQL-WITH">
<title id="sql-with-title"><literal>WITH</literal> Clause</title>
<para>
The <literal>WITH</literal> clause allows you to specify one or more
subqueries that can be referenced by name in the primary query.
The subqueries effectively act as temporary tables or views
for the duration of the primary query.
Each subquery can be a <command>SELECT</command>, <command>VALUES</command>,
<command>INSERT</command>, <command>UPDATE</command> or
<command>DELETE</command> statement.
When writing a data-modifying statement (<command>INSERT</command>,
<command>UPDATE</command> or <command>DELETE</command>) in
<literal>WITH</>, it is usual to include a <literal>RETURNING</> clause.
It is the output of <literal>RETURNING</>, <emphasis>not</> the underlying
table that the statement modifies, that forms the temporary table that is
read by the primary query. If <literal>RETURNING</> is omitted, the
statement is still executed, but it produces no output so it cannot be
referenced as a table by the primary query.
</para>
<para>
A name (without schema qualification) must be specified for each
<literal>WITH</literal> query. Optionally, a list of column names
can be specified; if this is omitted,
the column names are inferred from the subquery.
</para>
<para>
If <literal>RECURSIVE</literal> is specified, it allows a
<command>SELECT</command> subquery to reference itself by name. Such a
subquery must have the form
<synopsis>
<replaceable class="parameter">non_recursive_term</replaceable> UNION [ ALL | DISTINCT ] <replaceable class="parameter">recursive_term</replaceable>
</synopsis>
where the recursive self-reference must appear on the right-hand
side of the <literal>UNION</>. Only one recursive self-reference
is permitted per query. Recursive data-modifying statements are not
supported, but you can use the results of a recursive
<command>SELECT</command> query in
a data-modifying statement. See <xref linkend="queries-with"> for
an example.
</para>
<para>
Another effect of <literal>RECURSIVE</literal> is that
<literal>WITH</literal> queries need not be ordered: a query
can reference another one that is later in the list. (However,
circular references, or mutual recursion, are not implemented.)
Without <literal>RECURSIVE</literal>, <literal>WITH</literal> queries
can only reference sibling <literal>WITH</literal> queries
that are earlier in the <literal>WITH</literal> list.
</para>
<para>
A key property of <literal>WITH</literal> queries is that they
are evaluated only once per execution of the primary query,
even if the primary query refers to them more than once.
In particular, data-modifying statements are guaranteed to be
executed once and only once, regardless of whether the primary query
reads all or any of their output.
</para>
<para>
The primary query and the <literal>WITH</literal> queries are all
(notionally) executed at the same time. This implies that the effects of
a data-modifying statement in <literal>WITH</literal> cannot be seen from
other parts of the query, other than by reading its <literal>RETURNING</>
output. If two such data-modifying statements attempt to modify the same
row, the results are unspecified.
</para>
<para>
See <xref linkend="queries-with"> for additional information.
</para>
</refsect2>
<refsect2 id="SQL-FROM">
<title id="sql-from-title"><literal>FROM</literal> Clause</title>
<para>
The <literal>FROM</literal> clause specifies one or more source
tables for the <command>SELECT</command>. If multiple sources are
specified, the result is the Cartesian product (cross join) of all
the sources. But usually qualification conditions are added (via
<literal>WHERE</>) to restrict the returned rows to a small subset of the
Cartesian product.
</para>
<para>
The <literal>FROM</literal> clause can contain the following
elements:
<variablelist>
<varlistentry>
<term><replaceable class="parameter">table_name</replaceable></term>
<listitem>
<para>
The name (optionally schema-qualified) of an existing table or view.
If <literal>ONLY</> is specified before the table name, only that
table is scanned. If <literal>ONLY</> is not specified, the table
and all its descendant tables (if any) are scanned. Optionally,
<literal>*</> can be specified after the table name to explicitly
indicate that descendant tables are included.
</para>
</listitem>
</varlistentry>
<varlistentry>
<term><replaceable class="parameter">alias</replaceable></term>
<listitem>
<para>
A substitute name for the <literal>FROM</> item containing the
alias. An alias is used for brevity or to eliminate ambiguity
for self-joins (where the same table is scanned multiple
times). When an alias is provided, it completely hides the
actual name of the table or function; for example given
<literal>FROM foo AS f</>, the remainder of the
<command>SELECT</command> must refer to this <literal>FROM</>
item as <literal>f</> not <literal>foo</>. If an alias is
written, a column alias list can also be written to provide
substitute names for one or more columns of the table.
</para>
</listitem>
</varlistentry>
<varlistentry>
<term><replaceable class="parameter">select</replaceable></term>
<listitem>
<para>
A sub-<command>SELECT</command> can appear in the
<literal>FROM</literal> clause. This acts as though its
output were created as a temporary table for the duration of
this single <command>SELECT</command> command. Note that the
sub-<command>SELECT</command> must be surrounded by
parentheses, and an alias <emphasis>must</emphasis> be
provided for it. A
<xref linkend="sql-values"> command
can also be used here.
</para>
</listitem>
</varlistentry>
<varlistentry>
<term><replaceable class="parameter">with_query_name</replaceable></term>
<listitem>
<para>
A <literal>WITH</> query is referenced by writing its name,
just as though the query's name were a table name. (In fact,
the <literal>WITH</> query hides any real table of the same name
for the purposes of the primary query. If necessary, you can
refer to a real table of the same name by schema-qualifying
the table's name.)
An alias can be provided in the same way as for a table.
</para>
</listitem>
</varlistentry>
<varlistentry>
<term><replaceable class="parameter">function_name</replaceable></term>
<listitem>
<para>
Function calls can appear in the <literal>FROM</literal>
clause. (This is especially useful for functions that return
result sets, but any function can be used.) This acts as
Support multi-argument UNNEST(), and TABLE() syntax for multiple functions.
This patch adds the ability to write TABLE( function1(), function2(), ...)
as a single FROM-clause entry. The result is the concatenation of the
first row from each function, followed by the second row from each
function, etc; with NULLs inserted if any function produces fewer rows than
others. This is believed to be a much more useful behavior than what
Postgres currently does with multiple SRFs in a SELECT list.
This syntax also provides a reasonable way to combine use of column
definition lists with WITH ORDINALITY: put the column definition list
inside TABLE(), where it's clear that it doesn't control the ordinality
column as well.
Also implement SQL-compliant multiple-argument UNNEST(), by turning
UNNEST(a,b,c) into TABLE(unnest(a), unnest(b), unnest(c)).
The SQL standard specifies TABLE() with only a single function, not
multiple functions, and it seems to require an implicit UNNEST() which is
not what this patch does. There may be something wrong with that reading
of the spec, though, because if it's right then the spec's TABLE() is just
a pointless alternative spelling of UNNEST(). After further review of
that, we might choose to adopt a different syntax for what this patch does,
but in any case this functionality seems clearly worthwhile.
Andrew Gierth, reviewed by Zoltán Böszörményi and Heikki Linnakangas, and
significantly revised by me
12 years ago
though the function's output were created as a temporary table for the
duration of this single <command>SELECT</command> command.
Support multi-argument UNNEST(), and TABLE() syntax for multiple functions.
This patch adds the ability to write TABLE( function1(), function2(), ...)
as a single FROM-clause entry. The result is the concatenation of the
first row from each function, followed by the second row from each
function, etc; with NULLs inserted if any function produces fewer rows than
others. This is believed to be a much more useful behavior than what
Postgres currently does with multiple SRFs in a SELECT list.
This syntax also provides a reasonable way to combine use of column
definition lists with WITH ORDINALITY: put the column definition list
inside TABLE(), where it's clear that it doesn't control the ordinality
column as well.
Also implement SQL-compliant multiple-argument UNNEST(), by turning
UNNEST(a,b,c) into TABLE(unnest(a), unnest(b), unnest(c)).
The SQL standard specifies TABLE() with only a single function, not
multiple functions, and it seems to require an implicit UNNEST() which is
not what this patch does. There may be something wrong with that reading
of the spec, though, because if it's right then the spec's TABLE() is just
a pointless alternative spelling of UNNEST(). After further review of
that, we might choose to adopt a different syntax for what this patch does,
but in any case this functionality seems clearly worthwhile.
Andrew Gierth, reviewed by Zoltán Böszörményi and Heikki Linnakangas, and
significantly revised by me
12 years ago
When the optional <command>WITH ORDINALITY</command> clause is
added to the function call, a new column is appended after
all the function's output columns with numbering for each row.
</para>
<para>
An alias can be provided in the same way as for a table.
If an alias is written, a column
alias list can also be written to provide substitute names for
one or more attributes of the function's composite return
type, including the column added by <literal>ORDINALITY</literal>
if present.
Support multi-argument UNNEST(), and TABLE() syntax for multiple functions.
This patch adds the ability to write TABLE( function1(), function2(), ...)
as a single FROM-clause entry. The result is the concatenation of the
first row from each function, followed by the second row from each
function, etc; with NULLs inserted if any function produces fewer rows than
others. This is believed to be a much more useful behavior than what
Postgres currently does with multiple SRFs in a SELECT list.
This syntax also provides a reasonable way to combine use of column
definition lists with WITH ORDINALITY: put the column definition list
inside TABLE(), where it's clear that it doesn't control the ordinality
column as well.
Also implement SQL-compliant multiple-argument UNNEST(), by turning
UNNEST(a,b,c) into TABLE(unnest(a), unnest(b), unnest(c)).
The SQL standard specifies TABLE() with only a single function, not
multiple functions, and it seems to require an implicit UNNEST() which is
not what this patch does. There may be something wrong with that reading
of the spec, though, because if it's right then the spec's TABLE() is just
a pointless alternative spelling of UNNEST(). After further review of
that, we might choose to adopt a different syntax for what this patch does,
but in any case this functionality seems clearly worthwhile.
Andrew Gierth, reviewed by Zoltán Böszörményi and Heikki Linnakangas, and
significantly revised by me
12 years ago
</para>
<para>
Multiple function calls can be combined into a
single <literal>FROM</>-clause item by surrounding them
with <literal>ROWS FROM( ... )</>. The output of such an item is the
Support multi-argument UNNEST(), and TABLE() syntax for multiple functions.
This patch adds the ability to write TABLE( function1(), function2(), ...)
as a single FROM-clause entry. The result is the concatenation of the
first row from each function, followed by the second row from each
function, etc; with NULLs inserted if any function produces fewer rows than
others. This is believed to be a much more useful behavior than what
Postgres currently does with multiple SRFs in a SELECT list.
This syntax also provides a reasonable way to combine use of column
definition lists with WITH ORDINALITY: put the column definition list
inside TABLE(), where it's clear that it doesn't control the ordinality
column as well.
Also implement SQL-compliant multiple-argument UNNEST(), by turning
UNNEST(a,b,c) into TABLE(unnest(a), unnest(b), unnest(c)).
The SQL standard specifies TABLE() with only a single function, not
multiple functions, and it seems to require an implicit UNNEST() which is
not what this patch does. There may be something wrong with that reading
of the spec, though, because if it's right then the spec's TABLE() is just
a pointless alternative spelling of UNNEST(). After further review of
that, we might choose to adopt a different syntax for what this patch does,
but in any case this functionality seems clearly worthwhile.
Andrew Gierth, reviewed by Zoltán Böszörményi and Heikki Linnakangas, and
significantly revised by me
12 years ago
concatenation of the first row from each function, then the second
row from each function, etc. If some of the functions produce fewer
rows than others, NULLs are substituted for the missing data, so
that the total number of rows returned is always the same as for the
function that produced the most rows.
</para>
<para>
If the function has been defined as returning the
<type>record</> data type, then an alias or the key word
<literal>AS</> must be present, followed by a column
definition list in the form <literal>( <replaceable
class="parameter">column_name</replaceable> <replaceable
class="parameter">data_type</replaceable> <optional>, ...
</>)</literal>. The column definition list must match the
actual number and types of columns returned by the function.
Support multi-argument UNNEST(), and TABLE() syntax for multiple functions.
This patch adds the ability to write TABLE( function1(), function2(), ...)
as a single FROM-clause entry. The result is the concatenation of the
first row from each function, followed by the second row from each
function, etc; with NULLs inserted if any function produces fewer rows than
others. This is believed to be a much more useful behavior than what
Postgres currently does with multiple SRFs in a SELECT list.
This syntax also provides a reasonable way to combine use of column
definition lists with WITH ORDINALITY: put the column definition list
inside TABLE(), where it's clear that it doesn't control the ordinality
column as well.
Also implement SQL-compliant multiple-argument UNNEST(), by turning
UNNEST(a,b,c) into TABLE(unnest(a), unnest(b), unnest(c)).
The SQL standard specifies TABLE() with only a single function, not
multiple functions, and it seems to require an implicit UNNEST() which is
not what this patch does. There may be something wrong with that reading
of the spec, though, because if it's right then the spec's TABLE() is just
a pointless alternative spelling of UNNEST(). After further review of
that, we might choose to adopt a different syntax for what this patch does,
but in any case this functionality seems clearly worthwhile.
Andrew Gierth, reviewed by Zoltán Böszörményi and Heikki Linnakangas, and
significantly revised by me
12 years ago
</para>
<para>
When using the <literal>ROWS FROM( ... )</> syntax, if one of the
Support multi-argument UNNEST(), and TABLE() syntax for multiple functions.
This patch adds the ability to write TABLE( function1(), function2(), ...)
as a single FROM-clause entry. The result is the concatenation of the
first row from each function, followed by the second row from each
function, etc; with NULLs inserted if any function produces fewer rows than
others. This is believed to be a much more useful behavior than what
Postgres currently does with multiple SRFs in a SELECT list.
This syntax also provides a reasonable way to combine use of column
definition lists with WITH ORDINALITY: put the column definition list
inside TABLE(), where it's clear that it doesn't control the ordinality
column as well.
Also implement SQL-compliant multiple-argument UNNEST(), by turning
UNNEST(a,b,c) into TABLE(unnest(a), unnest(b), unnest(c)).
The SQL standard specifies TABLE() with only a single function, not
multiple functions, and it seems to require an implicit UNNEST() which is
not what this patch does. There may be something wrong with that reading
of the spec, though, because if it's right then the spec's TABLE() is just
a pointless alternative spelling of UNNEST(). After further review of
that, we might choose to adopt a different syntax for what this patch does,
but in any case this functionality seems clearly worthwhile.
Andrew Gierth, reviewed by Zoltán Böszörményi and Heikki Linnakangas, and
significantly revised by me
12 years ago
functions requires a column definition list, it's preferred to put
the column definition list after the function call inside
<literal>ROWS FROM( ... )</>. A column definition list can be placed
after the <literal>ROWS FROM( ... )</> construct only if there's just
a single function and no <literal>WITH ORDINALITY</> clause.
Support multi-argument UNNEST(), and TABLE() syntax for multiple functions.
This patch adds the ability to write TABLE( function1(), function2(), ...)
as a single FROM-clause entry. The result is the concatenation of the
first row from each function, followed by the second row from each
function, etc; with NULLs inserted if any function produces fewer rows than
others. This is believed to be a much more useful behavior than what
Postgres currently does with multiple SRFs in a SELECT list.
This syntax also provides a reasonable way to combine use of column
definition lists with WITH ORDINALITY: put the column definition list
inside TABLE(), where it's clear that it doesn't control the ordinality
column as well.
Also implement SQL-compliant multiple-argument UNNEST(), by turning
UNNEST(a,b,c) into TABLE(unnest(a), unnest(b), unnest(c)).
The SQL standard specifies TABLE() with only a single function, not
multiple functions, and it seems to require an implicit UNNEST() which is
not what this patch does. There may be something wrong with that reading
of the spec, though, because if it's right then the spec's TABLE() is just
a pointless alternative spelling of UNNEST(). After further review of
that, we might choose to adopt a different syntax for what this patch does,
but in any case this functionality seems clearly worthwhile.
Andrew Gierth, reviewed by Zoltán Böszörményi and Heikki Linnakangas, and
significantly revised by me
12 years ago
</para>
<para>
To use <literal>ORDINALITY</literal> together with a column definition
list, you must use the <literal>ROWS FROM( ... )</> syntax and put the
column definition list inside <literal>ROWS FROM( ... )</>.
</para>
</listitem>
</varlistentry>
<varlistentry>
<term><replaceable class="parameter">join_type</replaceable></term>
<listitem>
<para>
One of
<itemizedlist>
<listitem>
<para><literal>[ INNER ] JOIN</literal></para>
</listitem>
<listitem>
<para><literal>LEFT [ OUTER ] JOIN</literal></para>
</listitem>
<listitem>
<para><literal>RIGHT [ OUTER ] JOIN</literal></para>
</listitem>
<listitem>
<para><literal>FULL [ OUTER ] JOIN</literal></para>
</listitem>
<listitem>
<para><literal>CROSS JOIN</literal></para>
</listitem>
</itemizedlist>
For the <literal>INNER</> and <literal>OUTER</> join types, a
join condition must be specified, namely exactly one of
<literal>NATURAL</>, <literal>ON <replaceable
class="parameter">join_condition</replaceable></literal>, or
<literal>USING (<replaceable
class="parameter">join_column</replaceable> [, ...])</literal>.
See below for the meaning. For <literal>CROSS JOIN</literal>,
Update reference documentation on may/can/might:
Standard English uses "may", "can", and "might" in different ways:
may - permission, "You may borrow my rake."
can - ability, "I can lift that log."
might - possibility, "It might rain today."
Unfortunately, in conversational English, their use is often mixed, as
in, "You may use this variable to do X", when in fact, "can" is a better
choice. Similarly, "It may crash" is better stated, "It might crash".
19 years ago
none of these clauses can appear.
</para>
<para>
A <literal>JOIN</literal> clause combines two <literal>FROM</>
items, which for convenience we will refer to as <quote>tables</>,
though in reality they can be any type of <literal>FROM</> item.
Use parentheses if necessary to determine the order of nesting.
In the absence of parentheses, <literal>JOIN</literal>s nest
left-to-right. In any case <literal>JOIN</literal> binds more
tightly than the commas separating <literal>FROM</>-list items.
</para>
<para><literal>CROSS JOIN</> and <literal>INNER JOIN</literal>
produce a simple Cartesian product, the same result as you get from
listing the two tables at the top level of <literal>FROM</>,
but restricted by the join condition (if any).
<literal>CROSS JOIN</> is equivalent to <literal>INNER JOIN ON
(TRUE)</>, that is, no rows are removed by qualification.
These join types are just a notational convenience, since they
do nothing you couldn't do with plain <literal>FROM</> and
<literal>WHERE</>.
</para>
<para><literal>LEFT OUTER JOIN</> returns all rows in the qualified
Cartesian product (i.e., all combined rows that pass its join
condition), plus one copy of each row in the left-hand table
for which there was no right-hand row that passed the join
condition. This left-hand row is extended to the full width
of the joined table by inserting null values for the
right-hand columns. Note that only the <literal>JOIN</>
clause's own condition is considered while deciding which rows
have matches. Outer conditions are applied afterwards.
</para>
Attached are two patches to implement and document anonymous composite
types for Table Functions, as previously proposed on HACKERS. Here is a
brief explanation:
1. Creates a new pg_type typtype: 'p' for pseudo type (currently either
'b' for base or 'c' for catalog, i.e. a class).
2. Creates new builtin type of typtype='p' named RECORD. This is the
first of potentially several pseudo types.
3. Modify FROM clause grammer to accept:
SELECT * FROM my_func() AS m(colname1 type1, colname2 type1, ...)
where m is the table alias, colname1, etc are the column names, and
type1, etc are the column types.
4. When typtype == 'p' and the function return type is RECORD, a list
of column defs is required, and when typtype != 'p', it is
disallowed.
5. A check was added to ensure that the tupdesc provide via the parser
and the actual return tupdesc match in number and type of
attributes.
When creating a function you can do:
CREATE FUNCTION foo(text) RETURNS setof RECORD ...
When using it you can do:
SELECT * from foo(sqlstmt) AS (f1 int, f2 text, f3 timestamp)
or
SELECT * from foo(sqlstmt) AS f(f1 int, f2 text, f3 timestamp)
or
SELECT * from foo(sqlstmt) f(f1 int, f2 text, f3 timestamp)
Included in the patches are adjustments to the regression test sql and
expected files, and documentation.
p.s.
This potentially solves (or at least improves) the issue of builtin
Table Functions. They can be bootstrapped as returning RECORD, and
we can wrap system views around them with properly specified column
defs. For example:
CREATE VIEW pg_settings AS
SELECT s.name, s.setting
FROM show_all_settings()AS s(name text, setting text);
Then we can also add the UPDATE RULE that I previously posted to
pg_settings, and have pg_settings act like a virtual table, allowing
settings to be queried and set.
Joe Conway
23 years ago
<para>
Conversely, <literal>RIGHT OUTER JOIN</> returns all the
joined rows, plus one row for each unmatched right-hand row
(extended with nulls on the left). This is just a notational
convenience, since you could convert it to a <literal>LEFT
OUTER JOIN</> by switching the left and right tables.
Attached are two patches to implement and document anonymous composite
types for Table Functions, as previously proposed on HACKERS. Here is a
brief explanation:
1. Creates a new pg_type typtype: 'p' for pseudo type (currently either
'b' for base or 'c' for catalog, i.e. a class).
2. Creates new builtin type of typtype='p' named RECORD. This is the
first of potentially several pseudo types.
3. Modify FROM clause grammer to accept:
SELECT * FROM my_func() AS m(colname1 type1, colname2 type1, ...)
where m is the table alias, colname1, etc are the column names, and
type1, etc are the column types.
4. When typtype == 'p' and the function return type is RECORD, a list
of column defs is required, and when typtype != 'p', it is
disallowed.
5. A check was added to ensure that the tupdesc provide via the parser
and the actual return tupdesc match in number and type of
attributes.
When creating a function you can do:
CREATE FUNCTION foo(text) RETURNS setof RECORD ...
When using it you can do:
SELECT * from foo(sqlstmt) AS (f1 int, f2 text, f3 timestamp)
or
SELECT * from foo(sqlstmt) AS f(f1 int, f2 text, f3 timestamp)
or
SELECT * from foo(sqlstmt) f(f1 int, f2 text, f3 timestamp)
Included in the patches are adjustments to the regression test sql and
expected files, and documentation.
p.s.
This potentially solves (or at least improves) the issue of builtin
Table Functions. They can be bootstrapped as returning RECORD, and
we can wrap system views around them with properly specified column
defs. For example:
CREATE VIEW pg_settings AS
SELECT s.name, s.setting
FROM show_all_settings()AS s(name text, setting text);
Then we can also add the UPDATE RULE that I previously posted to
pg_settings, and have pg_settings act like a virtual table, allowing
settings to be queried and set.
Joe Conway
23 years ago
</para>
<para><literal>FULL OUTER JOIN</> returns all the joined rows, plus
one row for each unmatched left-hand row (extended with nulls
on the right), plus one row for each unmatched right-hand row
(extended with nulls on the left).
</para>
</listitem>
</varlistentry>
<varlistentry>
<term><literal>ON <replaceable class="parameter">join_condition</replaceable></literal></term>
<listitem>
<para><replaceable class="parameter">join_condition</replaceable> is
an expression resulting in a value of type
<type>boolean</type> (similar to a <literal>WHERE</literal>
clause) that specifies which rows in a join are considered to
match.
</para>
</listitem>
</varlistentry>
<varlistentry>
<term><literal>USING ( <replaceable class="parameter">join_column</replaceable> [, ...] )</literal></term>
<listitem>
<para>
A clause of the form <literal>USING ( a, b, ... )</literal> is
shorthand for <literal>ON left_table.a = right_table.a AND
left_table.b = right_table.b ...</literal>. Also,
<literal>USING</> implies that only one of each pair of
equivalent columns will be included in the join output, not
both.
</para>
</listitem>
</varlistentry>
<varlistentry>
<term><literal>NATURAL</literal></term>
<listitem>
<para><literal>NATURAL</literal> is shorthand for a
<literal>USING</> list that mentions all columns in the two
tables that have the same names.
</para>
</listitem>
</varlistentry>
<varlistentry>
<term><literal>LATERAL</literal></term>
<listitem>
<para>
The <literal>LATERAL</literal> key word can precede a
sub-<command>SELECT</command> <literal>FROM</> item. This allows the
sub-<command>SELECT</command> to refer to columns of <literal>FROM</>
items that appear before it in the <literal>FROM</> list. (Without
<literal>LATERAL</literal>, each sub-<command>SELECT</command> is
evaluated independently and so cannot cross-reference any other
<literal>FROM</> item.)
</para>
<para><literal>LATERAL</literal> can also precede a function-call
<literal>FROM</> item, but in this case it is a noise word, because
the function expression can refer to earlier <literal>FROM</> items
in any case.
</para>
<para>
A <literal>LATERAL</literal> item can appear at top level in the
<literal>FROM</> list, or within a <literal>JOIN</> tree. In the
latter case it can also refer to any items that are on the left-hand
side of a <literal>JOIN</> that it is on the right-hand side of.
</para>
<para>
When a <literal>FROM</> item contains <literal>LATERAL</literal>
cross-references, evaluation proceeds as follows: for each row of the
<literal>FROM</> item providing the cross-referenced column(s), or
set of rows of multiple <literal>FROM</> items providing the
columns, the <literal>LATERAL</literal> item is evaluated using that
row or row set's values of the columns. The resulting row(s) are
joined as usual with the rows they were computed from. This is
repeated for each row or set of rows from the column source table(s).
</para>
<para>
The column source table(s) must be <literal>INNER</> or
<literal>LEFT</> joined to the <literal>LATERAL</literal> item, else
there would not be a well-defined set of rows from which to compute
each set of rows for the <literal>LATERAL</literal> item. Thus,
although a construct such as <literal><replaceable>X</> RIGHT JOIN
LATERAL <replaceable>Y</></literal> is syntactically valid, it is
not actually allowed for <replaceable>Y</> to reference
<replaceable>X</>.
</para>
</listitem>
</varlistentry>
</variablelist>
</para>
</refsect2>
<refsect2 id="SQL-WHERE">
<title id="sql-where-title"><literal>WHERE</literal> Clause</title>
<para>
The optional <literal>WHERE</literal> clause has the general form
<synopsis>
WHERE <replaceable class="parameter">condition</replaceable>
</synopsis>
where <replaceable class="parameter">condition</replaceable> is
any expression that evaluates to a result of type
<type>boolean</type>. Any row that does not satisfy this
condition will be eliminated from the output. A row satisfies the
condition if it returns true when the actual row values are
substituted for any variable references.
</para>
</refsect2>
<refsect2 id="SQL-GROUPBY">
<title id="sql-groupby-title"><literal>GROUP BY</literal> Clause</title>
<para>
The optional <literal>GROUP BY</literal> clause has the general form
<synopsis>
GROUP BY <replaceable class="parameter">expression</replaceable> [, ...]
</synopsis>
</para>
<para>
<literal>GROUP BY</literal> will condense into a single row all
selected rows that share the same values for the grouped
expressions. <replaceable
class="parameter">expression</replaceable> can be an input column
name, or the name or ordinal number of an output column
(<command>SELECT</command> list item), or an arbitrary
expression formed from input-column values. In case of ambiguity,
a <literal>GROUP BY</literal> name will be interpreted as an
input-column name rather than an output column name.
</para>
Attached are two patches to implement and document anonymous composite
types for Table Functions, as previously proposed on HACKERS. Here is a
brief explanation:
1. Creates a new pg_type typtype: 'p' for pseudo type (currently either
'b' for base or 'c' for catalog, i.e. a class).
2. Creates new builtin type of typtype='p' named RECORD. This is the
first of potentially several pseudo types.
3. Modify FROM clause grammer to accept:
SELECT * FROM my_func() AS m(colname1 type1, colname2 type1, ...)
where m is the table alias, colname1, etc are the column names, and
type1, etc are the column types.
4. When typtype == 'p' and the function return type is RECORD, a list
of column defs is required, and when typtype != 'p', it is
disallowed.
5. A check was added to ensure that the tupdesc provide via the parser
and the actual return tupdesc match in number and type of
attributes.
When creating a function you can do:
CREATE FUNCTION foo(text) RETURNS setof RECORD ...
When using it you can do:
SELECT * from foo(sqlstmt) AS (f1 int, f2 text, f3 timestamp)
or
SELECT * from foo(sqlstmt) AS f(f1 int, f2 text, f3 timestamp)
or
SELECT * from foo(sqlstmt) f(f1 int, f2 text, f3 timestamp)
Included in the patches are adjustments to the regression test sql and
expected files, and documentation.
p.s.
This potentially solves (or at least improves) the issue of builtin
Table Functions. They can be bootstrapped as returning RECORD, and
we can wrap system views around them with properly specified column
defs. For example:
CREATE VIEW pg_settings AS
SELECT s.name, s.setting
FROM show_all_settings()AS s(name text, setting text);
Then we can also add the UPDATE RULE that I previously posted to
pg_settings, and have pg_settings act like a virtual table, allowing
settings to be queried and set.
Joe Conway
23 years ago
<para>
Aggregate functions, if any are used, are computed across all rows
making up each group, producing a separate value for each group
(whereas without <literal>GROUP BY</literal>, an aggregate
produces a single value computed across all the selected rows).
The set of rows fed to the aggregate function can be further filtered by
attaching a <literal>FILTER</literal> clause to the aggregate function
call; see <xref linkend="syntax-aggregates"> for more information. When
a <literal>FILTER</literal> clause is present, only those rows matching it
are included.
When <literal>GROUP BY</literal> is present, it is not valid for
the <command>SELECT</command> list expressions to refer to
ungrouped columns except within aggregate functions or if the
ungrouped column is functionally dependent on the grouped columns,
since there would otherwise be more than one possible value to
return for an ungrouped column. A functional dependency exists if
the grouped columns (or a subset thereof) are the primary key of
the table containing the ungrouped column.
</para>
<para>
Currently, <literal>FOR NO KEY UPDATE</>, <literal>FOR UPDATE</>,
<literal>FOR SHARE</> and <literal>FOR KEY SHARE</> cannot be
specified with <literal>GROUP BY</literal>.
</para>
</refsect2>
<refsect2 id="SQL-HAVING">
<title id="sql-having-title"><literal>HAVING</literal> Clause</title>
Attached are two patches to implement and document anonymous composite
types for Table Functions, as previously proposed on HACKERS. Here is a
brief explanation:
1. Creates a new pg_type typtype: 'p' for pseudo type (currently either
'b' for base or 'c' for catalog, i.e. a class).
2. Creates new builtin type of typtype='p' named RECORD. This is the
first of potentially several pseudo types.
3. Modify FROM clause grammer to accept:
SELECT * FROM my_func() AS m(colname1 type1, colname2 type1, ...)
where m is the table alias, colname1, etc are the column names, and
type1, etc are the column types.
4. When typtype == 'p' and the function return type is RECORD, a list
of column defs is required, and when typtype != 'p', it is
disallowed.
5. A check was added to ensure that the tupdesc provide via the parser
and the actual return tupdesc match in number and type of
attributes.
When creating a function you can do:
CREATE FUNCTION foo(text) RETURNS setof RECORD ...
When using it you can do:
SELECT * from foo(sqlstmt) AS (f1 int, f2 text, f3 timestamp)
or
SELECT * from foo(sqlstmt) AS f(f1 int, f2 text, f3 timestamp)
or
SELECT * from foo(sqlstmt) f(f1 int, f2 text, f3 timestamp)
Included in the patches are adjustments to the regression test sql and
expected files, and documentation.
p.s.
This potentially solves (or at least improves) the issue of builtin
Table Functions. They can be bootstrapped as returning RECORD, and
we can wrap system views around them with properly specified column
defs. For example:
CREATE VIEW pg_settings AS
SELECT s.name, s.setting
FROM show_all_settings()AS s(name text, setting text);
Then we can also add the UPDATE RULE that I previously posted to
pg_settings, and have pg_settings act like a virtual table, allowing
settings to be queried and set.
Joe Conway
23 years ago
<para>
The optional <literal>HAVING</literal> clause has the general form
<synopsis>
HAVING <replaceable class="parameter">condition</replaceable>
</synopsis>
where <replaceable class="parameter">condition</replaceable> is
the same as specified for the <literal>WHERE</literal> clause.
</para>
<para>
<literal>HAVING</literal> eliminates group rows that do not
satisfy the condition. <literal>HAVING</literal> is different
from <literal>WHERE</literal>: <literal>WHERE</literal> filters
individual rows before the application of <literal>GROUP
BY</literal>, while <literal>HAVING</literal> filters group rows
created by <literal>GROUP BY</literal>. Each column referenced in
<replaceable class="parameter">condition</replaceable> must
unambiguously reference a grouping column, unless the reference
appears within an aggregate function.
</para>
<para>
The presence of <literal>HAVING</literal> turns a query into a grouped
query even if there is no <literal>GROUP BY</> clause. This is the
same as what happens when the query contains aggregate functions but
no <literal>GROUP BY</> clause. All the selected rows are considered to
form a single group, and the <command>SELECT</command> list and
<literal>HAVING</literal> clause can only reference table columns from
within aggregate functions. Such a query will emit a single row if the
<literal>HAVING</literal> condition is true, zero rows if it is not true.
</para>
<para>
Currently, <literal>FOR NO KEY UPDATE</>, <literal>FOR UPDATE</>,
<literal>FOR SHARE</> and <literal>FOR KEY SHARE</> cannot be
specified with <literal>HAVING</literal>.
</para>
</refsect2>
<refsect2 id="SQL-WINDOW">
<title id="sql-window-title"><literal>WINDOW</literal> Clause</title>
<para>
The optional <literal>WINDOW</literal> clause has the general form
<synopsis>
WINDOW <replaceable class="parameter">window_name</replaceable> AS ( <replaceable class="parameter">window_definition</replaceable> ) [, ...]
</synopsis>
where <replaceable class="parameter">window_name</replaceable> is
a name that can be referenced from <literal>OVER</> clauses or
subsequent window definitions, and
<replaceable class="parameter">window_definition</replaceable> is
<synopsis>
[ <replaceable class="parameter">existing_window_name</replaceable> ]
[ PARTITION BY <replaceable class="parameter">expression</replaceable> [, ...] ]
[ ORDER BY <replaceable class="parameter">expression</replaceable> [ ASC | DESC | USING <replaceable class="parameter">operator</replaceable> ] [ NULLS { FIRST | LAST } ] [, ...] ]
[ <replaceable class="parameter">frame_clause</replaceable> ]
</synopsis>
</para>
<para>
If an <replaceable class="parameter">existing_window_name</replaceable>
is specified it must refer to an earlier entry in the <literal>WINDOW</>
list; the new window copies its partitioning clause from that entry,
as well as its ordering clause if any. In this case the new window cannot
specify its own <literal>PARTITION BY</> clause, and it can specify
<literal>ORDER BY</> only if the copied window does not have one.
The new window always uses its own frame clause; the copied window
must not specify a frame clause.
</para>
<para>
The elements of the <literal>PARTITION BY</> list are interpreted in
much the same fashion as elements of a
<xref linkend="sql-groupby" endterm="sql-groupby-title">, except that
they are always simple expressions and never the name or number of an
output column.
Another difference is that these expressions can contain aggregate
function calls, which are not allowed in a regular <literal>GROUP BY</>
clause. They are allowed here because windowing occurs after grouping
and aggregation.
</para>
<para>
Similarly, the elements of the <literal>ORDER BY</> list are interpreted
in much the same fashion as elements of an
<xref linkend="sql-orderby" endterm="sql-orderby-title">, except that
the expressions are always taken as simple expressions and never the name
or number of an output column.
</para>
<para>
The optional <replaceable class="parameter">frame_clause</> defines
the <firstterm>window frame</> for window functions that depend on the
frame (not all do). The window frame is a set of related rows for
each row of the query (called the <firstterm>current row</>).
The <replaceable class="parameter">frame_clause</> can be one of
<synopsis>
[ RANGE | ROWS ] <replaceable>frame_start</>
[ RANGE | ROWS ] BETWEEN <replaceable>frame_start</> AND <replaceable>frame_end</>
</synopsis>
where <replaceable>frame_start</> and <replaceable>frame_end</> can be
one of
<synopsis>
UNBOUNDED PRECEDING
<replaceable>value</replaceable> PRECEDING
CURRENT ROW
<replaceable>value</replaceable> FOLLOWING
UNBOUNDED FOLLOWING
</synopsis>
If <replaceable>frame_end</> is omitted it defaults to <literal>CURRENT
ROW</>. Restrictions are that
<replaceable>frame_start</> cannot be <literal>UNBOUNDED FOLLOWING</>,
<replaceable>frame_end</> cannot be <literal>UNBOUNDED PRECEDING</>,
and the <replaceable>frame_end</> choice cannot appear earlier in the
above list than the <replaceable>frame_start</> choice — for example
<literal>RANGE BETWEEN CURRENT ROW AND <replaceable>value</>
PRECEDING</literal> is not allowed.
</para>
<para>
The default framing option is <literal>RANGE UNBOUNDED PRECEDING</>,
which is the same as <literal>RANGE BETWEEN UNBOUNDED PRECEDING AND
CURRENT ROW</>; it sets the frame to be all rows from the partition start
up through the current row's last peer in the <literal>ORDER BY</>
ordering (which means all rows if there is no <literal>ORDER BY</>).
In general, <literal>UNBOUNDED PRECEDING</> means that the frame
starts with the first row of the partition, and similarly
<literal>UNBOUNDED FOLLOWING</> means that the frame ends with the last
row of the partition (regardless of <literal>RANGE</> or <literal>ROWS</>
mode). In <literal>ROWS</> mode, <literal>CURRENT ROW</>
means that the frame starts or ends with the current row; but in
<literal>RANGE</> mode it means that the frame starts or ends with
the current row's first or last peer in the <literal>ORDER BY</> ordering.
The <replaceable>value</> <literal>PRECEDING</> and
<replaceable>value</> <literal>FOLLOWING</> cases are currently only
allowed in <literal>ROWS</> mode. They indicate that the frame starts
or ends with the row that many rows before or after the current row.
<replaceable>value</replaceable> must be an integer expression not
containing any variables, aggregate functions, or window functions.
The value must not be null or negative; but it can be zero, which
selects the current row itself.
</para>
<para>
Beware that the <literal>ROWS</> options can produce unpredictable
results if the <literal>ORDER BY</> ordering does not order the rows
uniquely. The <literal>RANGE</> options are designed to ensure that
rows that are peers in the <literal>ORDER BY</> ordering are treated
alike; any two peer rows will be both in or both not in the frame.
</para>
<para>
The purpose of a <literal>WINDOW</literal> clause is to specify the
behavior of <firstterm>window functions</> appearing in the query's
<xref linkend="sql-select-list" endterm="sql-select-list-title"> or
<xref linkend="sql-orderby" endterm="sql-orderby-title">. These functions
can reference the <literal>WINDOW</literal> clause entries by name
in their <literal>OVER</> clauses. A <literal>WINDOW</literal> clause
entry does not have to be referenced anywhere, however; if it is not
used in the query it is simply ignored. It is possible to use window
functions without any <literal>WINDOW</literal> clause at all, since
a window function call can specify its window definition directly in
its <literal>OVER</> clause. However, the <literal>WINDOW</literal>
clause saves typing when the same window definition is needed for more
than one window function.
</para>
<para>
Currently, <literal>FOR NO KEY UPDATE</>, <literal>FOR UPDATE</>,
<literal>FOR SHARE</> and <literal>FOR KEY SHARE</> cannot be
specified with <literal>WINDOW</literal>.
</para>
<para>
Window functions are described in detail in
<xref linkend="tutorial-window">,
<xref linkend="syntax-window-functions">, and
<xref linkend="queries-window">.
</para>
</refsect2>
<refsect2 id="sql-select-list">
<title id="sql-select-list-title"><command>SELECT</command> List</title>
<para>
The <command>SELECT</command> list (between the key words
<literal>SELECT</> and <literal>FROM</>) specifies expressions
that form the output rows of the <command>SELECT</command>
statement. The expressions can (and usually do) refer to columns
computed in the <literal>FROM</> clause.
</para>
<para>
Just as in a table, every output column of a <command>SELECT</command>
has a name. In a simple <command>SELECT</command> this name is just
used to label the column for display, but when the <command>SELECT</>
is a sub-query of a larger query, the name is seen by the larger query
as the column name of the virtual table produced by the sub-query.
To specify the name to use for an output column, write
<literal>AS</> <replaceable class="parameter">output_name</replaceable>
after the column's expression. (You can omit <literal>AS</literal>,
but only if the desired output name does not match any
<productname>PostgreSQL</productname> keyword (see <xref
linkend="sql-keywords-appendix">). For protection against possible
future keyword additions, it is recommended that you always either
write <literal>AS</literal> or double-quote the output name.)
If you do not specify a column name, a name is chosen automatically
by <productname>PostgreSQL</productname>. If the column's expression
is a simple column reference then the chosen name is the same as that
column's name. In more complex cases a function or type name may be
used, or the system may fall back on a generated name such as
<literal>?column?</literal>.
</para>
<para>
An output column's name can be used to refer to the column's value in
<literal>ORDER BY</> and <literal>GROUP BY</> clauses, but not in the
<literal>WHERE</> or <literal>HAVING</> clauses; there you must write
out the expression instead.
</para>
<para>
Instead of an expression, <literal>*</literal> can be written in
the output list as a shorthand for all the columns of the selected
rows. Also, you can write <literal><replaceable
class="parameter">table_name</replaceable>.*</literal> as a
shorthand for the columns coming from just that table. In these
cases it is not possible to specify new names with <literal>AS</>;
the output column names will be the same as the table columns' names.
</para>
</refsect2>
<refsect2 id="sql-distinct">
<title id="sql-distinct-title"><literal>DISTINCT</literal> Clause</title>
<para>
If <literal>SELECT DISTINCT</> is specified, all duplicate rows are
removed from the result set (one row is kept from each group of
duplicates). <literal>SELECT ALL</> specifies the opposite: all rows are
kept; that is the default.
</para>
<para>
<literal>SELECT DISTINCT ON ( <replaceable
class="parameter">expression</replaceable> [, ...] )</literal>
keeps only the first row of each set of rows where the given
expressions evaluate to equal. The <literal>DISTINCT ON</literal>
expressions are interpreted using the same rules as for
<literal>ORDER BY</> (see above). Note that the <quote>first
row</quote> of each set is unpredictable unless <literal>ORDER
BY</> is used to ensure that the desired row appears first. For
example:
<programlisting>
SELECT DISTINCT ON (location) location, time, report
FROM weather_reports
ORDER BY location, time DESC;
</programlisting>
retrieves the most recent weather report for each location. But
if we had not used <literal>ORDER BY</> to force descending order
of time values for each location, we'd have gotten a report from
an unpredictable time for each location.
</para>
<para>
The <literal>DISTINCT ON</> expression(s) must match the leftmost
<literal>ORDER BY</> expression(s). The <literal>ORDER BY</> clause
will normally contain additional expression(s) that determine the
desired precedence of rows within each <literal>DISTINCT ON</> group.
</para>
<para>
Currently, <literal>FOR NO KEY UPDATE</>, <literal>FOR UPDATE</>,
<literal>FOR SHARE</> and <literal>FOR KEY SHARE</> cannot be
specified with <literal>DISTINCT</literal>.
</para>
</refsect2>
<refsect2 id="SQL-UNION">
<title id="sql-union-title"><literal>UNION</literal> Clause</title>
<para>
The <literal>UNION</literal> clause has this general form:
<synopsis>
<replaceable class="parameter">select_statement</replaceable> UNION [ ALL | DISTINCT ] <replaceable class="parameter">select_statement</replaceable>
</synopsis><replaceable class="parameter">select_statement</replaceable> is
any <command>SELECT</command> statement without an <literal>ORDER
Improve concurrency of foreign key locking
This patch introduces two additional lock modes for tuples: "SELECT FOR
KEY SHARE" and "SELECT FOR NO KEY UPDATE". These don't block each
other, in contrast with already existing "SELECT FOR SHARE" and "SELECT
FOR UPDATE". UPDATE commands that do not modify the values stored in
the columns that are part of the key of the tuple now grab a SELECT FOR
NO KEY UPDATE lock on the tuple, allowing them to proceed concurrently
with tuple locks of the FOR KEY SHARE variety.
Foreign key triggers now use FOR KEY SHARE instead of FOR SHARE; this
means the concurrency improvement applies to them, which is the whole
point of this patch.
The added tuple lock semantics require some rejiggering of the multixact
module, so that the locking level that each transaction is holding can
be stored alongside its Xid. Also, multixacts now need to persist
across server restarts and crashes, because they can now represent not
only tuple locks, but also tuple updates. This means we need more
careful tracking of lifetime of pg_multixact SLRU files; since they now
persist longer, we require more infrastructure to figure out when they
can be removed. pg_upgrade also needs to be careful to copy
pg_multixact files over from the old server to the new, or at least part
of multixact.c state, depending on the versions of the old and new
servers.
Tuple time qualification rules (HeapTupleSatisfies routines) need to be
careful not to consider tuples with the "is multi" infomask bit set as
being only locked; they might need to look up MultiXact values (i.e.
possibly do pg_multixact I/O) to find out the Xid that updated a tuple,
whereas they previously were assured to only use information readily
available from the tuple header. This is considered acceptable, because
the extra I/O would involve cases that would previously cause some
commands to block waiting for concurrent transactions to finish.
Another important change is the fact that locking tuples that have
previously been updated causes the future versions to be marked as
locked, too; this is essential for correctness of foreign key checks.
This causes additional WAL-logging, also (there was previously a single
WAL record for a locked tuple; now there are as many as updated copies
of the tuple there exist.)
With all this in place, contention related to tuples being checked by
foreign key rules should be much reduced.
As a bonus, the old behavior that a subtransaction grabbing a stronger
tuple lock than the parent (sub)transaction held on a given tuple and
later aborting caused the weaker lock to be lost, has been fixed.
Many new spec files were added for isolation tester framework, to ensure
overall behavior is sane. There's probably room for several more tests.
There were several reviewers of this patch; in particular, Noah Misch
and Andres Freund spent considerable time in it. Original idea for the
patch came from Simon Riggs, after a problem report by Joel Jacobson.
Most code is from me, with contributions from Marti Raudsepp, Alexander
Shulgin, Noah Misch and Andres Freund.
This patch was discussed in several pgsql-hackers threads; the most
important start at the following message-ids:
AANLkTimo9XVcEzfiBR-ut3KVNDkjm2Vxh+t8kAmWjPuv@mail.gmail.com
1290721684-sup-3951@alvh.no-ip.org
1294953201-sup-2099@alvh.no-ip.org
1320343602-sup-2290@alvh.no-ip.org
1339690386-sup-8927@alvh.no-ip.org
4FE5FF020200002500048A3D@gw.wicourts.gov
4FEAB90A0200002500048B7D@gw.wicourts.gov
13 years ago
BY</>, <literal>LIMIT</>, <literal>FOR NO KEY UPDATE</>, <literal>FOR UPDATE</literal>,
<literal>FOR SHARE</literal>, or <literal>FOR KEY SHARE</literal> clause.
(<literal>ORDER BY</> and <literal>LIMIT</> can be attached to a
subexpression if it is enclosed in parentheses. Without
parentheses, these clauses will be taken to apply to the result of
the <literal>UNION</literal>, not to its right-hand input
expression.)
</para>
<para>
The <literal>UNION</literal> operator computes the set union of
the rows returned by the involved <command>SELECT</command>
statements. A row is in the set union of two result sets if it
appears in at least one of the result sets. The two
<command>SELECT</command> statements that represent the direct
operands of the <literal>UNION</literal> must produce the same
number of columns, and corresponding columns must be of compatible
data types.
</para>
<para>
The result of <literal>UNION</> does not contain any duplicate
rows unless the <literal>ALL</> option is specified.
<literal>ALL</> prevents elimination of duplicates. (Therefore,
<literal>UNION ALL</> is usually significantly quicker than
<literal>UNION</>; use <literal>ALL</> when you can.)
<literal>DISTINCT</> can be written to explicitly specify the
default behavior of eliminating duplicate rows.
</para>
<para>
Multiple <literal>UNION</> operators in the same
<command>SELECT</command> statement are evaluated left to right,
unless otherwise indicated by parentheses.
</para>
<para>
Improve concurrency of foreign key locking
This patch introduces two additional lock modes for tuples: "SELECT FOR
KEY SHARE" and "SELECT FOR NO KEY UPDATE". These don't block each
other, in contrast with already existing "SELECT FOR SHARE" and "SELECT
FOR UPDATE". UPDATE commands that do not modify the values stored in
the columns that are part of the key of the tuple now grab a SELECT FOR
NO KEY UPDATE lock on the tuple, allowing them to proceed concurrently
with tuple locks of the FOR KEY SHARE variety.
Foreign key triggers now use FOR KEY SHARE instead of FOR SHARE; this
means the concurrency improvement applies to them, which is the whole
point of this patch.
The added tuple lock semantics require some rejiggering of the multixact
module, so that the locking level that each transaction is holding can
be stored alongside its Xid. Also, multixacts now need to persist
across server restarts and crashes, because they can now represent not
only tuple locks, but also tuple updates. This means we need more
careful tracking of lifetime of pg_multixact SLRU files; since they now
persist longer, we require more infrastructure to figure out when they
can be removed. pg_upgrade also needs to be careful to copy
pg_multixact files over from the old server to the new, or at least part
of multixact.c state, depending on the versions of the old and new
servers.
Tuple time qualification rules (HeapTupleSatisfies routines) need to be
careful not to consider tuples with the "is multi" infomask bit set as
being only locked; they might need to look up MultiXact values (i.e.
possibly do pg_multixact I/O) to find out the Xid that updated a tuple,
whereas they previously were assured to only use information readily
available from the tuple header. This is considered acceptable, because
the extra I/O would involve cases that would previously cause some
commands to block waiting for concurrent transactions to finish.
Another important change is the fact that locking tuples that have
previously been updated causes the future versions to be marked as
locked, too; this is essential for correctness of foreign key checks.
This causes additional WAL-logging, also (there was previously a single
WAL record for a locked tuple; now there are as many as updated copies
of the tuple there exist.)
With all this in place, contention related to tuples being checked by
foreign key rules should be much reduced.
As a bonus, the old behavior that a subtransaction grabbing a stronger
tuple lock than the parent (sub)transaction held on a given tuple and
later aborting caused the weaker lock to be lost, has been fixed.
Many new spec files were added for isolation tester framework, to ensure
overall behavior is sane. There's probably room for several more tests.
There were several reviewers of this patch; in particular, Noah Misch
and Andres Freund spent considerable time in it. Original idea for the
patch came from Simon Riggs, after a problem report by Joel Jacobson.
Most code is from me, with contributions from Marti Raudsepp, Alexander
Shulgin, Noah Misch and Andres Freund.
This patch was discussed in several pgsql-hackers threads; the most
important start at the following message-ids:
AANLkTimo9XVcEzfiBR-ut3KVNDkjm2Vxh+t8kAmWjPuv@mail.gmail.com
1290721684-sup-3951@alvh.no-ip.org
1294953201-sup-2099@alvh.no-ip.org
1320343602-sup-2290@alvh.no-ip.org
1339690386-sup-8927@alvh.no-ip.org
4FE5FF020200002500048A3D@gw.wicourts.gov
4FEAB90A0200002500048B7D@gw.wicourts.gov
13 years ago
Currently, <literal>FOR NO KEY UPDATE</>, <literal>FOR UPDATE</>, <literal>FOR SHARE</> and
<literal>FOR KEY SHARE</> cannot be
specified either for a <literal>UNION</> result or for any input of a
<literal>UNION</>.
</para>
</refsect2>
<refsect2 id="SQL-INTERSECT">
<title id="sql-intersect-title"><literal>INTERSECT</literal> Clause</title>
<para>
The <literal>INTERSECT</literal> clause has this general form:
<synopsis>
<replaceable class="parameter">select_statement</replaceable> INTERSECT [ ALL | DISTINCT ] <replaceable class="parameter">select_statement</replaceable>
</synopsis><replaceable class="parameter">select_statement</replaceable> is
any <command>SELECT</command> statement without an <literal>ORDER
Improve concurrency of foreign key locking
This patch introduces two additional lock modes for tuples: "SELECT FOR
KEY SHARE" and "SELECT FOR NO KEY UPDATE". These don't block each
other, in contrast with already existing "SELECT FOR SHARE" and "SELECT
FOR UPDATE". UPDATE commands that do not modify the values stored in
the columns that are part of the key of the tuple now grab a SELECT FOR
NO KEY UPDATE lock on the tuple, allowing them to proceed concurrently
with tuple locks of the FOR KEY SHARE variety.
Foreign key triggers now use FOR KEY SHARE instead of FOR SHARE; this
means the concurrency improvement applies to them, which is the whole
point of this patch.
The added tuple lock semantics require some rejiggering of the multixact
module, so that the locking level that each transaction is holding can
be stored alongside its Xid. Also, multixacts now need to persist
across server restarts and crashes, because they can now represent not
only tuple locks, but also tuple updates. This means we need more
careful tracking of lifetime of pg_multixact SLRU files; since they now
persist longer, we require more infrastructure to figure out when they
can be removed. pg_upgrade also needs to be careful to copy
pg_multixact files over from the old server to the new, or at least part
of multixact.c state, depending on the versions of the old and new
servers.
Tuple time qualification rules (HeapTupleSatisfies routines) need to be
careful not to consider tuples with the "is multi" infomask bit set as
being only locked; they might need to look up MultiXact values (i.e.
possibly do pg_multixact I/O) to find out the Xid that updated a tuple,
whereas they previously were assured to only use information readily
available from the tuple header. This is considered acceptable, because
the extra I/O would involve cases that would previously cause some
commands to block waiting for concurrent transactions to finish.
Another important change is the fact that locking tuples that have
previously been updated causes the future versions to be marked as
locked, too; this is essential for correctness of foreign key checks.
This causes additional WAL-logging, also (there was previously a single
WAL record for a locked tuple; now there are as many as updated copies
of the tuple there exist.)
With all this in place, contention related to tuples being checked by
foreign key rules should be much reduced.
As a bonus, the old behavior that a subtransaction grabbing a stronger
tuple lock than the parent (sub)transaction held on a given tuple and
later aborting caused the weaker lock to be lost, has been fixed.
Many new spec files were added for isolation tester framework, to ensure
overall behavior is sane. There's probably room for several more tests.
There were several reviewers of this patch; in particular, Noah Misch
and Andres Freund spent considerable time in it. Original idea for the
patch came from Simon Riggs, after a problem report by Joel Jacobson.
Most code is from me, with contributions from Marti Raudsepp, Alexander
Shulgin, Noah Misch and Andres Freund.
This patch was discussed in several pgsql-hackers threads; the most
important start at the following message-ids:
AANLkTimo9XVcEzfiBR-ut3KVNDkjm2Vxh+t8kAmWjPuv@mail.gmail.com
1290721684-sup-3951@alvh.no-ip.org
1294953201-sup-2099@alvh.no-ip.org
1320343602-sup-2290@alvh.no-ip.org
1339690386-sup-8927@alvh.no-ip.org
4FE5FF020200002500048A3D@gw.wicourts.gov
4FEAB90A0200002500048B7D@gw.wicourts.gov
13 years ago
BY</>, <literal>LIMIT</>, <literal>FOR NO KEY UPDATE</>, <literal>FOR UPDATE</literal>,
<literal>FOR SHARE</literal>, or <literal>FOR KEY SHARE</> clause.
</para>
<para>
The <literal>INTERSECT</literal> operator computes the set
intersection of the rows returned by the involved
<command>SELECT</command> statements. A row is in the
intersection of two result sets if it appears in both result sets.
</para>
<para>
The result of <literal>INTERSECT</literal> does not contain any
duplicate rows unless the <literal>ALL</> option is specified.
With <literal>ALL</>, a row that has <replaceable>m</> duplicates in the
left table and <replaceable>n</> duplicates in the right table will appear
min(<replaceable>m</>,<replaceable>n</>) times in the result set.
<literal>DISTINCT</> can be written to explicitly specify the
default behavior of eliminating duplicate rows.
</para>
<para>
Multiple <literal>INTERSECT</literal> operators in the same
<command>SELECT</command> statement are evaluated left to right,
unless parentheses dictate otherwise.
<literal>INTERSECT</literal> binds more tightly than
<literal>UNION</literal>. That is, <literal>A UNION B INTERSECT
C</literal> will be read as <literal>A UNION (B INTERSECT
C)</literal>.
</para>
<para>
Improve concurrency of foreign key locking
This patch introduces two additional lock modes for tuples: "SELECT FOR
KEY SHARE" and "SELECT FOR NO KEY UPDATE". These don't block each
other, in contrast with already existing "SELECT FOR SHARE" and "SELECT
FOR UPDATE". UPDATE commands that do not modify the values stored in
the columns that are part of the key of the tuple now grab a SELECT FOR
NO KEY UPDATE lock on the tuple, allowing them to proceed concurrently
with tuple locks of the FOR KEY SHARE variety.
Foreign key triggers now use FOR KEY SHARE instead of FOR SHARE; this
means the concurrency improvement applies to them, which is the whole
point of this patch.
The added tuple lock semantics require some rejiggering of the multixact
module, so that the locking level that each transaction is holding can
be stored alongside its Xid. Also, multixacts now need to persist
across server restarts and crashes, because they can now represent not
only tuple locks, but also tuple updates. This means we need more
careful tracking of lifetime of pg_multixact SLRU files; since they now
persist longer, we require more infrastructure to figure out when they
can be removed. pg_upgrade also needs to be careful to copy
pg_multixact files over from the old server to the new, or at least part
of multixact.c state, depending on the versions of the old and new
servers.
Tuple time qualification rules (HeapTupleSatisfies routines) need to be
careful not to consider tuples with the "is multi" infomask bit set as
being only locked; they might need to look up MultiXact values (i.e.
possibly do pg_multixact I/O) to find out the Xid that updated a tuple,
whereas they previously were assured to only use information readily
available from the tuple header. This is considered acceptable, because
the extra I/O would involve cases that would previously cause some
commands to block waiting for concurrent transactions to finish.
Another important change is the fact that locking tuples that have
previously been updated causes the future versions to be marked as
locked, too; this is essential for correctness of foreign key checks.
This causes additional WAL-logging, also (there was previously a single
WAL record for a locked tuple; now there are as many as updated copies
of the tuple there exist.)
With all this in place, contention related to tuples being checked by
foreign key rules should be much reduced.
As a bonus, the old behavior that a subtransaction grabbing a stronger
tuple lock than the parent (sub)transaction held on a given tuple and
later aborting caused the weaker lock to be lost, has been fixed.
Many new spec files were added for isolation tester framework, to ensure
overall behavior is sane. There's probably room for several more tests.
There were several reviewers of this patch; in particular, Noah Misch
and Andres Freund spent considerable time in it. Original idea for the
patch came from Simon Riggs, after a problem report by Joel Jacobson.
Most code is from me, with contributions from Marti Raudsepp, Alexander
Shulgin, Noah Misch and Andres Freund.
This patch was discussed in several pgsql-hackers threads; the most
important start at the following message-ids:
AANLkTimo9XVcEzfiBR-ut3KVNDkjm2Vxh+t8kAmWjPuv@mail.gmail.com
1290721684-sup-3951@alvh.no-ip.org
1294953201-sup-2099@alvh.no-ip.org
1320343602-sup-2290@alvh.no-ip.org
1339690386-sup-8927@alvh.no-ip.org
4FE5FF020200002500048A3D@gw.wicourts.gov
4FEAB90A0200002500048B7D@gw.wicourts.gov
13 years ago
Currently, <literal>FOR NO KEY UPDATE</>, <literal>FOR UPDATE</>, <literal>FOR SHARE</> and
<literal>FOR KEY SHARE</> cannot be
specified either for an <literal>INTERSECT</> result or for any input of
an <literal>INTERSECT</>.
</para>
</refsect2>
<refsect2 id="SQL-EXCEPT">
<title id="sql-except-title"><literal>EXCEPT</literal> Clause</title>
<para>
The <literal>EXCEPT</literal> clause has this general form:
<synopsis>
<replaceable class="parameter">select_statement</replaceable> EXCEPT [ ALL | DISTINCT ] <replaceable class="parameter">select_statement</replaceable>
</synopsis><replaceable class="parameter">select_statement</replaceable> is
any <command>SELECT</command> statement without an <literal>ORDER
Improve concurrency of foreign key locking
This patch introduces two additional lock modes for tuples: "SELECT FOR
KEY SHARE" and "SELECT FOR NO KEY UPDATE". These don't block each
other, in contrast with already existing "SELECT FOR SHARE" and "SELECT
FOR UPDATE". UPDATE commands that do not modify the values stored in
the columns that are part of the key of the tuple now grab a SELECT FOR
NO KEY UPDATE lock on the tuple, allowing them to proceed concurrently
with tuple locks of the FOR KEY SHARE variety.
Foreign key triggers now use FOR KEY SHARE instead of FOR SHARE; this
means the concurrency improvement applies to them, which is the whole
point of this patch.
The added tuple lock semantics require some rejiggering of the multixact
module, so that the locking level that each transaction is holding can
be stored alongside its Xid. Also, multixacts now need to persist
across server restarts and crashes, because they can now represent not
only tuple locks, but also tuple updates. This means we need more
careful tracking of lifetime of pg_multixact SLRU files; since they now
persist longer, we require more infrastructure to figure out when they
can be removed. pg_upgrade also needs to be careful to copy
pg_multixact files over from the old server to the new, or at least part
of multixact.c state, depending on the versions of the old and new
servers.
Tuple time qualification rules (HeapTupleSatisfies routines) need to be
careful not to consider tuples with the "is multi" infomask bit set as
being only locked; they might need to look up MultiXact values (i.e.
possibly do pg_multixact I/O) to find out the Xid that updated a tuple,
whereas they previously were assured to only use information readily
available from the tuple header. This is considered acceptable, because
the extra I/O would involve cases that would previously cause some
commands to block waiting for concurrent transactions to finish.
Another important change is the fact that locking tuples that have
previously been updated causes the future versions to be marked as
locked, too; this is essential for correctness of foreign key checks.
This causes additional WAL-logging, also (there was previously a single
WAL record for a locked tuple; now there are as many as updated copies
of the tuple there exist.)
With all this in place, contention related to tuples being checked by
foreign key rules should be much reduced.
As a bonus, the old behavior that a subtransaction grabbing a stronger
tuple lock than the parent (sub)transaction held on a given tuple and
later aborting caused the weaker lock to be lost, has been fixed.
Many new spec files were added for isolation tester framework, to ensure
overall behavior is sane. There's probably room for several more tests.
There were several reviewers of this patch; in particular, Noah Misch
and Andres Freund spent considerable time in it. Original idea for the
patch came from Simon Riggs, after a problem report by Joel Jacobson.
Most code is from me, with contributions from Marti Raudsepp, Alexander
Shulgin, Noah Misch and Andres Freund.
This patch was discussed in several pgsql-hackers threads; the most
important start at the following message-ids:
AANLkTimo9XVcEzfiBR-ut3KVNDkjm2Vxh+t8kAmWjPuv@mail.gmail.com
1290721684-sup-3951@alvh.no-ip.org
1294953201-sup-2099@alvh.no-ip.org
1320343602-sup-2290@alvh.no-ip.org
1339690386-sup-8927@alvh.no-ip.org
4FE5FF020200002500048A3D@gw.wicourts.gov
4FEAB90A0200002500048B7D@gw.wicourts.gov
13 years ago
BY</>, <literal>LIMIT</>, <literal>FOR NO KEY UPDATE</>, <literal>FOR UPDATE</literal>,
<literal>FOR SHARE</literal>, or <literal>FOR KEY SHARE</> clause.
</para>
<para>
The <literal>EXCEPT</literal> operator computes the set of rows
that are in the result of the left <command>SELECT</command>
statement but not in the result of the right one.
</para>
<para>
The result of <literal>EXCEPT</literal> does not contain any
duplicate rows unless the <literal>ALL</> option is specified.
With <literal>ALL</>, a row that has <replaceable>m</> duplicates in the
left table and <replaceable>n</> duplicates in the right table will appear
max(<replaceable>m</>-<replaceable>n</>,0) times in the result set.
<literal>DISTINCT</> can be written to explicitly specify the
default behavior of eliminating duplicate rows.
</para>
<para>
Multiple <literal>EXCEPT</literal> operators in the same
<command>SELECT</command> statement are evaluated left to right,
unless parentheses dictate otherwise. <literal>EXCEPT</> binds at
the same level as <literal>UNION</>.
</para>
<para>
Improve concurrency of foreign key locking
This patch introduces two additional lock modes for tuples: "SELECT FOR
KEY SHARE" and "SELECT FOR NO KEY UPDATE". These don't block each
other, in contrast with already existing "SELECT FOR SHARE" and "SELECT
FOR UPDATE". UPDATE commands that do not modify the values stored in
the columns that are part of the key of the tuple now grab a SELECT FOR
NO KEY UPDATE lock on the tuple, allowing them to proceed concurrently
with tuple locks of the FOR KEY SHARE variety.
Foreign key triggers now use FOR KEY SHARE instead of FOR SHARE; this
means the concurrency improvement applies to them, which is the whole
point of this patch.
The added tuple lock semantics require some rejiggering of the multixact
module, so that the locking level that each transaction is holding can
be stored alongside its Xid. Also, multixacts now need to persist
across server restarts and crashes, because they can now represent not
only tuple locks, but also tuple updates. This means we need more
careful tracking of lifetime of pg_multixact SLRU files; since they now
persist longer, we require more infrastructure to figure out when they
can be removed. pg_upgrade also needs to be careful to copy
pg_multixact files over from the old server to the new, or at least part
of multixact.c state, depending on the versions of the old and new
servers.
Tuple time qualification rules (HeapTupleSatisfies routines) need to be
careful not to consider tuples with the "is multi" infomask bit set as
being only locked; they might need to look up MultiXact values (i.e.
possibly do pg_multixact I/O) to find out the Xid that updated a tuple,
whereas they previously were assured to only use information readily
available from the tuple header. This is considered acceptable, because
the extra I/O would involve cases that would previously cause some
commands to block waiting for concurrent transactions to finish.
Another important change is the fact that locking tuples that have
previously been updated causes the future versions to be marked as
locked, too; this is essential for correctness of foreign key checks.
This causes additional WAL-logging, also (there was previously a single
WAL record for a locked tuple; now there are as many as updated copies
of the tuple there exist.)
With all this in place, contention related to tuples being checked by
foreign key rules should be much reduced.
As a bonus, the old behavior that a subtransaction grabbing a stronger
tuple lock than the parent (sub)transaction held on a given tuple and
later aborting caused the weaker lock to be lost, has been fixed.
Many new spec files were added for isolation tester framework, to ensure
overall behavior is sane. There's probably room for several more tests.
There were several reviewers of this patch; in particular, Noah Misch
and Andres Freund spent considerable time in it. Original idea for the
patch came from Simon Riggs, after a problem report by Joel Jacobson.
Most code is from me, with contributions from Marti Raudsepp, Alexander
Shulgin, Noah Misch and Andres Freund.
This patch was discussed in several pgsql-hackers threads; the most
important start at the following message-ids:
AANLkTimo9XVcEzfiBR-ut3KVNDkjm2Vxh+t8kAmWjPuv@mail.gmail.com
1290721684-sup-3951@alvh.no-ip.org
1294953201-sup-2099@alvh.no-ip.org
1320343602-sup-2290@alvh.no-ip.org
1339690386-sup-8927@alvh.no-ip.org
4FE5FF020200002500048A3D@gw.wicourts.gov
4FEAB90A0200002500048B7D@gw.wicourts.gov
13 years ago
Currently, <literal>FOR NO KEY UPDATE</>, <literal>FOR UPDATE</>, <literal>FOR SHARE</> and
<literal>FOR KEY SHARE</> cannot be
specified either for an <literal>EXCEPT</> result or for any input of
an <literal>EXCEPT</>.
</para>
</refsect2>
<refsect2 id="SQL-ORDERBY">
<title id="sql-orderby-title"><literal>ORDER BY</literal> Clause</title>
<para>
The optional <literal>ORDER BY</literal> clause has this general form:
<synopsis>
ORDER BY <replaceable class="parameter">expression</replaceable> [ ASC | DESC | USING <replaceable class="parameter">operator</replaceable> ] [ NULLS { FIRST | LAST } ] [, ...]
</synopsis>
The <literal>ORDER BY</literal> clause causes the result rows to
be sorted according to the specified expression(s). If two rows are
equal according to the leftmost expression, they are compared
according to the next expression and so on. If they are equal
according to all specified expressions, they are returned in
an implementation-dependent order.
</para>
<para>
Each <replaceable class="parameter">expression</replaceable> can be the
name or ordinal number of an output column
(<command>SELECT</command> list item), or it can be an arbitrary
expression formed from input-column values.
</para>
<para>
The ordinal number refers to the ordinal (left-to-right) position
of the output column. This feature makes it possible to define an
ordering on the basis of a column that does not have a unique
name. This is never absolutely necessary because it is always
possible to assign a name to an output column using the
<literal>AS</> clause.
</para>
<para>
It is also possible to use arbitrary expressions in the
<literal>ORDER BY</literal> clause, including columns that do not
appear in the <command>SELECT</command> output list. Thus the
following statement is valid:
<programlisting>
SELECT name FROM distributors ORDER BY code;
</programlisting>
A limitation of this feature is that an <literal>ORDER BY</>
clause applying to the result of a <literal>UNION</>,
Update reference documentation on may/can/might:
Standard English uses "may", "can", and "might" in different ways:
may - permission, "You may borrow my rake."
can - ability, "I can lift that log."
might - possibility, "It might rain today."
Unfortunately, in conversational English, their use is often mixed, as
in, "You may use this variable to do X", when in fact, "can" is a better
choice. Similarly, "It may crash" is better stated, "It might crash".
19 years ago
<literal>INTERSECT</>, or <literal>EXCEPT</> clause can only
specify an output column name or number, not an expression.
</para>
<para>
If an <literal>ORDER BY</> expression is a simple name that
matches both an output column name and an input column name,
<literal>ORDER BY</> will interpret it as the output column name.
This is the opposite of the choice that <literal>GROUP BY</> will
make in the same situation. This inconsistency is made to be
compatible with the SQL standard.
</para>
<para>
Update reference documentation on may/can/might:
Standard English uses "may", "can", and "might" in different ways:
may - permission, "You may borrow my rake."
can - ability, "I can lift that log."
might - possibility, "It might rain today."
Unfortunately, in conversational English, their use is often mixed, as
in, "You may use this variable to do X", when in fact, "can" is a better
choice. Similarly, "It may crash" is better stated, "It might crash".
19 years ago
Optionally one can add the key word <literal>ASC</> (ascending) or
<literal>DESC</> (descending) after any expression in the
<literal>ORDER BY</> clause. If not specified, <literal>ASC</> is
assumed by default. Alternatively, a specific ordering operator
Update reference documentation on may/can/might:
Standard English uses "may", "can", and "might" in different ways:
may - permission, "You may borrow my rake."
can - ability, "I can lift that log."
might - possibility, "It might rain today."
Unfortunately, in conversational English, their use is often mixed, as
in, "You may use this variable to do X", when in fact, "can" is a better
choice. Similarly, "It may crash" is better stated, "It might crash".
19 years ago
name can be specified in the <literal>USING</> clause.
An ordering operator must be a less-than or greater-than
member of some B-tree operator family.
<literal>ASC</> is usually equivalent to <literal>USING <</> and
<literal>DESC</> is usually equivalent to <literal>USING ></>.
(But the creator of a user-defined data type can define exactly what the
default sort ordering is, and it might correspond to operators with other
names.)
</para>
<para>
If <literal>NULLS LAST</> is specified, null values sort after all
non-null values; if <literal>NULLS FIRST</> is specified, null values
sort before all non-null values. If neither is specified, the default
behavior is <literal>NULLS LAST</> when <literal>ASC</> is specified
or implied, and <literal>NULLS FIRST</> when <literal>DESC</> is specified
(thus, the default is to act as though nulls are larger than non-nulls).
When <literal>USING</> is specified, the default nulls ordering depends
on whether the operator is a less-than or greater-than operator.
</para>
<para>
Note that ordering options apply only to the expression they follow;
for example <literal>ORDER BY x, y DESC</> does not mean
the same thing as <literal>ORDER BY x DESC, y DESC</>.
</para>
<para>
Character-string data is sorted according to the collation that applies
to the column being sorted. That can be overridden at need by including
a <literal>COLLATE</> clause in the
<replaceable class="parameter">expression</replaceable>, for example
<literal>ORDER BY mycolumn COLLATE "en_US"</>.
For more information see <xref linkend="sql-syntax-collate-exprs"> and
<xref linkend="collation">.
</para>
</refsect2>
<refsect2 id="SQL-LIMIT">
<title id="sql-limit-title"><literal>LIMIT</literal> Clause</title>
<para>
The <literal>LIMIT</literal> clause consists of two independent
sub-clauses:
<synopsis>
LIMIT { <replaceable class="parameter">count</replaceable> | ALL }
OFFSET <replaceable class="parameter">start</replaceable>
</synopsis>
<replaceable class="parameter">count</replaceable> specifies the
maximum number of rows to return, while <replaceable
class="parameter">start</replaceable> specifies the number of rows
to skip before starting to return rows. When both are specified,
<replaceable class="parameter">start</replaceable> rows are skipped
before starting to count the <replaceable
class="parameter">count</replaceable> rows to be returned.
</para>
<para>
If the <replaceable class="parameter">count</replaceable> expression
evaluates to NULL, it is treated as <literal>LIMIT ALL</>, i.e., no
limit. If <replaceable class="parameter">start</replaceable> evaluates
to NULL, it is treated the same as <literal>OFFSET 0</>.
</para>
<para>
SQL:2008 introduced a different syntax to achieve the same result,
which <productname>PostgreSQL</> also supports. It is:
<synopsis>
OFFSET <replaceable class="parameter">start</replaceable> { ROW | ROWS }
FETCH { FIRST | NEXT } [ <replaceable class="parameter">count</replaceable> ] { ROW | ROWS } ONLY
</synopsis>
In this syntax, to write anything except a simple integer constant for
<replaceable class="parameter">start</> or <replaceable
class="parameter">count</replaceable>, you must write parentheses
around it.
If <replaceable class="parameter">count</> is
omitted in a <literal>FETCH</> clause, it defaults to 1.
<literal>ROW</literal>
and <literal>ROWS</literal> as well as <literal>FIRST</literal>
and <literal>NEXT</literal> are noise words that don't influence
the effects of these clauses.
According to the standard, the <literal>OFFSET</literal> clause must come
before the <literal>FETCH</literal> clause if both are present; but
<productname>PostgreSQL</> is laxer and allows either order.
</para>
<para>
When using <literal>LIMIT</>, it is a good idea to use an
<literal>ORDER BY</> clause that constrains the result rows into a
unique order. Otherwise you will get an unpredictable subset of
Update reference documentation on may/can/might:
Standard English uses "may", "can", and "might" in different ways:
may - permission, "You may borrow my rake."
can - ability, "I can lift that log."
might - possibility, "It might rain today."
Unfortunately, in conversational English, their use is often mixed, as
in, "You may use this variable to do X", when in fact, "can" is a better
choice. Similarly, "It may crash" is better stated, "It might crash".
19 years ago
the query's rows — you might be asking for the tenth through
twentieth rows, but tenth through twentieth in what ordering? You
don't know what ordering unless you specify <literal>ORDER BY</>.
</para>
<para>
The query planner takes <literal>LIMIT</> into account when
generating a query plan, so you are very likely to get different
plans (yielding different row orders) depending on what you use
for <literal>LIMIT</> and <literal>OFFSET</>. Thus, using
different <literal>LIMIT</>/<literal>OFFSET</> values to select
different subsets of a query result <emphasis>will give
inconsistent results</emphasis> unless you enforce a predictable
result ordering with <literal>ORDER BY</>. This is not a bug; it
is an inherent consequence of the fact that SQL does not promise
to deliver the results of a query in any particular order unless
<literal>ORDER BY</> is used to constrain the order.
</para>
<para>
It is even possible for repeated executions of the same <literal>LIMIT</>
query to return different subsets of the rows of a table, if there
is not an <literal>ORDER BY</> to enforce selection of a deterministic
subset. Again, this is not a bug; determinism of the results is
simply not guaranteed in such a case.
</para>
</refsect2>
<refsect2 id="SQL-FOR-UPDATE-SHARE">
<title id="sql-for-update-share-title">The Locking Clause</title>
Improve concurrency of foreign key locking
This patch introduces two additional lock modes for tuples: "SELECT FOR
KEY SHARE" and "SELECT FOR NO KEY UPDATE". These don't block each
other, in contrast with already existing "SELECT FOR SHARE" and "SELECT
FOR UPDATE". UPDATE commands that do not modify the values stored in
the columns that are part of the key of the tuple now grab a SELECT FOR
NO KEY UPDATE lock on the tuple, allowing them to proceed concurrently
with tuple locks of the FOR KEY SHARE variety.
Foreign key triggers now use FOR KEY SHARE instead of FOR SHARE; this
means the concurrency improvement applies to them, which is the whole
point of this patch.
The added tuple lock semantics require some rejiggering of the multixact
module, so that the locking level that each transaction is holding can
be stored alongside its Xid. Also, multixacts now need to persist
across server restarts and crashes, because they can now represent not
only tuple locks, but also tuple updates. This means we need more
careful tracking of lifetime of pg_multixact SLRU files; since they now
persist longer, we require more infrastructure to figure out when they
can be removed. pg_upgrade also needs to be careful to copy
pg_multixact files over from the old server to the new, or at least part
of multixact.c state, depending on the versions of the old and new
servers.
Tuple time qualification rules (HeapTupleSatisfies routines) need to be
careful not to consider tuples with the "is multi" infomask bit set as
being only locked; they might need to look up MultiXact values (i.e.
possibly do pg_multixact I/O) to find out the Xid that updated a tuple,
whereas they previously were assured to only use information readily
available from the tuple header. This is considered acceptable, because
the extra I/O would involve cases that would previously cause some
commands to block waiting for concurrent transactions to finish.
Another important change is the fact that locking tuples that have
previously been updated causes the future versions to be marked as
locked, too; this is essential for correctness of foreign key checks.
This causes additional WAL-logging, also (there was previously a single
WAL record for a locked tuple; now there are as many as updated copies
of the tuple there exist.)
With all this in place, contention related to tuples being checked by
foreign key rules should be much reduced.
As a bonus, the old behavior that a subtransaction grabbing a stronger
tuple lock than the parent (sub)transaction held on a given tuple and
later aborting caused the weaker lock to be lost, has been fixed.
Many new spec files were added for isolation tester framework, to ensure
overall behavior is sane. There's probably room for several more tests.
There were several reviewers of this patch; in particular, Noah Misch
and Andres Freund spent considerable time in it. Original idea for the
patch came from Simon Riggs, after a problem report by Joel Jacobson.
Most code is from me, with contributions from Marti Raudsepp, Alexander
Shulgin, Noah Misch and Andres Freund.
This patch was discussed in several pgsql-hackers threads; the most
important start at the following message-ids:
AANLkTimo9XVcEzfiBR-ut3KVNDkjm2Vxh+t8kAmWjPuv@mail.gmail.com
1290721684-sup-3951@alvh.no-ip.org
1294953201-sup-2099@alvh.no-ip.org
1320343602-sup-2290@alvh.no-ip.org
1339690386-sup-8927@alvh.no-ip.org
4FE5FF020200002500048A3D@gw.wicourts.gov
4FEAB90A0200002500048B7D@gw.wicourts.gov
13 years ago
<para>
<literal>FOR UPDATE</>, <literal>FOR NO KEY UPDATE</>, <literal>FOR SHARE</>
and <literal>FOR KEY SHARE</>
are <firstterm>locking clauses</>; they affect how <literal>SELECT</>
locks rows as they are obtained from the table.
</para>
<para>
The locking clause has the general form
Improve concurrency of foreign key locking
This patch introduces two additional lock modes for tuples: "SELECT FOR
KEY SHARE" and "SELECT FOR NO KEY UPDATE". These don't block each
other, in contrast with already existing "SELECT FOR SHARE" and "SELECT
FOR UPDATE". UPDATE commands that do not modify the values stored in
the columns that are part of the key of the tuple now grab a SELECT FOR
NO KEY UPDATE lock on the tuple, allowing them to proceed concurrently
with tuple locks of the FOR KEY SHARE variety.
Foreign key triggers now use FOR KEY SHARE instead of FOR SHARE; this
means the concurrency improvement applies to them, which is the whole
point of this patch.
The added tuple lock semantics require some rejiggering of the multixact
module, so that the locking level that each transaction is holding can
be stored alongside its Xid. Also, multixacts now need to persist
across server restarts and crashes, because they can now represent not
only tuple locks, but also tuple updates. This means we need more
careful tracking of lifetime of pg_multixact SLRU files; since they now
persist longer, we require more infrastructure to figure out when they
can be removed. pg_upgrade also needs to be careful to copy
pg_multixact files over from the old server to the new, or at least part
of multixact.c state, depending on the versions of the old and new
servers.
Tuple time qualification rules (HeapTupleSatisfies routines) need to be
careful not to consider tuples with the "is multi" infomask bit set as
being only locked; they might need to look up MultiXact values (i.e.
possibly do pg_multixact I/O) to find out the Xid that updated a tuple,
whereas they previously were assured to only use information readily
available from the tuple header. This is considered acceptable, because
the extra I/O would involve cases that would previously cause some
commands to block waiting for concurrent transactions to finish.
Another important change is the fact that locking tuples that have
previously been updated causes the future versions to be marked as
locked, too; this is essential for correctness of foreign key checks.
This causes additional WAL-logging, also (there was previously a single
WAL record for a locked tuple; now there are as many as updated copies
of the tuple there exist.)
With all this in place, contention related to tuples being checked by
foreign key rules should be much reduced.
As a bonus, the old behavior that a subtransaction grabbing a stronger
tuple lock than the parent (sub)transaction held on a given tuple and
later aborting caused the weaker lock to be lost, has been fixed.
Many new spec files were added for isolation tester framework, to ensure
overall behavior is sane. There's probably room for several more tests.
There were several reviewers of this patch; in particular, Noah Misch
and Andres Freund spent considerable time in it. Original idea for the
patch came from Simon Riggs, after a problem report by Joel Jacobson.
Most code is from me, with contributions from Marti Raudsepp, Alexander
Shulgin, Noah Misch and Andres Freund.
This patch was discussed in several pgsql-hackers threads; the most
important start at the following message-ids:
AANLkTimo9XVcEzfiBR-ut3KVNDkjm2Vxh+t8kAmWjPuv@mail.gmail.com
1290721684-sup-3951@alvh.no-ip.org
1294953201-sup-2099@alvh.no-ip.org
1320343602-sup-2290@alvh.no-ip.org
1339690386-sup-8927@alvh.no-ip.org
4FE5FF020200002500048A3D@gw.wicourts.gov
4FEAB90A0200002500048B7D@gw.wicourts.gov
13 years ago
<synopsis>
FOR <replaceable>lock_strength</> [ OF <replaceable class="parameter">table_name</replaceable> [, ...] ] [ NOWAIT ]
Improve concurrency of foreign key locking
This patch introduces two additional lock modes for tuples: "SELECT FOR
KEY SHARE" and "SELECT FOR NO KEY UPDATE". These don't block each
other, in contrast with already existing "SELECT FOR SHARE" and "SELECT
FOR UPDATE". UPDATE commands that do not modify the values stored in
the columns that are part of the key of the tuple now grab a SELECT FOR
NO KEY UPDATE lock on the tuple, allowing them to proceed concurrently
with tuple locks of the FOR KEY SHARE variety.
Foreign key triggers now use FOR KEY SHARE instead of FOR SHARE; this
means the concurrency improvement applies to them, which is the whole
point of this patch.
The added tuple lock semantics require some rejiggering of the multixact
module, so that the locking level that each transaction is holding can
be stored alongside its Xid. Also, multixacts now need to persist
across server restarts and crashes, because they can now represent not
only tuple locks, but also tuple updates. This means we need more
careful tracking of lifetime of pg_multixact SLRU files; since they now
persist longer, we require more infrastructure to figure out when they
can be removed. pg_upgrade also needs to be careful to copy
pg_multixact files over from the old server to the new, or at least part
of multixact.c state, depending on the versions of the old and new
servers.
Tuple time qualification rules (HeapTupleSatisfies routines) need to be
careful not to consider tuples with the "is multi" infomask bit set as
being only locked; they might need to look up MultiXact values (i.e.
possibly do pg_multixact I/O) to find out the Xid that updated a tuple,
whereas they previously were assured to only use information readily
available from the tuple header. This is considered acceptable, because
the extra I/O would involve cases that would previously cause some
commands to block waiting for concurrent transactions to finish.
Another important change is the fact that locking tuples that have
previously been updated causes the future versions to be marked as
locked, too; this is essential for correctness of foreign key checks.
This causes additional WAL-logging, also (there was previously a single
WAL record for a locked tuple; now there are as many as updated copies
of the tuple there exist.)
With all this in place, contention related to tuples being checked by
foreign key rules should be much reduced.
As a bonus, the old behavior that a subtransaction grabbing a stronger
tuple lock than the parent (sub)transaction held on a given tuple and
later aborting caused the weaker lock to be lost, has been fixed.
Many new spec files were added for isolation tester framework, to ensure
overall behavior is sane. There's probably room for several more tests.
There were several reviewers of this patch; in particular, Noah Misch
and Andres Freund spent considerable time in it. Original idea for the
patch came from Simon Riggs, after a problem report by Joel Jacobson.
Most code is from me, with contributions from Marti Raudsepp, Alexander
Shulgin, Noah Misch and Andres Freund.
This patch was discussed in several pgsql-hackers threads; the most
important start at the following message-ids:
AANLkTimo9XVcEzfiBR-ut3KVNDkjm2Vxh+t8kAmWjPuv@mail.gmail.com
1290721684-sup-3951@alvh.no-ip.org
1294953201-sup-2099@alvh.no-ip.org
1320343602-sup-2290@alvh.no-ip.org
1339690386-sup-8927@alvh.no-ip.org
4FE5FF020200002500048A3D@gw.wicourts.gov
4FEAB90A0200002500048B7D@gw.wicourts.gov
13 years ago
</synopsis>
where <replaceable>lock_strength</> can be one of
Improve concurrency of foreign key locking
This patch introduces two additional lock modes for tuples: "SELECT FOR
KEY SHARE" and "SELECT FOR NO KEY UPDATE". These don't block each
other, in contrast with already existing "SELECT FOR SHARE" and "SELECT
FOR UPDATE". UPDATE commands that do not modify the values stored in
the columns that are part of the key of the tuple now grab a SELECT FOR
NO KEY UPDATE lock on the tuple, allowing them to proceed concurrently
with tuple locks of the FOR KEY SHARE variety.
Foreign key triggers now use FOR KEY SHARE instead of FOR SHARE; this
means the concurrency improvement applies to them, which is the whole
point of this patch.
The added tuple lock semantics require some rejiggering of the multixact
module, so that the locking level that each transaction is holding can
be stored alongside its Xid. Also, multixacts now need to persist
across server restarts and crashes, because they can now represent not
only tuple locks, but also tuple updates. This means we need more
careful tracking of lifetime of pg_multixact SLRU files; since they now
persist longer, we require more infrastructure to figure out when they
can be removed. pg_upgrade also needs to be careful to copy
pg_multixact files over from the old server to the new, or at least part
of multixact.c state, depending on the versions of the old and new
servers.
Tuple time qualification rules (HeapTupleSatisfies routines) need to be
careful not to consider tuples with the "is multi" infomask bit set as
being only locked; they might need to look up MultiXact values (i.e.
possibly do pg_multixact I/O) to find out the Xid that updated a tuple,
whereas they previously were assured to only use information readily
available from the tuple header. This is considered acceptable, because
the extra I/O would involve cases that would previously cause some
commands to block waiting for concurrent transactions to finish.
Another important change is the fact that locking tuples that have
previously been updated causes the future versions to be marked as
locked, too; this is essential for correctness of foreign key checks.
This causes additional WAL-logging, also (there was previously a single
WAL record for a locked tuple; now there are as many as updated copies
of the tuple there exist.)
With all this in place, contention related to tuples being checked by
foreign key rules should be much reduced.
As a bonus, the old behavior that a subtransaction grabbing a stronger
tuple lock than the parent (sub)transaction held on a given tuple and
later aborting caused the weaker lock to be lost, has been fixed.
Many new spec files were added for isolation tester framework, to ensure
overall behavior is sane. There's probably room for several more tests.
There were several reviewers of this patch; in particular, Noah Misch
and Andres Freund spent considerable time in it. Original idea for the
patch came from Simon Riggs, after a problem report by Joel Jacobson.
Most code is from me, with contributions from Marti Raudsepp, Alexander
Shulgin, Noah Misch and Andres Freund.
This patch was discussed in several pgsql-hackers threads; the most
important start at the following message-ids:
AANLkTimo9XVcEzfiBR-ut3KVNDkjm2Vxh+t8kAmWjPuv@mail.gmail.com
1290721684-sup-3951@alvh.no-ip.org
1294953201-sup-2099@alvh.no-ip.org
1320343602-sup-2290@alvh.no-ip.org
1339690386-sup-8927@alvh.no-ip.org
4FE5FF020200002500048A3D@gw.wicourts.gov
4FEAB90A0200002500048B7D@gw.wicourts.gov
13 years ago
<synopsis>
UPDATE
NO KEY UPDATE
SHARE
KEY SHARE
Improve concurrency of foreign key locking
This patch introduces two additional lock modes for tuples: "SELECT FOR
KEY SHARE" and "SELECT FOR NO KEY UPDATE". These don't block each
other, in contrast with already existing "SELECT FOR SHARE" and "SELECT
FOR UPDATE". UPDATE commands that do not modify the values stored in
the columns that are part of the key of the tuple now grab a SELECT FOR
NO KEY UPDATE lock on the tuple, allowing them to proceed concurrently
with tuple locks of the FOR KEY SHARE variety.
Foreign key triggers now use FOR KEY SHARE instead of FOR SHARE; this
means the concurrency improvement applies to them, which is the whole
point of this patch.
The added tuple lock semantics require some rejiggering of the multixact
module, so that the locking level that each transaction is holding can
be stored alongside its Xid. Also, multixacts now need to persist
across server restarts and crashes, because they can now represent not
only tuple locks, but also tuple updates. This means we need more
careful tracking of lifetime of pg_multixact SLRU files; since they now
persist longer, we require more infrastructure to figure out when they
can be removed. pg_upgrade also needs to be careful to copy
pg_multixact files over from the old server to the new, or at least part
of multixact.c state, depending on the versions of the old and new
servers.
Tuple time qualification rules (HeapTupleSatisfies routines) need to be
careful not to consider tuples with the "is multi" infomask bit set as
being only locked; they might need to look up MultiXact values (i.e.
possibly do pg_multixact I/O) to find out the Xid that updated a tuple,
whereas they previously were assured to only use information readily
available from the tuple header. This is considered acceptable, because
the extra I/O would involve cases that would previously cause some
commands to block waiting for concurrent transactions to finish.
Another important change is the fact that locking tuples that have
previously been updated causes the future versions to be marked as
locked, too; this is essential for correctness of foreign key checks.
This causes additional WAL-logging, also (there was previously a single
WAL record for a locked tuple; now there are as many as updated copies
of the tuple there exist.)
With all this in place, contention related to tuples being checked by
foreign key rules should be much reduced.
As a bonus, the old behavior that a subtransaction grabbing a stronger
tuple lock than the parent (sub)transaction held on a given tuple and
later aborting caused the weaker lock to be lost, has been fixed.
Many new spec files were added for isolation tester framework, to ensure
overall behavior is sane. There's probably room for several more tests.
There were several reviewers of this patch; in particular, Noah Misch
and Andres Freund spent considerable time in it. Original idea for the
patch came from Simon Riggs, after a problem report by Joel Jacobson.
Most code is from me, with contributions from Marti Raudsepp, Alexander
Shulgin, Noah Misch and Andres Freund.
This patch was discussed in several pgsql-hackers threads; the most
important start at the following message-ids:
AANLkTimo9XVcEzfiBR-ut3KVNDkjm2Vxh+t8kAmWjPuv@mail.gmail.com
1290721684-sup-3951@alvh.no-ip.org
1294953201-sup-2099@alvh.no-ip.org
1320343602-sup-2290@alvh.no-ip.org
1339690386-sup-8927@alvh.no-ip.org
4FE5FF020200002500048A3D@gw.wicourts.gov
4FEAB90A0200002500048B7D@gw.wicourts.gov
13 years ago
</synopsis>
</para>
<para>
<literal>FOR UPDATE</literal> causes the rows retrieved by the
<command>SELECT</command> statement to be locked as though for
update. This prevents them from being modified or deleted by
other transactions until the current transaction ends. That is,
other transactions that attempt <command>UPDATE</command>,
Improve concurrency of foreign key locking
This patch introduces two additional lock modes for tuples: "SELECT FOR
KEY SHARE" and "SELECT FOR NO KEY UPDATE". These don't block each
other, in contrast with already existing "SELECT FOR SHARE" and "SELECT
FOR UPDATE". UPDATE commands that do not modify the values stored in
the columns that are part of the key of the tuple now grab a SELECT FOR
NO KEY UPDATE lock on the tuple, allowing them to proceed concurrently
with tuple locks of the FOR KEY SHARE variety.
Foreign key triggers now use FOR KEY SHARE instead of FOR SHARE; this
means the concurrency improvement applies to them, which is the whole
point of this patch.
The added tuple lock semantics require some rejiggering of the multixact
module, so that the locking level that each transaction is holding can
be stored alongside its Xid. Also, multixacts now need to persist
across server restarts and crashes, because they can now represent not
only tuple locks, but also tuple updates. This means we need more
careful tracking of lifetime of pg_multixact SLRU files; since they now
persist longer, we require more infrastructure to figure out when they
can be removed. pg_upgrade also needs to be careful to copy
pg_multixact files over from the old server to the new, or at least part
of multixact.c state, depending on the versions of the old and new
servers.
Tuple time qualification rules (HeapTupleSatisfies routines) need to be
careful not to consider tuples with the "is multi" infomask bit set as
being only locked; they might need to look up MultiXact values (i.e.
possibly do pg_multixact I/O) to find out the Xid that updated a tuple,
whereas they previously were assured to only use information readily
available from the tuple header. This is considered acceptable, because
the extra I/O would involve cases that would previously cause some
commands to block waiting for concurrent transactions to finish.
Another important change is the fact that locking tuples that have
previously been updated causes the future versions to be marked as
locked, too; this is essential for correctness of foreign key checks.
This causes additional WAL-logging, also (there was previously a single
WAL record for a locked tuple; now there are as many as updated copies
of the tuple there exist.)
With all this in place, contention related to tuples being checked by
foreign key rules should be much reduced.
As a bonus, the old behavior that a subtransaction grabbing a stronger
tuple lock than the parent (sub)transaction held on a given tuple and
later aborting caused the weaker lock to be lost, has been fixed.
Many new spec files were added for isolation tester framework, to ensure
overall behavior is sane. There's probably room for several more tests.
There were several reviewers of this patch; in particular, Noah Misch
and Andres Freund spent considerable time in it. Original idea for the
patch came from Simon Riggs, after a problem report by Joel Jacobson.
Most code is from me, with contributions from Marti Raudsepp, Alexander
Shulgin, Noah Misch and Andres Freund.
This patch was discussed in several pgsql-hackers threads; the most
important start at the following message-ids:
AANLkTimo9XVcEzfiBR-ut3KVNDkjm2Vxh+t8kAmWjPuv@mail.gmail.com
1290721684-sup-3951@alvh.no-ip.org
1294953201-sup-2099@alvh.no-ip.org
1320343602-sup-2290@alvh.no-ip.org
1339690386-sup-8927@alvh.no-ip.org
4FE5FF020200002500048A3D@gw.wicourts.gov
4FEAB90A0200002500048B7D@gw.wicourts.gov
13 years ago
<command>DELETE</command>,
<command>SELECT FOR UPDATE</command>,
<command>SELECT FOR NO KEY UPDATE</command>,
Improve concurrency of foreign key locking
This patch introduces two additional lock modes for tuples: "SELECT FOR
KEY SHARE" and "SELECT FOR NO KEY UPDATE". These don't block each
other, in contrast with already existing "SELECT FOR SHARE" and "SELECT
FOR UPDATE". UPDATE commands that do not modify the values stored in
the columns that are part of the key of the tuple now grab a SELECT FOR
NO KEY UPDATE lock on the tuple, allowing them to proceed concurrently
with tuple locks of the FOR KEY SHARE variety.
Foreign key triggers now use FOR KEY SHARE instead of FOR SHARE; this
means the concurrency improvement applies to them, which is the whole
point of this patch.
The added tuple lock semantics require some rejiggering of the multixact
module, so that the locking level that each transaction is holding can
be stored alongside its Xid. Also, multixacts now need to persist
across server restarts and crashes, because they can now represent not
only tuple locks, but also tuple updates. This means we need more
careful tracking of lifetime of pg_multixact SLRU files; since they now
persist longer, we require more infrastructure to figure out when they
can be removed. pg_upgrade also needs to be careful to copy
pg_multixact files over from the old server to the new, or at least part
of multixact.c state, depending on the versions of the old and new
servers.
Tuple time qualification rules (HeapTupleSatisfies routines) need to be
careful not to consider tuples with the "is multi" infomask bit set as
being only locked; they might need to look up MultiXact values (i.e.
possibly do pg_multixact I/O) to find out the Xid that updated a tuple,
whereas they previously were assured to only use information readily
available from the tuple header. This is considered acceptable, because
the extra I/O would involve cases that would previously cause some
commands to block waiting for concurrent transactions to finish.
Another important change is the fact that locking tuples that have
previously been updated causes the future versions to be marked as
locked, too; this is essential for correctness of foreign key checks.
This causes additional WAL-logging, also (there was previously a single
WAL record for a locked tuple; now there are as many as updated copies
of the tuple there exist.)
With all this in place, contention related to tuples being checked by
foreign key rules should be much reduced.
As a bonus, the old behavior that a subtransaction grabbing a stronger
tuple lock than the parent (sub)transaction held on a given tuple and
later aborting caused the weaker lock to be lost, has been fixed.
Many new spec files were added for isolation tester framework, to ensure
overall behavior is sane. There's probably room for several more tests.
There were several reviewers of this patch; in particular, Noah Misch
and Andres Freund spent considerable time in it. Original idea for the
patch came from Simon Riggs, after a problem report by Joel Jacobson.
Most code is from me, with contributions from Marti Raudsepp, Alexander
Shulgin, Noah Misch and Andres Freund.
This patch was discussed in several pgsql-hackers threads; the most
important start at the following message-ids:
AANLkTimo9XVcEzfiBR-ut3KVNDkjm2Vxh+t8kAmWjPuv@mail.gmail.com
1290721684-sup-3951@alvh.no-ip.org
1294953201-sup-2099@alvh.no-ip.org
1320343602-sup-2290@alvh.no-ip.org
1339690386-sup-8927@alvh.no-ip.org
4FE5FF020200002500048A3D@gw.wicourts.gov
4FEAB90A0200002500048B7D@gw.wicourts.gov
13 years ago
<command>SELECT FOR SHARE</command> or
<command>SELECT FOR KEY SHARE</command>
of these rows will be blocked until the current transaction ends.
Improve concurrency of foreign key locking
This patch introduces two additional lock modes for tuples: "SELECT FOR
KEY SHARE" and "SELECT FOR NO KEY UPDATE". These don't block each
other, in contrast with already existing "SELECT FOR SHARE" and "SELECT
FOR UPDATE". UPDATE commands that do not modify the values stored in
the columns that are part of the key of the tuple now grab a SELECT FOR
NO KEY UPDATE lock on the tuple, allowing them to proceed concurrently
with tuple locks of the FOR KEY SHARE variety.
Foreign key triggers now use FOR KEY SHARE instead of FOR SHARE; this
means the concurrency improvement applies to them, which is the whole
point of this patch.
The added tuple lock semantics require some rejiggering of the multixact
module, so that the locking level that each transaction is holding can
be stored alongside its Xid. Also, multixacts now need to persist
across server restarts and crashes, because they can now represent not
only tuple locks, but also tuple updates. This means we need more
careful tracking of lifetime of pg_multixact SLRU files; since they now
persist longer, we require more infrastructure to figure out when they
can be removed. pg_upgrade also needs to be careful to copy
pg_multixact files over from the old server to the new, or at least part
of multixact.c state, depending on the versions of the old and new
servers.
Tuple time qualification rules (HeapTupleSatisfies routines) need to be
careful not to consider tuples with the "is multi" infomask bit set as
being only locked; they might need to look up MultiXact values (i.e.
possibly do pg_multixact I/O) to find out the Xid that updated a tuple,
whereas they previously were assured to only use information readily
available from the tuple header. This is considered acceptable, because
the extra I/O would involve cases that would previously cause some
commands to block waiting for concurrent transactions to finish.
Another important change is the fact that locking tuples that have
previously been updated causes the future versions to be marked as
locked, too; this is essential for correctness of foreign key checks.
This causes additional WAL-logging, also (there was previously a single
WAL record for a locked tuple; now there are as many as updated copies
of the tuple there exist.)
With all this in place, contention related to tuples being checked by
foreign key rules should be much reduced.
As a bonus, the old behavior that a subtransaction grabbing a stronger
tuple lock than the parent (sub)transaction held on a given tuple and
later aborting caused the weaker lock to be lost, has been fixed.
Many new spec files were added for isolation tester framework, to ensure
overall behavior is sane. There's probably room for several more tests.
There were several reviewers of this patch; in particular, Noah Misch
and Andres Freund spent considerable time in it. Original idea for the
patch came from Simon Riggs, after a problem report by Joel Jacobson.
Most code is from me, with contributions from Marti Raudsepp, Alexander
Shulgin, Noah Misch and Andres Freund.
This patch was discussed in several pgsql-hackers threads; the most
important start at the following message-ids:
AANLkTimo9XVcEzfiBR-ut3KVNDkjm2Vxh+t8kAmWjPuv@mail.gmail.com
1290721684-sup-3951@alvh.no-ip.org
1294953201-sup-2099@alvh.no-ip.org
1320343602-sup-2290@alvh.no-ip.org
1339690386-sup-8927@alvh.no-ip.org
4FE5FF020200002500048A3D@gw.wicourts.gov
4FEAB90A0200002500048B7D@gw.wicourts.gov
13 years ago
The <literal>FOR UPDATE</> lock mode
is also acquired by any <command>DELETE</> on a row, and also by an
<command>UPDATE</> that modifies the values on certain columns. Currently,
the set of columns considered for the <command>UPDATE</> case are those that
have an unique index on them that can be used in a foreign key (so partial
indexes and expressional indexes are not considered), but this may change
in the future.
Also, if an <command>UPDATE</command>, <command>DELETE</command>,
or <command>SELECT FOR UPDATE</command> from another transaction
has already locked a selected row or rows, <command>SELECT FOR
UPDATE</command> will wait for the other transaction to complete,
and will then lock and return the updated row (or no row, if the
Implement genuine serializable isolation level.
Until now, our Serializable mode has in fact been what's called Snapshot
Isolation, which allows some anomalies that could not occur in any
serialized ordering of the transactions. This patch fixes that using a
method called Serializable Snapshot Isolation, based on research papers by
Michael J. Cahill (see README-SSI for full references). In Serializable
Snapshot Isolation, transactions run like they do in Snapshot Isolation,
but a predicate lock manager observes the reads and writes performed and
aborts transactions if it detects that an anomaly might occur. This method
produces some false positives, ie. it sometimes aborts transactions even
though there is no anomaly.
To track reads we implement predicate locking, see storage/lmgr/predicate.c.
Whenever a tuple is read, a predicate lock is acquired on the tuple. Shared
memory is finite, so when a transaction takes many tuple-level locks on a
page, the locks are promoted to a single page-level lock, and further to a
single relation level lock if necessary. To lock key values with no matching
tuple, a sequential scan always takes a relation-level lock, and an index
scan acquires a page-level lock that covers the search key, whether or not
there are any matching keys at the moment.
A predicate lock doesn't conflict with any regular locks or with another
predicate locks in the normal sense. They're only used by the predicate lock
manager to detect the danger of anomalies. Only serializable transactions
participate in predicate locking, so there should be no extra overhead for
for other transactions.
Predicate locks can't be released at commit, but must be remembered until
all the transactions that overlapped with it have completed. That means that
we need to remember an unbounded amount of predicate locks, so we apply a
lossy but conservative method of tracking locks for committed transactions.
If we run short of shared memory, we overflow to a new "pg_serial" SLRU
pool.
We don't currently allow Serializable transactions in Hot Standby mode.
That would be hard, because even read-only transactions can cause anomalies
that wouldn't otherwise occur.
Serializable isolation mode now means the new fully serializable level.
Repeatable Read gives you the old Snapshot Isolation level that we have
always had.
Kevin Grittner and Dan Ports, reviewed by Jeff Davis, Heikki Linnakangas and
Anssi Kääriäinen
15 years ago
row was deleted). Within a <literal>REPEATABLE READ</> or <literal>SERIALIZABLE</> transaction,
however, an error will be thrown if a row to be locked has changed
since the transaction started. For further discussion see <xref
linkend="mvcc">.
</para>
Improve concurrency of foreign key locking
This patch introduces two additional lock modes for tuples: "SELECT FOR
KEY SHARE" and "SELECT FOR NO KEY UPDATE". These don't block each
other, in contrast with already existing "SELECT FOR SHARE" and "SELECT
FOR UPDATE". UPDATE commands that do not modify the values stored in
the columns that are part of the key of the tuple now grab a SELECT FOR
NO KEY UPDATE lock on the tuple, allowing them to proceed concurrently
with tuple locks of the FOR KEY SHARE variety.
Foreign key triggers now use FOR KEY SHARE instead of FOR SHARE; this
means the concurrency improvement applies to them, which is the whole
point of this patch.
The added tuple lock semantics require some rejiggering of the multixact
module, so that the locking level that each transaction is holding can
be stored alongside its Xid. Also, multixacts now need to persist
across server restarts and crashes, because they can now represent not
only tuple locks, but also tuple updates. This means we need more
careful tracking of lifetime of pg_multixact SLRU files; since they now
persist longer, we require more infrastructure to figure out when they
can be removed. pg_upgrade also needs to be careful to copy
pg_multixact files over from the old server to the new, or at least part
of multixact.c state, depending on the versions of the old and new
servers.
Tuple time qualification rules (HeapTupleSatisfies routines) need to be
careful not to consider tuples with the "is multi" infomask bit set as
being only locked; they might need to look up MultiXact values (i.e.
possibly do pg_multixact I/O) to find out the Xid that updated a tuple,
whereas they previously were assured to only use information readily
available from the tuple header. This is considered acceptable, because
the extra I/O would involve cases that would previously cause some
commands to block waiting for concurrent transactions to finish.
Another important change is the fact that locking tuples that have
previously been updated causes the future versions to be marked as
locked, too; this is essential for correctness of foreign key checks.
This causes additional WAL-logging, also (there was previously a single
WAL record for a locked tuple; now there are as many as updated copies
of the tuple there exist.)
With all this in place, contention related to tuples being checked by
foreign key rules should be much reduced.
As a bonus, the old behavior that a subtransaction grabbing a stronger
tuple lock than the parent (sub)transaction held on a given tuple and
later aborting caused the weaker lock to be lost, has been fixed.
Many new spec files were added for isolation tester framework, to ensure
overall behavior is sane. There's probably room for several more tests.
There were several reviewers of this patch; in particular, Noah Misch
and Andres Freund spent considerable time in it. Original idea for the
patch came from Simon Riggs, after a problem report by Joel Jacobson.
Most code is from me, with contributions from Marti Raudsepp, Alexander
Shulgin, Noah Misch and Andres Freund.
This patch was discussed in several pgsql-hackers threads; the most
important start at the following message-ids:
AANLkTimo9XVcEzfiBR-ut3KVNDkjm2Vxh+t8kAmWjPuv@mail.gmail.com
1290721684-sup-3951@alvh.no-ip.org
1294953201-sup-2099@alvh.no-ip.org
1320343602-sup-2290@alvh.no-ip.org
1339690386-sup-8927@alvh.no-ip.org
4FE5FF020200002500048A3D@gw.wicourts.gov
4FEAB90A0200002500048B7D@gw.wicourts.gov
13 years ago
<para>
<literal>FOR NO KEY UPDATE</> behaves similarly, except that the lock
acquired is weaker: this lock will not block
<literal>SELECT FOR KEY SHARE</> commands that attempt to acquire
a lock on the same rows. This lock mode is also acquired by any
<command>UPDATE</> that does not acquire a <literal>FOR UPDATE</> lock.
Improve concurrency of foreign key locking
This patch introduces two additional lock modes for tuples: "SELECT FOR
KEY SHARE" and "SELECT FOR NO KEY UPDATE". These don't block each
other, in contrast with already existing "SELECT FOR SHARE" and "SELECT
FOR UPDATE". UPDATE commands that do not modify the values stored in
the columns that are part of the key of the tuple now grab a SELECT FOR
NO KEY UPDATE lock on the tuple, allowing them to proceed concurrently
with tuple locks of the FOR KEY SHARE variety.
Foreign key triggers now use FOR KEY SHARE instead of FOR SHARE; this
means the concurrency improvement applies to them, which is the whole
point of this patch.
The added tuple lock semantics require some rejiggering of the multixact
module, so that the locking level that each transaction is holding can
be stored alongside its Xid. Also, multixacts now need to persist
across server restarts and crashes, because they can now represent not
only tuple locks, but also tuple updates. This means we need more
careful tracking of lifetime of pg_multixact SLRU files; since they now
persist longer, we require more infrastructure to figure out when they
can be removed. pg_upgrade also needs to be careful to copy
pg_multixact files over from the old server to the new, or at least part
of multixact.c state, depending on the versions of the old and new
servers.
Tuple time qualification rules (HeapTupleSatisfies routines) need to be
careful not to consider tuples with the "is multi" infomask bit set as
being only locked; they might need to look up MultiXact values (i.e.
possibly do pg_multixact I/O) to find out the Xid that updated a tuple,
whereas they previously were assured to only use information readily
available from the tuple header. This is considered acceptable, because
the extra I/O would involve cases that would previously cause some
commands to block waiting for concurrent transactions to finish.
Another important change is the fact that locking tuples that have
previously been updated causes the future versions to be marked as
locked, too; this is essential for correctness of foreign key checks.
This causes additional WAL-logging, also (there was previously a single
WAL record for a locked tuple; now there are as many as updated copies
of the tuple there exist.)
With all this in place, contention related to tuples being checked by
foreign key rules should be much reduced.
As a bonus, the old behavior that a subtransaction grabbing a stronger
tuple lock than the parent (sub)transaction held on a given tuple and
later aborting caused the weaker lock to be lost, has been fixed.
Many new spec files were added for isolation tester framework, to ensure
overall behavior is sane. There's probably room for several more tests.
There were several reviewers of this patch; in particular, Noah Misch
and Andres Freund spent considerable time in it. Original idea for the
patch came from Simon Riggs, after a problem report by Joel Jacobson.
Most code is from me, with contributions from Marti Raudsepp, Alexander
Shulgin, Noah Misch and Andres Freund.
This patch was discussed in several pgsql-hackers threads; the most
important start at the following message-ids:
AANLkTimo9XVcEzfiBR-ut3KVNDkjm2Vxh+t8kAmWjPuv@mail.gmail.com
1290721684-sup-3951@alvh.no-ip.org
1294953201-sup-2099@alvh.no-ip.org
1320343602-sup-2290@alvh.no-ip.org
1339690386-sup-8927@alvh.no-ip.org
4FE5FF020200002500048A3D@gw.wicourts.gov
4FEAB90A0200002500048B7D@gw.wicourts.gov
13 years ago
</para>
<para>
<literal>FOR SHARE</literal> behaves similarly, except that it
acquires a shared rather than exclusive lock on each retrieved
row. A shared lock blocks other transactions from performing
<command>UPDATE</command>, <command>DELETE</command>, <command>SELECT
FOR UPDATE</command> or <command>SELECT FOR NO KEY UPDATE</>
on these rows, but it does not prevent them
Improve concurrency of foreign key locking
This patch introduces two additional lock modes for tuples: "SELECT FOR
KEY SHARE" and "SELECT FOR NO KEY UPDATE". These don't block each
other, in contrast with already existing "SELECT FOR SHARE" and "SELECT
FOR UPDATE". UPDATE commands that do not modify the values stored in
the columns that are part of the key of the tuple now grab a SELECT FOR
NO KEY UPDATE lock on the tuple, allowing them to proceed concurrently
with tuple locks of the FOR KEY SHARE variety.
Foreign key triggers now use FOR KEY SHARE instead of FOR SHARE; this
means the concurrency improvement applies to them, which is the whole
point of this patch.
The added tuple lock semantics require some rejiggering of the multixact
module, so that the locking level that each transaction is holding can
be stored alongside its Xid. Also, multixacts now need to persist
across server restarts and crashes, because they can now represent not
only tuple locks, but also tuple updates. This means we need more
careful tracking of lifetime of pg_multixact SLRU files; since they now
persist longer, we require more infrastructure to figure out when they
can be removed. pg_upgrade also needs to be careful to copy
pg_multixact files over from the old server to the new, or at least part
of multixact.c state, depending on the versions of the old and new
servers.
Tuple time qualification rules (HeapTupleSatisfies routines) need to be
careful not to consider tuples with the "is multi" infomask bit set as
being only locked; they might need to look up MultiXact values (i.e.
possibly do pg_multixact I/O) to find out the Xid that updated a tuple,
whereas they previously were assured to only use information readily
available from the tuple header. This is considered acceptable, because
the extra I/O would involve cases that would previously cause some
commands to block waiting for concurrent transactions to finish.
Another important change is the fact that locking tuples that have
previously been updated causes the future versions to be marked as
locked, too; this is essential for correctness of foreign key checks.
This causes additional WAL-logging, also (there was previously a single
WAL record for a locked tuple; now there are as many as updated copies
of the tuple there exist.)
With all this in place, contention related to tuples being checked by
foreign key rules should be much reduced.
As a bonus, the old behavior that a subtransaction grabbing a stronger
tuple lock than the parent (sub)transaction held on a given tuple and
later aborting caused the weaker lock to be lost, has been fixed.
Many new spec files were added for isolation tester framework, to ensure
overall behavior is sane. There's probably room for several more tests.
There were several reviewers of this patch; in particular, Noah Misch
and Andres Freund spent considerable time in it. Original idea for the
patch came from Simon Riggs, after a problem report by Joel Jacobson.
Most code is from me, with contributions from Marti Raudsepp, Alexander
Shulgin, Noah Misch and Andres Freund.
This patch was discussed in several pgsql-hackers threads; the most
important start at the following message-ids:
AANLkTimo9XVcEzfiBR-ut3KVNDkjm2Vxh+t8kAmWjPuv@mail.gmail.com
1290721684-sup-3951@alvh.no-ip.org
1294953201-sup-2099@alvh.no-ip.org
1320343602-sup-2290@alvh.no-ip.org
1339690386-sup-8927@alvh.no-ip.org
4FE5FF020200002500048A3D@gw.wicourts.gov
4FEAB90A0200002500048B7D@gw.wicourts.gov
13 years ago
from performing <command>SELECT FOR SHARE</command> or
<command>SELECT FOR KEY SHARE</command>.
</para>
<para>
<literal>FOR KEY SHARE</> behaves similarly to <literal>FOR SHARE</literal>,
except that the lock
is weaker: <literal>SELECT FOR UPDATE</> is blocked, but
not <literal>SELECT FOR NO KEY UPDATE</>. A key-shared
lock blocks other transactions from performing <command>DELETE</command>
or any <command>UPDATE</command> that changes the key values, but not
other <command>UPDATE</>, and neither does it prevent
<command>SELECT FOR NO KEY UPDATE</>, <command>SELECT FOR SHARE</>, or
Improve concurrency of foreign key locking
This patch introduces two additional lock modes for tuples: "SELECT FOR
KEY SHARE" and "SELECT FOR NO KEY UPDATE". These don't block each
other, in contrast with already existing "SELECT FOR SHARE" and "SELECT
FOR UPDATE". UPDATE commands that do not modify the values stored in
the columns that are part of the key of the tuple now grab a SELECT FOR
NO KEY UPDATE lock on the tuple, allowing them to proceed concurrently
with tuple locks of the FOR KEY SHARE variety.
Foreign key triggers now use FOR KEY SHARE instead of FOR SHARE; this
means the concurrency improvement applies to them, which is the whole
point of this patch.
The added tuple lock semantics require some rejiggering of the multixact
module, so that the locking level that each transaction is holding can
be stored alongside its Xid. Also, multixacts now need to persist
across server restarts and crashes, because they can now represent not
only tuple locks, but also tuple updates. This means we need more
careful tracking of lifetime of pg_multixact SLRU files; since they now
persist longer, we require more infrastructure to figure out when they
can be removed. pg_upgrade also needs to be careful to copy
pg_multixact files over from the old server to the new, or at least part
of multixact.c state, depending on the versions of the old and new
servers.
Tuple time qualification rules (HeapTupleSatisfies routines) need to be
careful not to consider tuples with the "is multi" infomask bit set as
being only locked; they might need to look up MultiXact values (i.e.
possibly do pg_multixact I/O) to find out the Xid that updated a tuple,
whereas they previously were assured to only use information readily
available from the tuple header. This is considered acceptable, because
the extra I/O would involve cases that would previously cause some
commands to block waiting for concurrent transactions to finish.
Another important change is the fact that locking tuples that have
previously been updated causes the future versions to be marked as
locked, too; this is essential for correctness of foreign key checks.
This causes additional WAL-logging, also (there was previously a single
WAL record for a locked tuple; now there are as many as updated copies
of the tuple there exist.)
With all this in place, contention related to tuples being checked by
foreign key rules should be much reduced.
As a bonus, the old behavior that a subtransaction grabbing a stronger
tuple lock than the parent (sub)transaction held on a given tuple and
later aborting caused the weaker lock to be lost, has been fixed.
Many new spec files were added for isolation tester framework, to ensure
overall behavior is sane. There's probably room for several more tests.
There were several reviewers of this patch; in particular, Noah Misch
and Andres Freund spent considerable time in it. Original idea for the
patch came from Simon Riggs, after a problem report by Joel Jacobson.
Most code is from me, with contributions from Marti Raudsepp, Alexander
Shulgin, Noah Misch and Andres Freund.
This patch was discussed in several pgsql-hackers threads; the most
important start at the following message-ids:
AANLkTimo9XVcEzfiBR-ut3KVNDkjm2Vxh+t8kAmWjPuv@mail.gmail.com
1290721684-sup-3951@alvh.no-ip.org
1294953201-sup-2099@alvh.no-ip.org
1320343602-sup-2290@alvh.no-ip.org
1339690386-sup-8927@alvh.no-ip.org
4FE5FF020200002500048A3D@gw.wicourts.gov
4FEAB90A0200002500048B7D@gw.wicourts.gov
13 years ago
<command>SELECT FOR KEY SHARE</>.
</para>
<para>
To prevent the operation from waiting for other transactions to commit,
use the <literal>NOWAIT</> option. With <literal>NOWAIT</>, the statement
reports an error, rather than waiting, if a selected row
cannot be locked immediately. Note that <literal>NOWAIT</> applies only
to the row-level lock(s) — the required <literal>ROW SHARE</literal>
table-level lock is still taken in the ordinary way (see
<xref linkend="mvcc">). You can use
<xref linkend="sql-lock">
with the <literal>NOWAIT</> option first,
if you need to acquire the table-level lock without waiting.
</para>
<para>
Improve concurrency of foreign key locking
This patch introduces two additional lock modes for tuples: "SELECT FOR
KEY SHARE" and "SELECT FOR NO KEY UPDATE". These don't block each
other, in contrast with already existing "SELECT FOR SHARE" and "SELECT
FOR UPDATE". UPDATE commands that do not modify the values stored in
the columns that are part of the key of the tuple now grab a SELECT FOR
NO KEY UPDATE lock on the tuple, allowing them to proceed concurrently
with tuple locks of the FOR KEY SHARE variety.
Foreign key triggers now use FOR KEY SHARE instead of FOR SHARE; this
means the concurrency improvement applies to them, which is the whole
point of this patch.
The added tuple lock semantics require some rejiggering of the multixact
module, so that the locking level that each transaction is holding can
be stored alongside its Xid. Also, multixacts now need to persist
across server restarts and crashes, because they can now represent not
only tuple locks, but also tuple updates. This means we need more
careful tracking of lifetime of pg_multixact SLRU files; since they now
persist longer, we require more infrastructure to figure out when they
can be removed. pg_upgrade also needs to be careful to copy
pg_multixact files over from the old server to the new, or at least part
of multixact.c state, depending on the versions of the old and new
servers.
Tuple time qualification rules (HeapTupleSatisfies routines) need to be
careful not to consider tuples with the "is multi" infomask bit set as
being only locked; they might need to look up MultiXact values (i.e.
possibly do pg_multixact I/O) to find out the Xid that updated a tuple,
whereas they previously were assured to only use information readily
available from the tuple header. This is considered acceptable, because
the extra I/O would involve cases that would previously cause some
commands to block waiting for concurrent transactions to finish.
Another important change is the fact that locking tuples that have
previously been updated causes the future versions to be marked as
locked, too; this is essential for correctness of foreign key checks.
This causes additional WAL-logging, also (there was previously a single
WAL record for a locked tuple; now there are as many as updated copies
of the tuple there exist.)
With all this in place, contention related to tuples being checked by
foreign key rules should be much reduced.
As a bonus, the old behavior that a subtransaction grabbing a stronger
tuple lock than the parent (sub)transaction held on a given tuple and
later aborting caused the weaker lock to be lost, has been fixed.
Many new spec files were added for isolation tester framework, to ensure
overall behavior is sane. There's probably room for several more tests.
There were several reviewers of this patch; in particular, Noah Misch
and Andres Freund spent considerable time in it. Original idea for the
patch came from Simon Riggs, after a problem report by Joel Jacobson.
Most code is from me, with contributions from Marti Raudsepp, Alexander
Shulgin, Noah Misch and Andres Freund.
This patch was discussed in several pgsql-hackers threads; the most
important start at the following message-ids:
AANLkTimo9XVcEzfiBR-ut3KVNDkjm2Vxh+t8kAmWjPuv@mail.gmail.com
1290721684-sup-3951@alvh.no-ip.org
1294953201-sup-2099@alvh.no-ip.org
1320343602-sup-2290@alvh.no-ip.org
1339690386-sup-8927@alvh.no-ip.org
4FE5FF020200002500048A3D@gw.wicourts.gov
4FEAB90A0200002500048B7D@gw.wicourts.gov
13 years ago
If specific tables are named in a locking clause,
then only rows coming from those tables are locked; any other
tables used in the <command>SELECT</command> are simply read as
Improve concurrency of foreign key locking
This patch introduces two additional lock modes for tuples: "SELECT FOR
KEY SHARE" and "SELECT FOR NO KEY UPDATE". These don't block each
other, in contrast with already existing "SELECT FOR SHARE" and "SELECT
FOR UPDATE". UPDATE commands that do not modify the values stored in
the columns that are part of the key of the tuple now grab a SELECT FOR
NO KEY UPDATE lock on the tuple, allowing them to proceed concurrently
with tuple locks of the FOR KEY SHARE variety.
Foreign key triggers now use FOR KEY SHARE instead of FOR SHARE; this
means the concurrency improvement applies to them, which is the whole
point of this patch.
The added tuple lock semantics require some rejiggering of the multixact
module, so that the locking level that each transaction is holding can
be stored alongside its Xid. Also, multixacts now need to persist
across server restarts and crashes, because they can now represent not
only tuple locks, but also tuple updates. This means we need more
careful tracking of lifetime of pg_multixact SLRU files; since they now
persist longer, we require more infrastructure to figure out when they
can be removed. pg_upgrade also needs to be careful to copy
pg_multixact files over from the old server to the new, or at least part
of multixact.c state, depending on the versions of the old and new
servers.
Tuple time qualification rules (HeapTupleSatisfies routines) need to be
careful not to consider tuples with the "is multi" infomask bit set as
being only locked; they might need to look up MultiXact values (i.e.
possibly do pg_multixact I/O) to find out the Xid that updated a tuple,
whereas they previously were assured to only use information readily
available from the tuple header. This is considered acceptable, because
the extra I/O would involve cases that would previously cause some
commands to block waiting for concurrent transactions to finish.
Another important change is the fact that locking tuples that have
previously been updated causes the future versions to be marked as
locked, too; this is essential for correctness of foreign key checks.
This causes additional WAL-logging, also (there was previously a single
WAL record for a locked tuple; now there are as many as updated copies
of the tuple there exist.)
With all this in place, contention related to tuples being checked by
foreign key rules should be much reduced.
As a bonus, the old behavior that a subtransaction grabbing a stronger
tuple lock than the parent (sub)transaction held on a given tuple and
later aborting caused the weaker lock to be lost, has been fixed.
Many new spec files were added for isolation tester framework, to ensure
overall behavior is sane. There's probably room for several more tests.
There were several reviewers of this patch; in particular, Noah Misch
and Andres Freund spent considerable time in it. Original idea for the
patch came from Simon Riggs, after a problem report by Joel Jacobson.
Most code is from me, with contributions from Marti Raudsepp, Alexander
Shulgin, Noah Misch and Andres Freund.
This patch was discussed in several pgsql-hackers threads; the most
important start at the following message-ids:
AANLkTimo9XVcEzfiBR-ut3KVNDkjm2Vxh+t8kAmWjPuv@mail.gmail.com
1290721684-sup-3951@alvh.no-ip.org
1294953201-sup-2099@alvh.no-ip.org
1320343602-sup-2290@alvh.no-ip.org
1339690386-sup-8927@alvh.no-ip.org
4FE5FF020200002500048A3D@gw.wicourts.gov
4FEAB90A0200002500048B7D@gw.wicourts.gov
13 years ago
usual. A locking
clause without a table list affects all tables used in the statement.
Improve concurrency of foreign key locking
This patch introduces two additional lock modes for tuples: "SELECT FOR
KEY SHARE" and "SELECT FOR NO KEY UPDATE". These don't block each
other, in contrast with already existing "SELECT FOR SHARE" and "SELECT
FOR UPDATE". UPDATE commands that do not modify the values stored in
the columns that are part of the key of the tuple now grab a SELECT FOR
NO KEY UPDATE lock on the tuple, allowing them to proceed concurrently
with tuple locks of the FOR KEY SHARE variety.
Foreign key triggers now use FOR KEY SHARE instead of FOR SHARE; this
means the concurrency improvement applies to them, which is the whole
point of this patch.
The added tuple lock semantics require some rejiggering of the multixact
module, so that the locking level that each transaction is holding can
be stored alongside its Xid. Also, multixacts now need to persist
across server restarts and crashes, because they can now represent not
only tuple locks, but also tuple updates. This means we need more
careful tracking of lifetime of pg_multixact SLRU files; since they now
persist longer, we require more infrastructure to figure out when they
can be removed. pg_upgrade also needs to be careful to copy
pg_multixact files over from the old server to the new, or at least part
of multixact.c state, depending on the versions of the old and new
servers.
Tuple time qualification rules (HeapTupleSatisfies routines) need to be
careful not to consider tuples with the "is multi" infomask bit set as
being only locked; they might need to look up MultiXact values (i.e.
possibly do pg_multixact I/O) to find out the Xid that updated a tuple,
whereas they previously were assured to only use information readily
available from the tuple header. This is considered acceptable, because
the extra I/O would involve cases that would previously cause some
commands to block waiting for concurrent transactions to finish.
Another important change is the fact that locking tuples that have
previously been updated causes the future versions to be marked as
locked, too; this is essential for correctness of foreign key checks.
This causes additional WAL-logging, also (there was previously a single
WAL record for a locked tuple; now there are as many as updated copies
of the tuple there exist.)
With all this in place, contention related to tuples being checked by
foreign key rules should be much reduced.
As a bonus, the old behavior that a subtransaction grabbing a stronger
tuple lock than the parent (sub)transaction held on a given tuple and
later aborting caused the weaker lock to be lost, has been fixed.
Many new spec files were added for isolation tester framework, to ensure
overall behavior is sane. There's probably room for several more tests.
There were several reviewers of this patch; in particular, Noah Misch
and Andres Freund spent considerable time in it. Original idea for the
patch came from Simon Riggs, after a problem report by Joel Jacobson.
Most code is from me, with contributions from Marti Raudsepp, Alexander
Shulgin, Noah Misch and Andres Freund.
This patch was discussed in several pgsql-hackers threads; the most
important start at the following message-ids:
AANLkTimo9XVcEzfiBR-ut3KVNDkjm2Vxh+t8kAmWjPuv@mail.gmail.com
1290721684-sup-3951@alvh.no-ip.org
1294953201-sup-2099@alvh.no-ip.org
1320343602-sup-2290@alvh.no-ip.org
1339690386-sup-8927@alvh.no-ip.org
4FE5FF020200002500048A3D@gw.wicourts.gov
4FEAB90A0200002500048B7D@gw.wicourts.gov
13 years ago
If a locking clause is
applied to a view or sub-query, it affects all tables used in
the view or sub-query.
Improve concurrency of foreign key locking
This patch introduces two additional lock modes for tuples: "SELECT FOR
KEY SHARE" and "SELECT FOR NO KEY UPDATE". These don't block each
other, in contrast with already existing "SELECT FOR SHARE" and "SELECT
FOR UPDATE". UPDATE commands that do not modify the values stored in
the columns that are part of the key of the tuple now grab a SELECT FOR
NO KEY UPDATE lock on the tuple, allowing them to proceed concurrently
with tuple locks of the FOR KEY SHARE variety.
Foreign key triggers now use FOR KEY SHARE instead of FOR SHARE; this
means the concurrency improvement applies to them, which is the whole
point of this patch.
The added tuple lock semantics require some rejiggering of the multixact
module, so that the locking level that each transaction is holding can
be stored alongside its Xid. Also, multixacts now need to persist
across server restarts and crashes, because they can now represent not
only tuple locks, but also tuple updates. This means we need more
careful tracking of lifetime of pg_multixact SLRU files; since they now
persist longer, we require more infrastructure to figure out when they
can be removed. pg_upgrade also needs to be careful to copy
pg_multixact files over from the old server to the new, or at least part
of multixact.c state, depending on the versions of the old and new
servers.
Tuple time qualification rules (HeapTupleSatisfies routines) need to be
careful not to consider tuples with the "is multi" infomask bit set as
being only locked; they might need to look up MultiXact values (i.e.
possibly do pg_multixact I/O) to find out the Xid that updated a tuple,
whereas they previously were assured to only use information readily
available from the tuple header. This is considered acceptable, because
the extra I/O would involve cases that would previously cause some
commands to block waiting for concurrent transactions to finish.
Another important change is the fact that locking tuples that have
previously been updated causes the future versions to be marked as
locked, too; this is essential for correctness of foreign key checks.
This causes additional WAL-logging, also (there was previously a single
WAL record for a locked tuple; now there are as many as updated copies
of the tuple there exist.)
With all this in place, contention related to tuples being checked by
foreign key rules should be much reduced.
As a bonus, the old behavior that a subtransaction grabbing a stronger
tuple lock than the parent (sub)transaction held on a given tuple and
later aborting caused the weaker lock to be lost, has been fixed.
Many new spec files were added for isolation tester framework, to ensure
overall behavior is sane. There's probably room for several more tests.
There were several reviewers of this patch; in particular, Noah Misch
and Andres Freund spent considerable time in it. Original idea for the
patch came from Simon Riggs, after a problem report by Joel Jacobson.
Most code is from me, with contributions from Marti Raudsepp, Alexander
Shulgin, Noah Misch and Andres Freund.
This patch was discussed in several pgsql-hackers threads; the most
important start at the following message-ids:
AANLkTimo9XVcEzfiBR-ut3KVNDkjm2Vxh+t8kAmWjPuv@mail.gmail.com
1290721684-sup-3951@alvh.no-ip.org
1294953201-sup-2099@alvh.no-ip.org
1320343602-sup-2290@alvh.no-ip.org
1339690386-sup-8927@alvh.no-ip.org
4FE5FF020200002500048A3D@gw.wicourts.gov
4FEAB90A0200002500048B7D@gw.wicourts.gov
13 years ago
However, these clauses
do not apply to <literal>WITH</> queries referenced by the primary query.
If you want row locking to occur within a <literal>WITH</> query, specify
Improve concurrency of foreign key locking
This patch introduces two additional lock modes for tuples: "SELECT FOR
KEY SHARE" and "SELECT FOR NO KEY UPDATE". These don't block each
other, in contrast with already existing "SELECT FOR SHARE" and "SELECT
FOR UPDATE". UPDATE commands that do not modify the values stored in
the columns that are part of the key of the tuple now grab a SELECT FOR
NO KEY UPDATE lock on the tuple, allowing them to proceed concurrently
with tuple locks of the FOR KEY SHARE variety.
Foreign key triggers now use FOR KEY SHARE instead of FOR SHARE; this
means the concurrency improvement applies to them, which is the whole
point of this patch.
The added tuple lock semantics require some rejiggering of the multixact
module, so that the locking level that each transaction is holding can
be stored alongside its Xid. Also, multixacts now need to persist
across server restarts and crashes, because they can now represent not
only tuple locks, but also tuple updates. This means we need more
careful tracking of lifetime of pg_multixact SLRU files; since they now
persist longer, we require more infrastructure to figure out when they
can be removed. pg_upgrade also needs to be careful to copy
pg_multixact files over from the old server to the new, or at least part
of multixact.c state, depending on the versions of the old and new
servers.
Tuple time qualification rules (HeapTupleSatisfies routines) need to be
careful not to consider tuples with the "is multi" infomask bit set as
being only locked; they might need to look up MultiXact values (i.e.
possibly do pg_multixact I/O) to find out the Xid that updated a tuple,
whereas they previously were assured to only use information readily
available from the tuple header. This is considered acceptable, because
the extra I/O would involve cases that would previously cause some
commands to block waiting for concurrent transactions to finish.
Another important change is the fact that locking tuples that have
previously been updated causes the future versions to be marked as
locked, too; this is essential for correctness of foreign key checks.
This causes additional WAL-logging, also (there was previously a single
WAL record for a locked tuple; now there are as many as updated copies
of the tuple there exist.)
With all this in place, contention related to tuples being checked by
foreign key rules should be much reduced.
As a bonus, the old behavior that a subtransaction grabbing a stronger
tuple lock than the parent (sub)transaction held on a given tuple and
later aborting caused the weaker lock to be lost, has been fixed.
Many new spec files were added for isolation tester framework, to ensure
overall behavior is sane. There's probably room for several more tests.
There were several reviewers of this patch; in particular, Noah Misch
and Andres Freund spent considerable time in it. Original idea for the
patch came from Simon Riggs, after a problem report by Joel Jacobson.
Most code is from me, with contributions from Marti Raudsepp, Alexander
Shulgin, Noah Misch and Andres Freund.
This patch was discussed in several pgsql-hackers threads; the most
important start at the following message-ids:
AANLkTimo9XVcEzfiBR-ut3KVNDkjm2Vxh+t8kAmWjPuv@mail.gmail.com
1290721684-sup-3951@alvh.no-ip.org
1294953201-sup-2099@alvh.no-ip.org
1320343602-sup-2290@alvh.no-ip.org
1339690386-sup-8927@alvh.no-ip.org
4FE5FF020200002500048A3D@gw.wicourts.gov
4FEAB90A0200002500048B7D@gw.wicourts.gov
13 years ago
a locking clause within the <literal>WITH</> query.
</para>
<para>
Improve concurrency of foreign key locking
This patch introduces two additional lock modes for tuples: "SELECT FOR
KEY SHARE" and "SELECT FOR NO KEY UPDATE". These don't block each
other, in contrast with already existing "SELECT FOR SHARE" and "SELECT
FOR UPDATE". UPDATE commands that do not modify the values stored in
the columns that are part of the key of the tuple now grab a SELECT FOR
NO KEY UPDATE lock on the tuple, allowing them to proceed concurrently
with tuple locks of the FOR KEY SHARE variety.
Foreign key triggers now use FOR KEY SHARE instead of FOR SHARE; this
means the concurrency improvement applies to them, which is the whole
point of this patch.
The added tuple lock semantics require some rejiggering of the multixact
module, so that the locking level that each transaction is holding can
be stored alongside its Xid. Also, multixacts now need to persist
across server restarts and crashes, because they can now represent not
only tuple locks, but also tuple updates. This means we need more
careful tracking of lifetime of pg_multixact SLRU files; since they now
persist longer, we require more infrastructure to figure out when they
can be removed. pg_upgrade also needs to be careful to copy
pg_multixact files over from the old server to the new, or at least part
of multixact.c state, depending on the versions of the old and new
servers.
Tuple time qualification rules (HeapTupleSatisfies routines) need to be
careful not to consider tuples with the "is multi" infomask bit set as
being only locked; they might need to look up MultiXact values (i.e.
possibly do pg_multixact I/O) to find out the Xid that updated a tuple,
whereas they previously were assured to only use information readily
available from the tuple header. This is considered acceptable, because
the extra I/O would involve cases that would previously cause some
commands to block waiting for concurrent transactions to finish.
Another important change is the fact that locking tuples that have
previously been updated causes the future versions to be marked as
locked, too; this is essential for correctness of foreign key checks.
This causes additional WAL-logging, also (there was previously a single
WAL record for a locked tuple; now there are as many as updated copies
of the tuple there exist.)
With all this in place, contention related to tuples being checked by
foreign key rules should be much reduced.
As a bonus, the old behavior that a subtransaction grabbing a stronger
tuple lock than the parent (sub)transaction held on a given tuple and
later aborting caused the weaker lock to be lost, has been fixed.
Many new spec files were added for isolation tester framework, to ensure
overall behavior is sane. There's probably room for several more tests.
There were several reviewers of this patch; in particular, Noah Misch
and Andres Freund spent considerable time in it. Original idea for the
patch came from Simon Riggs, after a problem report by Joel Jacobson.
Most code is from me, with contributions from Marti Raudsepp, Alexander
Shulgin, Noah Misch and Andres Freund.
This patch was discussed in several pgsql-hackers threads; the most
important start at the following message-ids:
AANLkTimo9XVcEzfiBR-ut3KVNDkjm2Vxh+t8kAmWjPuv@mail.gmail.com
1290721684-sup-3951@alvh.no-ip.org
1294953201-sup-2099@alvh.no-ip.org
1320343602-sup-2290@alvh.no-ip.org
1339690386-sup-8927@alvh.no-ip.org
4FE5FF020200002500048A3D@gw.wicourts.gov
4FEAB90A0200002500048B7D@gw.wicourts.gov
13 years ago
Multiple locking
clauses can be written if it is necessary to specify different locking
behavior for different tables. If the same table is mentioned (or
Improve concurrency of foreign key locking
This patch introduces two additional lock modes for tuples: "SELECT FOR
KEY SHARE" and "SELECT FOR NO KEY UPDATE". These don't block each
other, in contrast with already existing "SELECT FOR SHARE" and "SELECT
FOR UPDATE". UPDATE commands that do not modify the values stored in
the columns that are part of the key of the tuple now grab a SELECT FOR
NO KEY UPDATE lock on the tuple, allowing them to proceed concurrently
with tuple locks of the FOR KEY SHARE variety.
Foreign key triggers now use FOR KEY SHARE instead of FOR SHARE; this
means the concurrency improvement applies to them, which is the whole
point of this patch.
The added tuple lock semantics require some rejiggering of the multixact
module, so that the locking level that each transaction is holding can
be stored alongside its Xid. Also, multixacts now need to persist
across server restarts and crashes, because they can now represent not
only tuple locks, but also tuple updates. This means we need more
careful tracking of lifetime of pg_multixact SLRU files; since they now
persist longer, we require more infrastructure to figure out when they
can be removed. pg_upgrade also needs to be careful to copy
pg_multixact files over from the old server to the new, or at least part
of multixact.c state, depending on the versions of the old and new
servers.
Tuple time qualification rules (HeapTupleSatisfies routines) need to be
careful not to consider tuples with the "is multi" infomask bit set as
being only locked; they might need to look up MultiXact values (i.e.
possibly do pg_multixact I/O) to find out the Xid that updated a tuple,
whereas they previously were assured to only use information readily
available from the tuple header. This is considered acceptable, because
the extra I/O would involve cases that would previously cause some
commands to block waiting for concurrent transactions to finish.
Another important change is the fact that locking tuples that have
previously been updated causes the future versions to be marked as
locked, too; this is essential for correctness of foreign key checks.
This causes additional WAL-logging, also (there was previously a single
WAL record for a locked tuple; now there are as many as updated copies
of the tuple there exist.)
With all this in place, contention related to tuples being checked by
foreign key rules should be much reduced.
As a bonus, the old behavior that a subtransaction grabbing a stronger
tuple lock than the parent (sub)transaction held on a given tuple and
later aborting caused the weaker lock to be lost, has been fixed.
Many new spec files were added for isolation tester framework, to ensure
overall behavior is sane. There's probably room for several more tests.
There were several reviewers of this patch; in particular, Noah Misch
and Andres Freund spent considerable time in it. Original idea for the
patch came from Simon Riggs, after a problem report by Joel Jacobson.
Most code is from me, with contributions from Marti Raudsepp, Alexander
Shulgin, Noah Misch and Andres Freund.
This patch was discussed in several pgsql-hackers threads; the most
important start at the following message-ids:
AANLkTimo9XVcEzfiBR-ut3KVNDkjm2Vxh+t8kAmWjPuv@mail.gmail.com
1290721684-sup-3951@alvh.no-ip.org
1294953201-sup-2099@alvh.no-ip.org
1320343602-sup-2290@alvh.no-ip.org
1339690386-sup-8927@alvh.no-ip.org
4FE5FF020200002500048A3D@gw.wicourts.gov
4FEAB90A0200002500048B7D@gw.wicourts.gov
13 years ago
implicitly affected) by more than one locking clause,
then it is processed as if it was only specified by the strongest one.
Similarly, a table is processed
as <literal>NOWAIT</> if that is specified in any of the clauses
affecting it.
</para>
<para>
Improve concurrency of foreign key locking
This patch introduces two additional lock modes for tuples: "SELECT FOR
KEY SHARE" and "SELECT FOR NO KEY UPDATE". These don't block each
other, in contrast with already existing "SELECT FOR SHARE" and "SELECT
FOR UPDATE". UPDATE commands that do not modify the values stored in
the columns that are part of the key of the tuple now grab a SELECT FOR
NO KEY UPDATE lock on the tuple, allowing them to proceed concurrently
with tuple locks of the FOR KEY SHARE variety.
Foreign key triggers now use FOR KEY SHARE instead of FOR SHARE; this
means the concurrency improvement applies to them, which is the whole
point of this patch.
The added tuple lock semantics require some rejiggering of the multixact
module, so that the locking level that each transaction is holding can
be stored alongside its Xid. Also, multixacts now need to persist
across server restarts and crashes, because they can now represent not
only tuple locks, but also tuple updates. This means we need more
careful tracking of lifetime of pg_multixact SLRU files; since they now
persist longer, we require more infrastructure to figure out when they
can be removed. pg_upgrade also needs to be careful to copy
pg_multixact files over from the old server to the new, or at least part
of multixact.c state, depending on the versions of the old and new
servers.
Tuple time qualification rules (HeapTupleSatisfies routines) need to be
careful not to consider tuples with the "is multi" infomask bit set as
being only locked; they might need to look up MultiXact values (i.e.
possibly do pg_multixact I/O) to find out the Xid that updated a tuple,
whereas they previously were assured to only use information readily
available from the tuple header. This is considered acceptable, because
the extra I/O would involve cases that would previously cause some
commands to block waiting for concurrent transactions to finish.
Another important change is the fact that locking tuples that have
previously been updated causes the future versions to be marked as
locked, too; this is essential for correctness of foreign key checks.
This causes additional WAL-logging, also (there was previously a single
WAL record for a locked tuple; now there are as many as updated copies
of the tuple there exist.)
With all this in place, contention related to tuples being checked by
foreign key rules should be much reduced.
As a bonus, the old behavior that a subtransaction grabbing a stronger
tuple lock than the parent (sub)transaction held on a given tuple and
later aborting caused the weaker lock to be lost, has been fixed.
Many new spec files were added for isolation tester framework, to ensure
overall behavior is sane. There's probably room for several more tests.
There were several reviewers of this patch; in particular, Noah Misch
and Andres Freund spent considerable time in it. Original idea for the
patch came from Simon Riggs, after a problem report by Joel Jacobson.
Most code is from me, with contributions from Marti Raudsepp, Alexander
Shulgin, Noah Misch and Andres Freund.
This patch was discussed in several pgsql-hackers threads; the most
important start at the following message-ids:
AANLkTimo9XVcEzfiBR-ut3KVNDkjm2Vxh+t8kAmWjPuv@mail.gmail.com
1290721684-sup-3951@alvh.no-ip.org
1294953201-sup-2099@alvh.no-ip.org
1320343602-sup-2290@alvh.no-ip.org
1339690386-sup-8927@alvh.no-ip.org
4FE5FF020200002500048A3D@gw.wicourts.gov
4FEAB90A0200002500048B7D@gw.wicourts.gov
13 years ago
The locking clauses cannot be
Wording cleanup for error messages. Also change can't -> cannot.
Standard English uses "may", "can", and "might" in different ways:
may - permission, "You may borrow my rake."
can - ability, "I can lift that log."
might - possibility, "It might rain today."
Unfortunately, in conversational English, their use is often mixed, as
in, "You may use this variable to do X", when in fact, "can" is a better
choice. Similarly, "It may crash" is better stated, "It might crash".
19 years ago
used in contexts where returned rows cannot be clearly identified with
individual table rows; for example they cannot be used with aggregation.
</para>
<para>
Improve concurrency of foreign key locking
This patch introduces two additional lock modes for tuples: "SELECT FOR
KEY SHARE" and "SELECT FOR NO KEY UPDATE". These don't block each
other, in contrast with already existing "SELECT FOR SHARE" and "SELECT
FOR UPDATE". UPDATE commands that do not modify the values stored in
the columns that are part of the key of the tuple now grab a SELECT FOR
NO KEY UPDATE lock on the tuple, allowing them to proceed concurrently
with tuple locks of the FOR KEY SHARE variety.
Foreign key triggers now use FOR KEY SHARE instead of FOR SHARE; this
means the concurrency improvement applies to them, which is the whole
point of this patch.
The added tuple lock semantics require some rejiggering of the multixact
module, so that the locking level that each transaction is holding can
be stored alongside its Xid. Also, multixacts now need to persist
across server restarts and crashes, because they can now represent not
only tuple locks, but also tuple updates. This means we need more
careful tracking of lifetime of pg_multixact SLRU files; since they now
persist longer, we require more infrastructure to figure out when they
can be removed. pg_upgrade also needs to be careful to copy
pg_multixact files over from the old server to the new, or at least part
of multixact.c state, depending on the versions of the old and new
servers.
Tuple time qualification rules (HeapTupleSatisfies routines) need to be
careful not to consider tuples with the "is multi" infomask bit set as
being only locked; they might need to look up MultiXact values (i.e.
possibly do pg_multixact I/O) to find out the Xid that updated a tuple,
whereas they previously were assured to only use information readily
available from the tuple header. This is considered acceptable, because
the extra I/O would involve cases that would previously cause some
commands to block waiting for concurrent transactions to finish.
Another important change is the fact that locking tuples that have
previously been updated causes the future versions to be marked as
locked, too; this is essential for correctness of foreign key checks.
This causes additional WAL-logging, also (there was previously a single
WAL record for a locked tuple; now there are as many as updated copies
of the tuple there exist.)
With all this in place, contention related to tuples being checked by
foreign key rules should be much reduced.
As a bonus, the old behavior that a subtransaction grabbing a stronger
tuple lock than the parent (sub)transaction held on a given tuple and
later aborting caused the weaker lock to be lost, has been fixed.
Many new spec files were added for isolation tester framework, to ensure
overall behavior is sane. There's probably room for several more tests.
There were several reviewers of this patch; in particular, Noah Misch
and Andres Freund spent considerable time in it. Original idea for the
patch came from Simon Riggs, after a problem report by Joel Jacobson.
Most code is from me, with contributions from Marti Raudsepp, Alexander
Shulgin, Noah Misch and Andres Freund.
This patch was discussed in several pgsql-hackers threads; the most
important start at the following message-ids:
AANLkTimo9XVcEzfiBR-ut3KVNDkjm2Vxh+t8kAmWjPuv@mail.gmail.com
1290721684-sup-3951@alvh.no-ip.org
1294953201-sup-2099@alvh.no-ip.org
1320343602-sup-2290@alvh.no-ip.org
1339690386-sup-8927@alvh.no-ip.org
4FE5FF020200002500048A3D@gw.wicourts.gov
4FEAB90A0200002500048B7D@gw.wicourts.gov
13 years ago
When a locking clause
appears at the top level of a <command>SELECT</> query, the rows that
are locked are exactly those that are returned by the query; in the
case of a join query, the rows locked are those that contribute to
returned join rows. In addition, rows that satisfied the query
conditions as of the query snapshot will be locked, although they
will not be returned if they were updated after the snapshot
and no longer satisfy the query conditions. If a
<literal>LIMIT</> is used, locking stops
once enough rows have been returned to satisfy the limit (but note that
rows skipped over by <literal>OFFSET</> will get locked). Similarly,
Improve concurrency of foreign key locking
This patch introduces two additional lock modes for tuples: "SELECT FOR
KEY SHARE" and "SELECT FOR NO KEY UPDATE". These don't block each
other, in contrast with already existing "SELECT FOR SHARE" and "SELECT
FOR UPDATE". UPDATE commands that do not modify the values stored in
the columns that are part of the key of the tuple now grab a SELECT FOR
NO KEY UPDATE lock on the tuple, allowing them to proceed concurrently
with tuple locks of the FOR KEY SHARE variety.
Foreign key triggers now use FOR KEY SHARE instead of FOR SHARE; this
means the concurrency improvement applies to them, which is the whole
point of this patch.
The added tuple lock semantics require some rejiggering of the multixact
module, so that the locking level that each transaction is holding can
be stored alongside its Xid. Also, multixacts now need to persist
across server restarts and crashes, because they can now represent not
only tuple locks, but also tuple updates. This means we need more
careful tracking of lifetime of pg_multixact SLRU files; since they now
persist longer, we require more infrastructure to figure out when they
can be removed. pg_upgrade also needs to be careful to copy
pg_multixact files over from the old server to the new, or at least part
of multixact.c state, depending on the versions of the old and new
servers.
Tuple time qualification rules (HeapTupleSatisfies routines) need to be
careful not to consider tuples with the "is multi" infomask bit set as
being only locked; they might need to look up MultiXact values (i.e.
possibly do pg_multixact I/O) to find out the Xid that updated a tuple,
whereas they previously were assured to only use information readily
available from the tuple header. This is considered acceptable, because
the extra I/O would involve cases that would previously cause some
commands to block waiting for concurrent transactions to finish.
Another important change is the fact that locking tuples that have
previously been updated causes the future versions to be marked as
locked, too; this is essential for correctness of foreign key checks.
This causes additional WAL-logging, also (there was previously a single
WAL record for a locked tuple; now there are as many as updated copies
of the tuple there exist.)
With all this in place, contention related to tuples being checked by
foreign key rules should be much reduced.
As a bonus, the old behavior that a subtransaction grabbing a stronger
tuple lock than the parent (sub)transaction held on a given tuple and
later aborting caused the weaker lock to be lost, has been fixed.
Many new spec files were added for isolation tester framework, to ensure
overall behavior is sane. There's probably room for several more tests.
There were several reviewers of this patch; in particular, Noah Misch
and Andres Freund spent considerable time in it. Original idea for the
patch came from Simon Riggs, after a problem report by Joel Jacobson.
Most code is from me, with contributions from Marti Raudsepp, Alexander
Shulgin, Noah Misch and Andres Freund.
This patch was discussed in several pgsql-hackers threads; the most
important start at the following message-ids:
AANLkTimo9XVcEzfiBR-ut3KVNDkjm2Vxh+t8kAmWjPuv@mail.gmail.com
1290721684-sup-3951@alvh.no-ip.org
1294953201-sup-2099@alvh.no-ip.org
1320343602-sup-2290@alvh.no-ip.org
1339690386-sup-8927@alvh.no-ip.org
4FE5FF020200002500048A3D@gw.wicourts.gov
4FEAB90A0200002500048B7D@gw.wicourts.gov
13 years ago
if a locking clause
is used in a cursor's query, only rows actually fetched or stepped past
by the cursor will be locked.
</para>
<para>
Improve concurrency of foreign key locking
This patch introduces two additional lock modes for tuples: "SELECT FOR
KEY SHARE" and "SELECT FOR NO KEY UPDATE". These don't block each
other, in contrast with already existing "SELECT FOR SHARE" and "SELECT
FOR UPDATE". UPDATE commands that do not modify the values stored in
the columns that are part of the key of the tuple now grab a SELECT FOR
NO KEY UPDATE lock on the tuple, allowing them to proceed concurrently
with tuple locks of the FOR KEY SHARE variety.
Foreign key triggers now use FOR KEY SHARE instead of FOR SHARE; this
means the concurrency improvement applies to them, which is the whole
point of this patch.
The added tuple lock semantics require some rejiggering of the multixact
module, so that the locking level that each transaction is holding can
be stored alongside its Xid. Also, multixacts now need to persist
across server restarts and crashes, because they can now represent not
only tuple locks, but also tuple updates. This means we need more
careful tracking of lifetime of pg_multixact SLRU files; since they now
persist longer, we require more infrastructure to figure out when they
can be removed. pg_upgrade also needs to be careful to copy
pg_multixact files over from the old server to the new, or at least part
of multixact.c state, depending on the versions of the old and new
servers.
Tuple time qualification rules (HeapTupleSatisfies routines) need to be
careful not to consider tuples with the "is multi" infomask bit set as
being only locked; they might need to look up MultiXact values (i.e.
possibly do pg_multixact I/O) to find out the Xid that updated a tuple,
whereas they previously were assured to only use information readily
available from the tuple header. This is considered acceptable, because
the extra I/O would involve cases that would previously cause some
commands to block waiting for concurrent transactions to finish.
Another important change is the fact that locking tuples that have
previously been updated causes the future versions to be marked as
locked, too; this is essential for correctness of foreign key checks.
This causes additional WAL-logging, also (there was previously a single
WAL record for a locked tuple; now there are as many as updated copies
of the tuple there exist.)
With all this in place, contention related to tuples being checked by
foreign key rules should be much reduced.
As a bonus, the old behavior that a subtransaction grabbing a stronger
tuple lock than the parent (sub)transaction held on a given tuple and
later aborting caused the weaker lock to be lost, has been fixed.
Many new spec files were added for isolation tester framework, to ensure
overall behavior is sane. There's probably room for several more tests.
There were several reviewers of this patch; in particular, Noah Misch
and Andres Freund spent considerable time in it. Original idea for the
patch came from Simon Riggs, after a problem report by Joel Jacobson.
Most code is from me, with contributions from Marti Raudsepp, Alexander
Shulgin, Noah Misch and Andres Freund.
This patch was discussed in several pgsql-hackers threads; the most
important start at the following message-ids:
AANLkTimo9XVcEzfiBR-ut3KVNDkjm2Vxh+t8kAmWjPuv@mail.gmail.com
1290721684-sup-3951@alvh.no-ip.org
1294953201-sup-2099@alvh.no-ip.org
1320343602-sup-2290@alvh.no-ip.org
1339690386-sup-8927@alvh.no-ip.org
4FE5FF020200002500048A3D@gw.wicourts.gov
4FEAB90A0200002500048B7D@gw.wicourts.gov
13 years ago
When a locking clause
appears in a sub-<command>SELECT</>, the rows locked are those
returned to the outer query by the sub-query. This might involve
fewer rows than inspection of the sub-query alone would suggest,
since conditions from the outer query might be used to optimize
execution of the sub-query. For example,
<programlisting>
SELECT * FROM (SELECT * FROM mytable FOR UPDATE) ss WHERE col1 = 5;
</programlisting>
will lock only rows having <literal>col1 = 5</>, even though that
condition is not textually within the sub-query.
</para>
Improve concurrency of foreign key locking
This patch introduces two additional lock modes for tuples: "SELECT FOR
KEY SHARE" and "SELECT FOR NO KEY UPDATE". These don't block each
other, in contrast with already existing "SELECT FOR SHARE" and "SELECT
FOR UPDATE". UPDATE commands that do not modify the values stored in
the columns that are part of the key of the tuple now grab a SELECT FOR
NO KEY UPDATE lock on the tuple, allowing them to proceed concurrently
with tuple locks of the FOR KEY SHARE variety.
Foreign key triggers now use FOR KEY SHARE instead of FOR SHARE; this
means the concurrency improvement applies to them, which is the whole
point of this patch.
The added tuple lock semantics require some rejiggering of the multixact
module, so that the locking level that each transaction is holding can
be stored alongside its Xid. Also, multixacts now need to persist
across server restarts and crashes, because they can now represent not
only tuple locks, but also tuple updates. This means we need more
careful tracking of lifetime of pg_multixact SLRU files; since they now
persist longer, we require more infrastructure to figure out when they
can be removed. pg_upgrade also needs to be careful to copy
pg_multixact files over from the old server to the new, or at least part
of multixact.c state, depending on the versions of the old and new
servers.
Tuple time qualification rules (HeapTupleSatisfies routines) need to be
careful not to consider tuples with the "is multi" infomask bit set as
being only locked; they might need to look up MultiXact values (i.e.
possibly do pg_multixact I/O) to find out the Xid that updated a tuple,
whereas they previously were assured to only use information readily
available from the tuple header. This is considered acceptable, because
the extra I/O would involve cases that would previously cause some
commands to block waiting for concurrent transactions to finish.
Another important change is the fact that locking tuples that have
previously been updated causes the future versions to be marked as
locked, too; this is essential for correctness of foreign key checks.
This causes additional WAL-logging, also (there was previously a single
WAL record for a locked tuple; now there are as many as updated copies
of the tuple there exist.)
With all this in place, contention related to tuples being checked by
foreign key rules should be much reduced.
As a bonus, the old behavior that a subtransaction grabbing a stronger
tuple lock than the parent (sub)transaction held on a given tuple and
later aborting caused the weaker lock to be lost, has been fixed.
Many new spec files were added for isolation tester framework, to ensure
overall behavior is sane. There's probably room for several more tests.
There were several reviewers of this patch; in particular, Noah Misch
and Andres Freund spent considerable time in it. Original idea for the
patch came from Simon Riggs, after a problem report by Joel Jacobson.
Most code is from me, with contributions from Marti Raudsepp, Alexander
Shulgin, Noah Misch and Andres Freund.
This patch was discussed in several pgsql-hackers threads; the most
important start at the following message-ids:
AANLkTimo9XVcEzfiBR-ut3KVNDkjm2Vxh+t8kAmWjPuv@mail.gmail.com
1290721684-sup-3951@alvh.no-ip.org
1294953201-sup-2099@alvh.no-ip.org
1320343602-sup-2290@alvh.no-ip.org
1339690386-sup-8927@alvh.no-ip.org
4FE5FF020200002500048A3D@gw.wicourts.gov
4FEAB90A0200002500048B7D@gw.wicourts.gov
13 years ago
<para>
Previous releases failed to preserve a lock which is upgraded by a later
savepoint. For example, this code:
<programlisting>
BEGIN;
SELECT * FROM mytable WHERE key = 1 FOR UPDATE;
SAVEPOINT s;
UPDATE mytable SET ... WHERE key = 1;
ROLLBACK TO s;
</programlisting>
Improve concurrency of foreign key locking
This patch introduces two additional lock modes for tuples: "SELECT FOR
KEY SHARE" and "SELECT FOR NO KEY UPDATE". These don't block each
other, in contrast with already existing "SELECT FOR SHARE" and "SELECT
FOR UPDATE". UPDATE commands that do not modify the values stored in
the columns that are part of the key of the tuple now grab a SELECT FOR
NO KEY UPDATE lock on the tuple, allowing them to proceed concurrently
with tuple locks of the FOR KEY SHARE variety.
Foreign key triggers now use FOR KEY SHARE instead of FOR SHARE; this
means the concurrency improvement applies to them, which is the whole
point of this patch.
The added tuple lock semantics require some rejiggering of the multixact
module, so that the locking level that each transaction is holding can
be stored alongside its Xid. Also, multixacts now need to persist
across server restarts and crashes, because they can now represent not
only tuple locks, but also tuple updates. This means we need more
careful tracking of lifetime of pg_multixact SLRU files; since they now
persist longer, we require more infrastructure to figure out when they
can be removed. pg_upgrade also needs to be careful to copy
pg_multixact files over from the old server to the new, or at least part
of multixact.c state, depending on the versions of the old and new
servers.
Tuple time qualification rules (HeapTupleSatisfies routines) need to be
careful not to consider tuples with the "is multi" infomask bit set as
being only locked; they might need to look up MultiXact values (i.e.
possibly do pg_multixact I/O) to find out the Xid that updated a tuple,
whereas they previously were assured to only use information readily
available from the tuple header. This is considered acceptable, because
the extra I/O would involve cases that would previously cause some
commands to block waiting for concurrent transactions to finish.
Another important change is the fact that locking tuples that have
previously been updated causes the future versions to be marked as
locked, too; this is essential for correctness of foreign key checks.
This causes additional WAL-logging, also (there was previously a single
WAL record for a locked tuple; now there are as many as updated copies
of the tuple there exist.)
With all this in place, contention related to tuples being checked by
foreign key rules should be much reduced.
As a bonus, the old behavior that a subtransaction grabbing a stronger
tuple lock than the parent (sub)transaction held on a given tuple and
later aborting caused the weaker lock to be lost, has been fixed.
Many new spec files were added for isolation tester framework, to ensure
overall behavior is sane. There's probably room for several more tests.
There were several reviewers of this patch; in particular, Noah Misch
and Andres Freund spent considerable time in it. Original idea for the
patch came from Simon Riggs, after a problem report by Joel Jacobson.
Most code is from me, with contributions from Marti Raudsepp, Alexander
Shulgin, Noah Misch and Andres Freund.
This patch was discussed in several pgsql-hackers threads; the most
important start at the following message-ids:
AANLkTimo9XVcEzfiBR-ut3KVNDkjm2Vxh+t8kAmWjPuv@mail.gmail.com
1290721684-sup-3951@alvh.no-ip.org
1294953201-sup-2099@alvh.no-ip.org
1320343602-sup-2290@alvh.no-ip.org
1339690386-sup-8927@alvh.no-ip.org
4FE5FF020200002500048A3D@gw.wicourts.gov
4FEAB90A0200002500048B7D@gw.wicourts.gov
13 years ago
would fail to preserve the <literal>FOR UPDATE</> lock after the
<command>ROLLBACK TO</>. This has been fixed in release 9.3.
Improve concurrency of foreign key locking
This patch introduces two additional lock modes for tuples: "SELECT FOR
KEY SHARE" and "SELECT FOR NO KEY UPDATE". These don't block each
other, in contrast with already existing "SELECT FOR SHARE" and "SELECT
FOR UPDATE". UPDATE commands that do not modify the values stored in
the columns that are part of the key of the tuple now grab a SELECT FOR
NO KEY UPDATE lock on the tuple, allowing them to proceed concurrently
with tuple locks of the FOR KEY SHARE variety.
Foreign key triggers now use FOR KEY SHARE instead of FOR SHARE; this
means the concurrency improvement applies to them, which is the whole
point of this patch.
The added tuple lock semantics require some rejiggering of the multixact
module, so that the locking level that each transaction is holding can
be stored alongside its Xid. Also, multixacts now need to persist
across server restarts and crashes, because they can now represent not
only tuple locks, but also tuple updates. This means we need more
careful tracking of lifetime of pg_multixact SLRU files; since they now
persist longer, we require more infrastructure to figure out when they
can be removed. pg_upgrade also needs to be careful to copy
pg_multixact files over from the old server to the new, or at least part
of multixact.c state, depending on the versions of the old and new
servers.
Tuple time qualification rules (HeapTupleSatisfies routines) need to be
careful not to consider tuples with the "is multi" infomask bit set as
being only locked; they might need to look up MultiXact values (i.e.
possibly do pg_multixact I/O) to find out the Xid that updated a tuple,
whereas they previously were assured to only use information readily
available from the tuple header. This is considered acceptable, because
the extra I/O would involve cases that would previously cause some
commands to block waiting for concurrent transactions to finish.
Another important change is the fact that locking tuples that have
previously been updated causes the future versions to be marked as
locked, too; this is essential for correctness of foreign key checks.
This causes additional WAL-logging, also (there was previously a single
WAL record for a locked tuple; now there are as many as updated copies
of the tuple there exist.)
With all this in place, contention related to tuples being checked by
foreign key rules should be much reduced.
As a bonus, the old behavior that a subtransaction grabbing a stronger
tuple lock than the parent (sub)transaction held on a given tuple and
later aborting caused the weaker lock to be lost, has been fixed.
Many new spec files were added for isolation tester framework, to ensure
overall behavior is sane. There's probably room for several more tests.
There were several reviewers of this patch; in particular, Noah Misch
and Andres Freund spent considerable time in it. Original idea for the
patch came from Simon Riggs, after a problem report by Joel Jacobson.
Most code is from me, with contributions from Marti Raudsepp, Alexander
Shulgin, Noah Misch and Andres Freund.
This patch was discussed in several pgsql-hackers threads; the most
important start at the following message-ids:
AANLkTimo9XVcEzfiBR-ut3KVNDkjm2Vxh+t8kAmWjPuv@mail.gmail.com
1290721684-sup-3951@alvh.no-ip.org
1294953201-sup-2099@alvh.no-ip.org
1320343602-sup-2290@alvh.no-ip.org
1339690386-sup-8927@alvh.no-ip.org
4FE5FF020200002500048A3D@gw.wicourts.gov
4FEAB90A0200002500048B7D@gw.wicourts.gov
13 years ago
</para>
<caution>
<para>
It is possible for a <command>SELECT</> command running at the <literal>READ
COMMITTED</literal> transaction isolation level and using <literal>ORDER
Improve concurrency of foreign key locking
This patch introduces two additional lock modes for tuples: "SELECT FOR
KEY SHARE" and "SELECT FOR NO KEY UPDATE". These don't block each
other, in contrast with already existing "SELECT FOR SHARE" and "SELECT
FOR UPDATE". UPDATE commands that do not modify the values stored in
the columns that are part of the key of the tuple now grab a SELECT FOR
NO KEY UPDATE lock on the tuple, allowing them to proceed concurrently
with tuple locks of the FOR KEY SHARE variety.
Foreign key triggers now use FOR KEY SHARE instead of FOR SHARE; this
means the concurrency improvement applies to them, which is the whole
point of this patch.
The added tuple lock semantics require some rejiggering of the multixact
module, so that the locking level that each transaction is holding can
be stored alongside its Xid. Also, multixacts now need to persist
across server restarts and crashes, because they can now represent not
only tuple locks, but also tuple updates. This means we need more
careful tracking of lifetime of pg_multixact SLRU files; since they now
persist longer, we require more infrastructure to figure out when they
can be removed. pg_upgrade also needs to be careful to copy
pg_multixact files over from the old server to the new, or at least part
of multixact.c state, depending on the versions of the old and new
servers.
Tuple time qualification rules (HeapTupleSatisfies routines) need to be
careful not to consider tuples with the "is multi" infomask bit set as
being only locked; they might need to look up MultiXact values (i.e.
possibly do pg_multixact I/O) to find out the Xid that updated a tuple,
whereas they previously were assured to only use information readily
available from the tuple header. This is considered acceptable, because
the extra I/O would involve cases that would previously cause some
commands to block waiting for concurrent transactions to finish.
Another important change is the fact that locking tuples that have
previously been updated causes the future versions to be marked as
locked, too; this is essential for correctness of foreign key checks.
This causes additional WAL-logging, also (there was previously a single
WAL record for a locked tuple; now there are as many as updated copies
of the tuple there exist.)
With all this in place, contention related to tuples being checked by
foreign key rules should be much reduced.
As a bonus, the old behavior that a subtransaction grabbing a stronger
tuple lock than the parent (sub)transaction held on a given tuple and
later aborting caused the weaker lock to be lost, has been fixed.
Many new spec files were added for isolation tester framework, to ensure
overall behavior is sane. There's probably room for several more tests.
There were several reviewers of this patch; in particular, Noah Misch
and Andres Freund spent considerable time in it. Original idea for the
patch came from Simon Riggs, after a problem report by Joel Jacobson.
Most code is from me, with contributions from Marti Raudsepp, Alexander
Shulgin, Noah Misch and Andres Freund.
This patch was discussed in several pgsql-hackers threads; the most
important start at the following message-ids:
AANLkTimo9XVcEzfiBR-ut3KVNDkjm2Vxh+t8kAmWjPuv@mail.gmail.com
1290721684-sup-3951@alvh.no-ip.org
1294953201-sup-2099@alvh.no-ip.org
1320343602-sup-2290@alvh.no-ip.org
1339690386-sup-8927@alvh.no-ip.org
4FE5FF020200002500048A3D@gw.wicourts.gov
4FEAB90A0200002500048B7D@gw.wicourts.gov
13 years ago
BY</literal> and a locking clause to return rows out of
order. This is because <literal>ORDER BY</> is applied first.
The command sorts the result, but might then block trying to obtain a lock
on one or more of the rows. Once the <literal>SELECT</> unblocks, some
of the ordering column values might have been modified, leading to those
rows appearing to be out of order (though they are in order in terms
of the original column values). This can be worked around at need by
placing the <literal>FOR UPDATE/SHARE</literal> clause in a sub-query,
for example
<programlisting>
SELECT * FROM (SELECT * FROM mytable FOR UPDATE) ss ORDER BY column1;
</programlisting>
Note that this will result in locking all rows of <structname>mytable</>,
whereas <literal>FOR UPDATE</> at the top level would lock only the
actually returned rows. This can make for a significant performance
difference, particularly if the <literal>ORDER BY</> is combined with
<literal>LIMIT</> or other restrictions. So this technique is recommended
only if concurrent updates of the ordering columns are expected and a
strictly sorted result is required.
</para>
<para>
At the <literal>REPEATABLE READ</literal> or <literal>SERIALIZABLE</literal>
transaction isolation level this would cause a serialization failure (with
a <literal>SQLSTATE</literal> of <literal>'40001'</literal>), so there is
no possibility of receiving rows out of order under these isolation levels.
</para>
</caution>
</refsect2>
<refsect2 id="SQL-TABLE">
<title><literal>TABLE</literal> Command</title>
<para>
The command
<programlisting>
TABLE <replaceable class="parameter">name</replaceable>
</programlisting>
is completely equivalent to
<programlisting>
SELECT * FROM <replaceable class="parameter">name</replaceable>
</programlisting>
It can be used as a top-level command or as a space-saving syntax
variant in parts of complex queries.
</para>
</refsect2>
</refsect1>
<refsect1>
<title>Examples</title>
<para>
To join the table <literal>films</literal> with the table
<literal>distributors</literal>:
<programlisting>
SELECT f.title, f.did, d.name, f.date_prod, f.kind
FROM distributors d, films f
WHERE f.did = d.did
title | did | name | date_prod | kind
-------------------+-----+--------------+------------+----------
The Third Man | 101 | British Lion | 1949-12-23 | Drama
The African Queen | 101 | British Lion | 1951-08-11 | Romantic
...
</programlisting>
</para>
<para>
To sum the column <literal>len</literal> of all films and group
the results by <literal>kind</literal>:
<programlisting>
SELECT kind, sum(len) AS total FROM films GROUP BY kind;
kind | total
----------+-------
Action | 07:34
Comedy | 02:58
Drama | 14:28
Musical | 06:42
Romantic | 04:38
</programlisting>
</para>
<para>
To sum the column <literal>len</literal> of all films, group
the results by <literal>kind</literal> and show those group totals
that are less than 5 hours:
<programlisting>
SELECT kind, sum(len) AS total
FROM films
GROUP BY kind
HAVING sum(len) < interval '5 hours';
kind | total
----------+-------
Comedy | 02:58
Romantic | 04:38
</programlisting>
</para>
<para>
The following two examples are identical ways of sorting the individual
results according to the contents of the second column
(<literal>name</literal>):
<programlisting>
SELECT * FROM distributors ORDER BY name;
SELECT * FROM distributors ORDER BY 2;
did | name
-----+------------------
109 | 20th Century Fox
110 | Bavaria Atelier
101 | British Lion
107 | Columbia
102 | Jean Luc Godard
113 | Luso films
104 | Mosfilm
103 | Paramount
106 | Toho
105 | United Artists
111 | Walt Disney
112 | Warner Bros.
108 | Westward
</programlisting>
</para>
<para>
The next example shows how to obtain the union of the tables
<literal>distributors</literal> and
<literal>actors</literal>, restricting the results to those that begin
with the letter W in each table. Only distinct rows are wanted, so the
key word <literal>ALL</literal> is omitted.
<programlisting>
distributors: actors:
did | name id | name
-----+-------------- ----+----------------
108 | Westward 1 | Woody Allen
111 | Walt Disney 2 | Warren Beatty
112 | Warner Bros. 3 | Walter Matthau
... ...
SELECT distributors.name
FROM distributors
WHERE distributors.name LIKE 'W%'
UNION
SELECT actors.name
FROM actors
WHERE actors.name LIKE 'W%';
name
----------------
Walt Disney
Walter Matthau
Warner Bros.
Warren Beatty
Westward
Woody Allen
Attached are two patches to implement and document anonymous composite
types for Table Functions, as previously proposed on HACKERS. Here is a
brief explanation:
1. Creates a new pg_type typtype: 'p' for pseudo type (currently either
'b' for base or 'c' for catalog, i.e. a class).
2. Creates new builtin type of typtype='p' named RECORD. This is the
first of potentially several pseudo types.
3. Modify FROM clause grammer to accept:
SELECT * FROM my_func() AS m(colname1 type1, colname2 type1, ...)
where m is the table alias, colname1, etc are the column names, and
type1, etc are the column types.
4. When typtype == 'p' and the function return type is RECORD, a list
of column defs is required, and when typtype != 'p', it is
disallowed.
5. A check was added to ensure that the tupdesc provide via the parser
and the actual return tupdesc match in number and type of
attributes.
When creating a function you can do:
CREATE FUNCTION foo(text) RETURNS setof RECORD ...
When using it you can do:
SELECT * from foo(sqlstmt) AS (f1 int, f2 text, f3 timestamp)
or
SELECT * from foo(sqlstmt) AS f(f1 int, f2 text, f3 timestamp)
or
SELECT * from foo(sqlstmt) f(f1 int, f2 text, f3 timestamp)
Included in the patches are adjustments to the regression test sql and
expected files, and documentation.
p.s.
This potentially solves (or at least improves) the issue of builtin
Table Functions. They can be bootstrapped as returning RECORD, and
we can wrap system views around them with properly specified column
defs. For example:
CREATE VIEW pg_settings AS
SELECT s.name, s.setting
FROM show_all_settings()AS s(name text, setting text);
Then we can also add the UPDATE RULE that I previously posted to
pg_settings, and have pg_settings act like a virtual table, allowing
settings to be queried and set.
Joe Conway
23 years ago
</programlisting>
</para>
<para>
This example shows how to use a function in the <literal>FROM</>
clause, both with and without a column definition list:
Attached are two patches to implement and document anonymous composite
types for Table Functions, as previously proposed on HACKERS. Here is a
brief explanation:
1. Creates a new pg_type typtype: 'p' for pseudo type (currently either
'b' for base or 'c' for catalog, i.e. a class).
2. Creates new builtin type of typtype='p' named RECORD. This is the
first of potentially several pseudo types.
3. Modify FROM clause grammer to accept:
SELECT * FROM my_func() AS m(colname1 type1, colname2 type1, ...)
where m is the table alias, colname1, etc are the column names, and
type1, etc are the column types.
4. When typtype == 'p' and the function return type is RECORD, a list
of column defs is required, and when typtype != 'p', it is
disallowed.
5. A check was added to ensure that the tupdesc provide via the parser
and the actual return tupdesc match in number and type of
attributes.
When creating a function you can do:
CREATE FUNCTION foo(text) RETURNS setof RECORD ...
When using it you can do:
SELECT * from foo(sqlstmt) AS (f1 int, f2 text, f3 timestamp)
or
SELECT * from foo(sqlstmt) AS f(f1 int, f2 text, f3 timestamp)
or
SELECT * from foo(sqlstmt) f(f1 int, f2 text, f3 timestamp)
Included in the patches are adjustments to the regression test sql and
expected files, and documentation.
p.s.
This potentially solves (or at least improves) the issue of builtin
Table Functions. They can be bootstrapped as returning RECORD, and
we can wrap system views around them with properly specified column
defs. For example:
CREATE VIEW pg_settings AS
SELECT s.name, s.setting
FROM show_all_settings()AS s(name text, setting text);
Then we can also add the UPDATE RULE that I previously posted to
pg_settings, and have pg_settings act like a virtual table, allowing
settings to be queried and set.
Joe Conway
23 years ago
<programlisting>
CREATE FUNCTION distributors(int) RETURNS SETOF distributors AS $$
SELECT * FROM distributors WHERE did = $1;
$$ LANGUAGE SQL;
Attached are two patches to implement and document anonymous composite
types for Table Functions, as previously proposed on HACKERS. Here is a
brief explanation:
1. Creates a new pg_type typtype: 'p' for pseudo type (currently either
'b' for base or 'c' for catalog, i.e. a class).
2. Creates new builtin type of typtype='p' named RECORD. This is the
first of potentially several pseudo types.
3. Modify FROM clause grammer to accept:
SELECT * FROM my_func() AS m(colname1 type1, colname2 type1, ...)
where m is the table alias, colname1, etc are the column names, and
type1, etc are the column types.
4. When typtype == 'p' and the function return type is RECORD, a list
of column defs is required, and when typtype != 'p', it is
disallowed.
5. A check was added to ensure that the tupdesc provide via the parser
and the actual return tupdesc match in number and type of
attributes.
When creating a function you can do:
CREATE FUNCTION foo(text) RETURNS setof RECORD ...
When using it you can do:
SELECT * from foo(sqlstmt) AS (f1 int, f2 text, f3 timestamp)
or
SELECT * from foo(sqlstmt) AS f(f1 int, f2 text, f3 timestamp)
or
SELECT * from foo(sqlstmt) f(f1 int, f2 text, f3 timestamp)
Included in the patches are adjustments to the regression test sql and
expected files, and documentation.
p.s.
This potentially solves (or at least improves) the issue of builtin
Table Functions. They can be bootstrapped as returning RECORD, and
we can wrap system views around them with properly specified column
defs. For example:
CREATE VIEW pg_settings AS
SELECT s.name, s.setting
FROM show_all_settings()AS s(name text, setting text);
Then we can also add the UPDATE RULE that I previously posted to
pg_settings, and have pg_settings act like a virtual table, allowing
settings to be queried and set.
Joe Conway
23 years ago
SELECT * FROM distributors(111);
did | name
-----+-------------
111 | Walt Disney
CREATE FUNCTION distributors_2(int) RETURNS SETOF record AS $$
SELECT * FROM distributors WHERE did = $1;
$$ LANGUAGE SQL;
Attached are two patches to implement and document anonymous composite
types for Table Functions, as previously proposed on HACKERS. Here is a
brief explanation:
1. Creates a new pg_type typtype: 'p' for pseudo type (currently either
'b' for base or 'c' for catalog, i.e. a class).
2. Creates new builtin type of typtype='p' named RECORD. This is the
first of potentially several pseudo types.
3. Modify FROM clause grammer to accept:
SELECT * FROM my_func() AS m(colname1 type1, colname2 type1, ...)
where m is the table alias, colname1, etc are the column names, and
type1, etc are the column types.
4. When typtype == 'p' and the function return type is RECORD, a list
of column defs is required, and when typtype != 'p', it is
disallowed.
5. A check was added to ensure that the tupdesc provide via the parser
and the actual return tupdesc match in number and type of
attributes.
When creating a function you can do:
CREATE FUNCTION foo(text) RETURNS setof RECORD ...
When using it you can do:
SELECT * from foo(sqlstmt) AS (f1 int, f2 text, f3 timestamp)
or
SELECT * from foo(sqlstmt) AS f(f1 int, f2 text, f3 timestamp)
or
SELECT * from foo(sqlstmt) f(f1 int, f2 text, f3 timestamp)
Included in the patches are adjustments to the regression test sql and
expected files, and documentation.
p.s.
This potentially solves (or at least improves) the issue of builtin
Table Functions. They can be bootstrapped as returning RECORD, and
we can wrap system views around them with properly specified column
defs. For example:
CREATE VIEW pg_settings AS
SELECT s.name, s.setting
FROM show_all_settings()AS s(name text, setting text);
Then we can also add the UPDATE RULE that I previously posted to
pg_settings, and have pg_settings act like a virtual table, allowing
settings to be queried and set.
Joe Conway
23 years ago
SELECT * FROM distributors_2(111) AS (f1 int, f2 text);
f1 | f2
-----+-------------
111 | Walt Disney
</programlisting>
</para>
Support multi-argument UNNEST(), and TABLE() syntax for multiple functions.
This patch adds the ability to write TABLE( function1(), function2(), ...)
as a single FROM-clause entry. The result is the concatenation of the
first row from each function, followed by the second row from each
function, etc; with NULLs inserted if any function produces fewer rows than
others. This is believed to be a much more useful behavior than what
Postgres currently does with multiple SRFs in a SELECT list.
This syntax also provides a reasonable way to combine use of column
definition lists with WITH ORDINALITY: put the column definition list
inside TABLE(), where it's clear that it doesn't control the ordinality
column as well.
Also implement SQL-compliant multiple-argument UNNEST(), by turning
UNNEST(a,b,c) into TABLE(unnest(a), unnest(b), unnest(c)).
The SQL standard specifies TABLE() with only a single function, not
multiple functions, and it seems to require an implicit UNNEST() which is
not what this patch does. There may be something wrong with that reading
of the spec, though, because if it's right then the spec's TABLE() is just
a pointless alternative spelling of UNNEST(). After further review of
that, we might choose to adopt a different syntax for what this patch does,
but in any case this functionality seems clearly worthwhile.
Andrew Gierth, reviewed by Zoltán Böszörményi and Heikki Linnakangas, and
significantly revised by me
12 years ago
<para>
Here is an example of a function with an ordinality column added:
<programlisting>
SELECT * FROM unnest(ARRAY['a','b','c','d','e','f']) WITH ORDINALITY;
unnest | ordinality
--------+----------
a | 1
b | 2
c | 3
d | 4
e | 5
f | 6
(6 rows)
</programlisting>
</para>
<para>
This example shows how to use a simple <literal>WITH</> clause:
<programlisting>
WITH t AS (
SELECT random() as x FROM generate_series(1, 3)
)
SELECT * FROM t
UNION ALL
SELECT * FROM t
x
--------------------
0.534150459803641
0.520092216785997
0.0735620250925422
0.534150459803641
0.520092216785997
0.0735620250925422
</programlisting>
Notice that the <literal>WITH</> query was evaluated only once,
so that we got two sets of the same three random values.
</para>
<para>
This example uses <literal>WITH RECURSIVE</literal> to find all
subordinates (direct or indirect) of the employee Mary, and their
level of indirectness, from a table that shows only direct
subordinates:
<programlisting>
WITH RECURSIVE employee_recursive(distance, employee_name, manager_name) AS (
SELECT 1, employee_name, manager_name
FROM employee
WHERE manager_name = 'Mary'
UNION ALL
SELECT er.distance + 1, e.employee_name, e.manager_name
FROM employee_recursive er, employee e
WHERE er.employee_name = e.manager_name
)
SELECT distance, employee_name FROM employee_recursive;
</programlisting>
Notice the typical form of recursive queries:
an initial condition, followed by <literal>UNION</literal>,
followed by the recursive part of the query. Be sure that the
recursive part of the query will eventually return no tuples, or
else the query will loop indefinitely. (See <xref linkend="queries-with">
for more examples.)
</para>
<para>
This example uses <literal>LATERAL</> to apply a set-returning function
<function>get_product_names()</> for each row of the
<structname>manufacturers</> table:
<programlisting>
SELECT m.name AS mname, pname
FROM manufacturers m, LATERAL get_product_names(m.id) pname;
</programlisting>
Manufacturers not currently having any products would not appear in the
result, since it is an inner join. If we wished to include the names of
such manufacturers in the result, we could do:
<programlisting>
SELECT m.name AS mname, pname
FROM manufacturers m LEFT JOIN LATERAL get_product_names(m.id) pname ON true;
</programlisting></para>
</refsect1>
<refsect1>
<title>Compatibility</title>
<para>
Of course, the <command>SELECT</command> statement is compatible
with the SQL standard. But there are some extensions and some
missing features.
</para>
<refsect2>
<title>Omitted <literal>FROM</literal> Clauses</title>
<para>
<productname>PostgreSQL</productname> allows one to omit the
<literal>FROM</literal> clause. It has a straightforward use to
compute the results of simple expressions:
<programlisting>
SELECT 2+2;
?column?
----------
4
</programlisting>
Some other <acronym>SQL</acronym> databases cannot do this except
by introducing a dummy one-row table from which to do the
<command>SELECT</command>.
</para>
<para>
Note that if a <literal>FROM</literal> clause is not specified,
the query cannot reference any database tables. For example, the
following query is invalid:
<programlisting>
SELECT distributors.* WHERE distributors.name = 'Westward';
</programlisting>
<productname>PostgreSQL</productname> releases prior to
8.1 would accept queries of this form, and add an implicit entry
to the query's <literal>FROM</literal> clause for each table
referenced by the query. This is no longer allowed.
</para>
</refsect2>
<refsect2>
<title>Empty <literal>SELECT</literal> Lists</title>
<para>
The list of output expressions after <literal>SELECT</literal> can be
empty, producing a zero-column result table.
This is not valid syntax according to the SQL standard.
<productname>PostgreSQL</productname> allows it to be consistent with
allowing zero-column tables.
However, an empty list is not allowed when <literal>DISTINCT</> is used.
</para>
</refsect2>
<refsect2>
<title>Omitting the <literal>AS</literal> Key Word</title>
<para>
In the SQL standard, the optional key word <literal>AS</> can be
omitted before an output column name whenever the new column name
is a valid column name (that is, not the same as any reserved
keyword). <productname>PostgreSQL</productname> is slightly more
restrictive: <literal>AS</> is required if the new column name
matches any keyword at all, reserved or not. Recommended practice is
to use <literal>AS</> or double-quote output column names, to prevent
any possible conflict against future keyword additions.
</para>
<para>
In <literal>FROM</literal> items, both the standard and
<productname>PostgreSQL</productname> allow <literal>AS</> to
be omitted before an alias that is an unreserved keyword. But
this is impractical for output column names, because of syntactic
ambiguities.
</para>
</refsect2>
<refsect2>
<title><literal>ONLY</literal> and Inheritance</title>
<para>
The SQL standard requires parentheses around the table name when
writing <literal>ONLY</literal>, for example <literal>SELECT * FROM ONLY
(tab1), ONLY (tab2) WHERE ...</literal>. <productname>PostgreSQL</>
considers these parentheses to be optional.
</para>
<para>
<productname>PostgreSQL</> allows a trailing <literal>*</> to be written to
explicitly specify the non-<literal>ONLY</literal> behavior of including
child tables. The standard does not allow this.
</para>
<para>
(These points apply equally to all SQL commands supporting the
<literal>ONLY</literal> option.)
</para>
</refsect2>
<refsect2>
<title>Function Calls in <literal>FROM</literal></title>
<para>
<productname>PostgreSQL</productname> allows a function call to be
written directly as a member of the <literal>FROM</> list. In the SQL
standard it would be necessary to wrap such a function call in a
sub-<command>SELECT</command>; that is, the syntax
<literal>FROM <replaceable>func</>(...) <replaceable>alias</></literal>
is approximately equivalent to
<literal>FROM LATERAL (SELECT <replaceable>func</>(...)) <replaceable>alias</></literal>.
Note that <literal>LATERAL</> is considered to be implicit; this is
because the standard requires <literal>LATERAL</> semantics for an
<literal>UNNEST()</> item in <literal>FROM</>.
<productname>PostgreSQL</productname> treats <literal>UNNEST()</> the
same as other set-returning functions.
</para>
</refsect2>
<refsect2>
<title>Namespace Available to <literal>GROUP BY</literal> and <literal>ORDER BY</literal></title>
<para>
Update reference documentation on may/can/might:
Standard English uses "may", "can", and "might" in different ways:
may - permission, "You may borrow my rake."
can - ability, "I can lift that log."
might - possibility, "It might rain today."
Unfortunately, in conversational English, their use is often mixed, as
in, "You may use this variable to do X", when in fact, "can" is a better
choice. Similarly, "It may crash" is better stated, "It might crash".
19 years ago
In the SQL-92 standard, an <literal>ORDER BY</literal> clause can
only use output column names or numbers, while a <literal>GROUP
Update reference documentation on may/can/might:
Standard English uses "may", "can", and "might" in different ways:
may - permission, "You may borrow my rake."
can - ability, "I can lift that log."
might - possibility, "It might rain today."
Unfortunately, in conversational English, their use is often mixed, as
in, "You may use this variable to do X", when in fact, "can" is a better
choice. Similarly, "It may crash" is better stated, "It might crash".
19 years ago
BY</literal> clause can only use expressions based on input column
names. <productname>PostgreSQL</productname> extends each of
these clauses to allow the other choice as well (but it uses the
standard's interpretation if there is ambiguity).
<productname>PostgreSQL</productname> also allows both clauses to
specify arbitrary expressions. Note that names appearing in an
expression will always be taken as input-column names, not as
output-column names.
</para>
<para>
SQL:1999 and later use a slightly different definition which is not
entirely upward compatible with SQL-92.
In most cases, however, <productname>PostgreSQL</productname>
will interpret an <literal>ORDER BY</literal> or <literal>GROUP
BY</literal> expression the same way SQL:1999 does.
</para>
</refsect2>
<refsect2>
<title>Functional Dependencies</title>
<para>
<productname>PostgreSQL</productname> recognizes functional dependency
(allowing columns to be omitted from <literal>GROUP BY</>) only when
a table's primary key is included in the <literal>GROUP BY</> list.
The SQL standard specifies additional conditions that should be
recognized.
</para>
</refsect2>
<refsect2>
<title><literal>WINDOW</literal> Clause Restrictions</title>
<para>
The SQL standard provides additional options for the window
<replaceable class="parameter">frame_clause</>.
<productname>PostgreSQL</productname> currently supports only the
options listed above.
</para>
</refsect2>
<refsect2>
<title><literal>LIMIT</literal> and <literal>OFFSET</literal></title>
<para>
The clauses <literal>LIMIT</literal> and <literal>OFFSET</literal>
are <productname>PostgreSQL</productname>-specific syntax, also
used by <productname>MySQL</productname>. The SQL:2008 standard
has introduced the clauses <literal>OFFSET ... FETCH {FIRST|NEXT}
...</literal> for the same functionality, as shown above
in <xref linkend="sql-limit" endterm="sql-limit-title">. This
syntax is also used by <productname>IBM DB2</productname>.
(Applications written for <productname>Oracle</productname>
frequently use a workaround involving the automatically
generated <literal>rownum</literal> column, which is not available in
PostgreSQL, to implement the effects of these clauses.)
</para>
</refsect2>
<refsect2>
Improve concurrency of foreign key locking
This patch introduces two additional lock modes for tuples: "SELECT FOR
KEY SHARE" and "SELECT FOR NO KEY UPDATE". These don't block each
other, in contrast with already existing "SELECT FOR SHARE" and "SELECT
FOR UPDATE". UPDATE commands that do not modify the values stored in
the columns that are part of the key of the tuple now grab a SELECT FOR
NO KEY UPDATE lock on the tuple, allowing them to proceed concurrently
with tuple locks of the FOR KEY SHARE variety.
Foreign key triggers now use FOR KEY SHARE instead of FOR SHARE; this
means the concurrency improvement applies to them, which is the whole
point of this patch.
The added tuple lock semantics require some rejiggering of the multixact
module, so that the locking level that each transaction is holding can
be stored alongside its Xid. Also, multixacts now need to persist
across server restarts and crashes, because they can now represent not
only tuple locks, but also tuple updates. This means we need more
careful tracking of lifetime of pg_multixact SLRU files; since they now
persist longer, we require more infrastructure to figure out when they
can be removed. pg_upgrade also needs to be careful to copy
pg_multixact files over from the old server to the new, or at least part
of multixact.c state, depending on the versions of the old and new
servers.
Tuple time qualification rules (HeapTupleSatisfies routines) need to be
careful not to consider tuples with the "is multi" infomask bit set as
being only locked; they might need to look up MultiXact values (i.e.
possibly do pg_multixact I/O) to find out the Xid that updated a tuple,
whereas they previously were assured to only use information readily
available from the tuple header. This is considered acceptable, because
the extra I/O would involve cases that would previously cause some
commands to block waiting for concurrent transactions to finish.
Another important change is the fact that locking tuples that have
previously been updated causes the future versions to be marked as
locked, too; this is essential for correctness of foreign key checks.
This causes additional WAL-logging, also (there was previously a single
WAL record for a locked tuple; now there are as many as updated copies
of the tuple there exist.)
With all this in place, contention related to tuples being checked by
foreign key rules should be much reduced.
As a bonus, the old behavior that a subtransaction grabbing a stronger
tuple lock than the parent (sub)transaction held on a given tuple and
later aborting caused the weaker lock to be lost, has been fixed.
Many new spec files were added for isolation tester framework, to ensure
overall behavior is sane. There's probably room for several more tests.
There were several reviewers of this patch; in particular, Noah Misch
and Andres Freund spent considerable time in it. Original idea for the
patch came from Simon Riggs, after a problem report by Joel Jacobson.
Most code is from me, with contributions from Marti Raudsepp, Alexander
Shulgin, Noah Misch and Andres Freund.
This patch was discussed in several pgsql-hackers threads; the most
important start at the following message-ids:
AANLkTimo9XVcEzfiBR-ut3KVNDkjm2Vxh+t8kAmWjPuv@mail.gmail.com
1290721684-sup-3951@alvh.no-ip.org
1294953201-sup-2099@alvh.no-ip.org
1320343602-sup-2290@alvh.no-ip.org
1339690386-sup-8927@alvh.no-ip.org
4FE5FF020200002500048A3D@gw.wicourts.gov
4FEAB90A0200002500048B7D@gw.wicourts.gov
13 years ago
<title><literal>FOR NO KEY UPDATE</>, <literal>FOR UPDATE</>, <literal>FOR SHARE</>, <literal>FOR KEY SHARE</></title>
<para>
Although <literal>FOR UPDATE</> appears in the SQL standard, the
standard allows it only as an option of <command>DECLARE CURSOR</>.
<productname>PostgreSQL</productname> allows it in any <command>SELECT</>
query as well as in sub-<command>SELECT</>s, but this is an extension.
Improve concurrency of foreign key locking
This patch introduces two additional lock modes for tuples: "SELECT FOR
KEY SHARE" and "SELECT FOR NO KEY UPDATE". These don't block each
other, in contrast with already existing "SELECT FOR SHARE" and "SELECT
FOR UPDATE". UPDATE commands that do not modify the values stored in
the columns that are part of the key of the tuple now grab a SELECT FOR
NO KEY UPDATE lock on the tuple, allowing them to proceed concurrently
with tuple locks of the FOR KEY SHARE variety.
Foreign key triggers now use FOR KEY SHARE instead of FOR SHARE; this
means the concurrency improvement applies to them, which is the whole
point of this patch.
The added tuple lock semantics require some rejiggering of the multixact
module, so that the locking level that each transaction is holding can
be stored alongside its Xid. Also, multixacts now need to persist
across server restarts and crashes, because they can now represent not
only tuple locks, but also tuple updates. This means we need more
careful tracking of lifetime of pg_multixact SLRU files; since they now
persist longer, we require more infrastructure to figure out when they
can be removed. pg_upgrade also needs to be careful to copy
pg_multixact files over from the old server to the new, or at least part
of multixact.c state, depending on the versions of the old and new
servers.
Tuple time qualification rules (HeapTupleSatisfies routines) need to be
careful not to consider tuples with the "is multi" infomask bit set as
being only locked; they might need to look up MultiXact values (i.e.
possibly do pg_multixact I/O) to find out the Xid that updated a tuple,
whereas they previously were assured to only use information readily
available from the tuple header. This is considered acceptable, because
the extra I/O would involve cases that would previously cause some
commands to block waiting for concurrent transactions to finish.
Another important change is the fact that locking tuples that have
previously been updated causes the future versions to be marked as
locked, too; this is essential for correctness of foreign key checks.
This causes additional WAL-logging, also (there was previously a single
WAL record for a locked tuple; now there are as many as updated copies
of the tuple there exist.)
With all this in place, contention related to tuples being checked by
foreign key rules should be much reduced.
As a bonus, the old behavior that a subtransaction grabbing a stronger
tuple lock than the parent (sub)transaction held on a given tuple and
later aborting caused the weaker lock to be lost, has been fixed.
Many new spec files were added for isolation tester framework, to ensure
overall behavior is sane. There's probably room for several more tests.
There were several reviewers of this patch; in particular, Noah Misch
and Andres Freund spent considerable time in it. Original idea for the
patch came from Simon Riggs, after a problem report by Joel Jacobson.
Most code is from me, with contributions from Marti Raudsepp, Alexander
Shulgin, Noah Misch and Andres Freund.
This patch was discussed in several pgsql-hackers threads; the most
important start at the following message-ids:
AANLkTimo9XVcEzfiBR-ut3KVNDkjm2Vxh+t8kAmWjPuv@mail.gmail.com
1290721684-sup-3951@alvh.no-ip.org
1294953201-sup-2099@alvh.no-ip.org
1320343602-sup-2290@alvh.no-ip.org
1339690386-sup-8927@alvh.no-ip.org
4FE5FF020200002500048A3D@gw.wicourts.gov
4FEAB90A0200002500048B7D@gw.wicourts.gov
13 years ago
The <literal>FOR NO KEY UPDATE</>, <literal>FOR SHARE</> and
<literal>FOR KEY SHARE</> variants,
as well as the <literal>NOWAIT</> option,
do not appear in the standard.
</para>
</refsect2>
<refsect2>
<title>Data-Modifying Statements in <literal>WITH</></title>
<para>
<productname>PostgreSQL</productname> allows <command>INSERT</>,
<command>UPDATE</>, and <command>DELETE</> to be used as <literal>WITH</>
queries. This is not found in the SQL standard.
</para>
</refsect2>
<refsect2>
<title>Nonstandard Clauses</title>
<para>
<literal>DISTINCT ON ( ... )</literal> is an extension of the
SQL standard.
</para>
<para>
<literal>ROWS FROM( ... )</> is an extension of the SQL standard.
</para>
</refsect2>
</refsect1>
</refentry>
