It's always been possible for index AMs to cache data across successive
amgettuple calls within a single SQL command: the IndexScanDesc.opaque
field is meant for precisely that. However, no comparable facility
exists for amortizing setup work across successive aminsert calls.
This patch adds such a feature and teaches GIN, GIST, and BRIN to use it
to amortize catalog lookups they'd previously been doing on every call.
(The other standard index AMs keep everything they need in the relcache,
so there's little to improve there.)
For GIN, the overall improvement in a statement that inserts many rows
can be as much as 10%, though it seems a bit less for the other two.
In addition, this makes a really significant difference in runtime
for CLOBBER_CACHE_ALWAYS tests, since in those builds the repeated
catalog lookups are vastly more expensive.
The reason this has been hard up to now is that the aminsert function is
not passed any useful place to cache per-statement data. What I chose to
do is to add suitable fields to struct IndexInfo and pass that to aminsert.
That's not widening the index AM API very much because IndexInfo is already
within the ken of ambuild; in fact, by passing the same info to aminsert
as to ambuild, this is really removing an inconsistency in the AM API.
Discussion: https://postgr.es/m/27568.1486508680@sss.pgh.pa.us
This is infrastructure for a pending patch to allow parallel bitmap
heap scans.
Dilip Kumar, reviewed (in earlier versions) by Andres Freund and
(more recently) by me. Some further renaming by me, also.
When I wrote commit ab1f0c822, I really missed the castNode() macro that
Peter E. had proposed shortly before. This back-fills the uses I would
have put it to. It's probably not all that significant, but there are
more assertions here than there were before, and conceivably they will
help catch any bugs associated with those representation changes.
I left behind a number of usages like "(Query *) copyObject(query_var)".
Those could have been converted as well, but Peter has proposed another
notational improvement that would handle copyObject cases automatically,
so I let that be for now.
Previously, ExecInitModifyTable was missing handling for WITH CHECK
OPTION, and view_query_is_auto_updatable was missing handling for
RELKIND_PARTITIONED_TABLE.
Amit Langote, reviewed by me.
In 2ac3ef7a01, we changed things so that
it's possible for a different TupleTableSlot to be used for partitioned
tables at successively lower levels. If we do end up changing the slot
from the original, we must update ecxt_scantuple to point to the new one
for partition key of the tuple to be computed correctly.
Reported by Rajkumar Raghuwanshi. Patch by Amit Langote.
Discussion: http://postgr.es/m/CAKcux6%3Dm1qyqB2k6cjniuMMrYXb75O-MB4qGQMu8zg-iGGLjDw%40mail.gmail.com
We've accumulated quite a bit of stuff with which pgindent is not
quite happy in this code; clean it up to provide a less-annoying base
for future pgindent runs.
Coverity complained quite properly that commit ea15e1867 had introduced
unreachable code into ExecGather(); to wit, it was no longer possible to
iterate the final for-loop more or less than once. So remove the for().
In passing, clean up a couple of comments, and make better use of a local
variable.
Since 69f4b9c plain expression evaluation (and thus normal projection)
can't return sets of tuples anymore. Thus remove code dealing with
that possibility.
This will require adjustments in external code using
ExecEvalExpr()/ExecProject() - that should neither be hard nor very
common.
Author: Andres Freund and Tom Lane
Discussion: https://postgr.es/m/20160822214023.aaxz5l4igypowyri@alap3.anarazel.de
In ExecInsert(), do not switch back to the root partitioned table
ResultRelInfo until after we finish ExecProcessReturning(), so that
RETURNING projection is done using the partition's descriptor. For
the projection to work correctly, we must initialize the same for each
leaf partition during ModifyTableState initialization.
Amit Langote
When a tuple is inherited into a partitioning root, no partition
constraints need to be enforced; when it is inserted into a leaf, the
parent's partitioning quals needed to be enforced. The previous
coding got both of those cases right. When a tuple is inserted into
an intermediate level of the partitioning hierarchy (i.e. a table
which is both a partition itself and in turn partitioned), it must
enforce the partitioning qual inherited from its parent. That case
got overlooked; repair.
Amit Langote
Evaluation of set returning functions (SRFs_ in the targetlist (like SELECT
generate_series(1,5)) so far was done in the expression evaluation (i.e.
ExecEvalExpr()) and projection (i.e. ExecProject/ExecTargetList) code.
This meant that most executor nodes performing projection, and most
expression evaluation functions, had to deal with the possibility that an
evaluated expression could return a set of return values.
That's bad because it leads to repeated code in a lot of places. It also,
and that's my (Andres's) motivation, made it a lot harder to implement a
more efficient way of doing expression evaluation.
To fix this, introduce a new executor node (ProjectSet) that can evaluate
targetlists containing one or more SRFs. To avoid the complexity of the old
way of handling nested expressions returning sets (e.g. having to pass up
ExprDoneCond, and dealing with arguments to functions returning sets etc.),
those SRFs can only be at the top level of the node's targetlist. The
planner makes sure (via split_pathtarget_at_srfs()) that SRF evaluation is
only necessary in ProjectSet nodes and that SRFs are only present at the
top level of the node's targetlist. If there are nested SRFs the planner
creates multiple stacked ProjectSet nodes. The ProjectSet nodes always get
input from an underlying node.
We also discussed and prototyped evaluating targetlist SRFs using ROWS
FROM(), but that turned out to be more complicated than we'd hoped.
While moving SRF evaluation to ProjectSet would allow to retain the old
"least common multiple" behavior when multiple SRFs are present in one
targetlist (i.e. continue returning rows until all SRFs are at the end of
their input at the same time), we decided to instead only return rows till
all SRFs are exhausted, returning NULL for already exhausted ones. We
deemed the previous behavior to be too confusing, unexpected and actually
not particularly useful.
As a side effect, the previously prohibited case of multiple set returning
arguments to a function, is now allowed. Not because it's particularly
desirable, but because it ends up working and there seems to be no argument
for adding code to prohibit it.
Currently the behavior for COALESCE and CASE containing SRFs has changed,
returning multiple rows from the expression, even when the SRF containing
"arm" of the expression is not evaluated. That's because the SRFs are
evaluated in a separate ProjectSet node. As that's quite confusing, we're
likely to instead prohibit SRFs in those places. But that's still being
discussed, and the code would reside in places not touched here, so that's
a task for later.
There's a lot of, now superfluous, code dealing with set return expressions
around. But as the changes to get rid of those are verbose largely boring,
it seems better for readability to keep the cleanup as a separate commit.
Author: Tom Lane and Andres Freund
Discussion: https://postgr.es/m/20160822214023.aaxz5l4igypowyri@alap3.anarazel.de
Gen_fmgrtab.pl creates a new file fmgrprotos.h, which contains
prototypes for all functions registered in pg_proc.h. This avoids
having to manually maintain these prototypes across a random variety of
header files. It also automatically enforces a correct function
signature, and since there are warnings about missing prototypes, it
will detect functions that are defined but not registered in
pg_proc.h (or otherwise used).
Reviewed-by: Pavel Stehule <pavel.stehule@gmail.com>
This patch makes several changes that improve the consistency of
representation of lists of statements. It's always been the case
that the output of parse analysis is a list of Query nodes, whatever
the types of the individual statements in the list. This patch brings
similar consistency to the outputs of raw parsing and planning steps:
* The output of raw parsing is now always a list of RawStmt nodes;
the statement-type-dependent nodes are one level down from that.
* The output of pg_plan_queries() is now always a list of PlannedStmt
nodes, even for utility statements. In the case of a utility statement,
"planning" just consists of wrapping a CMD_UTILITY PlannedStmt around
the utility node. This list representation is now used in Portal and
CachedPlan plan lists, replacing the former convention of intermixing
PlannedStmts with bare utility-statement nodes.
Now, every list of statements has a consistent head-node type depending
on how far along it is in processing. This allows changing many places
that formerly used generic "Node *" pointers to use a more specific
pointer type, thus reducing the number of IsA() tests and casts needed,
as well as improving code clarity.
Also, the post-parse-analysis representation of DECLARE CURSOR is changed
so that it looks more like EXPLAIN, PREPARE, etc. That is, the contained
SELECT remains a child of the DeclareCursorStmt rather than getting flipped
around to be the other way. It's now true for both Query and PlannedStmt
that utilityStmt is non-null if and only if commandType is CMD_UTILITY.
That allows simplifying a lot of places that were testing both fields.
(I think some of those were just defensive programming, but in many places,
it was actually necessary to avoid confusing DECLARE CURSOR with SELECT.)
Because PlannedStmt carries a canSetTag field, we're also able to get rid
of some ad-hoc rules about how to reconstruct canSetTag for a bare utility
statement; specifically, the assumption that a utility is canSetTag if and
only if it's the only one in its list. While I see no near-term need for
relaxing that restriction, it's nice to get rid of the ad-hocery.
The API of ProcessUtility() is changed so that what it's passed is the
wrapper PlannedStmt not just the bare utility statement. This will affect
all users of ProcessUtility_hook, but the changes are pretty trivial; see
the affected contrib modules for examples of the minimum change needed.
(Most compilers should give pointer-type-mismatch warnings for uncorrected
code.)
There's also a change in the API of ExplainOneQuery_hook, to pass through
cursorOptions instead of expecting hook functions to know what to pick.
This is needed because of the DECLARE CURSOR changes, but really should
have been done in 9.6; it's unlikely that any extant hook functions
know about using CURSOR_OPT_PARALLEL_OK.
Finally, teach gram.y to save statement boundary locations in RawStmt
nodes, and pass those through to Query and PlannedStmt nodes. This allows
more intelligent handling of cases where a source query string contains
multiple statements. This patch doesn't actually do anything with the
information, but a follow-on patch will. (Passing this information through
cleanly is the true motivation for these changes; while I think this is all
good cleanup, it's unlikely we'd have bothered without this end goal.)
catversion bump because addition of location fields to struct Query
affects stored rules.
This patch is by me, but it owes a good deal to Fabien Coelho who did
a lot of preliminary work on the problem, and also reviewed the patch.
Discussion: https://postgr.es/m/alpine.DEB.2.20.1612200926310.29821@lancre
A client copy can't work inside a function because the FE/BE wire protocol
doesn't support nesting of a COPY operation within query results. (Maybe
it could, but the protocol spec doesn't suggest that clients should support
this, and libpq for one certainly doesn't.)
In most PLs, this prohibition is enforced by spi.c, but SQL functions don't
use SPI. A comparison of _SPI_execute_plan() and init_execution_state()
shows that rejecting client COPY is the only discrepancy in what they
allow, so there's no other similar bugs.
This is an astonishingly ancient oversight, so back-patch to all supported
branches.
Report: https://postgr.es/m/BY2PR05MB2309EABA3DEFA0143F50F0D593780@BY2PR05MB2309.namprd05.prod.outlook.com
This commit purported to use a variable hash seed for Partial
HashAggregate, but actually did the opposite - it made us use a
variable seed for any HashAggregate that is NOT partial. Woops.
With the old code, a backend that read pg_stat_activity without ever
having executed a parallel query might see a backend in the midst of
executing one waiting on a DSA LWLock, resulting in a crash. The
solution is for backends to register the tranche at startup time, not
the first time a parallel query is executed.
Report by Andreas Seltenreich. Patch by me, reviewed by Thomas Munro.
RelationGetPartitionQual() and generate_partition_qual() are always
called with recurse = true, so we don't need an argument for that.
Extracted by me from a larger patch by Amit Langote.
After a tuple is routed to a partition, it has been converted from the
root table's row type to the partition's row type. ExecConstraints
needs to report the failure using the original tuple and the parent's
tuple descriptor rather than the ones for the selected partition.
Amit Langote
Commit 2ac3ef7a01 added a TupleTapleSlot
for partition tuple slot to EState (es_partition_tuple_slot) but it's
more logical to have it as part of ModifyTableState
(mt_partition_tuple_slot) and CopyState (partition_tuple_slot).
Discussion: http://postgr.es/m/1bd459d9-4c0c-197a-346e-e5e59e217d97@lab.ntt.co.jp
Amit Langote, per a gripe from me
I got a little annoyed by reading documentation paragraphs containing
both spellings within a few lines of each other. My dictionary says
"descendant" is the preferred spelling, and it's certainly the majority
usage in our tree, so standardize on that.
For one usage in parallel.sgml, I thought it better to rewrite to avoid
the term altogether.
The previous coding failed to work correctly when we have a
multi-level partitioned hierarchy where tables at successive levels
have different attribute numbers for the partition key attributes. To
fix, have each PartitionDispatch object store a standalone
TupleTableSlot initialized with the TupleDesc of the corresponding
partitioned table, along with a TupleConversionMap to map tuples from
the its parent's rowtype to own rowtype. After tuple routing chooses
a leaf partition, we must use the leaf partition's tuple descriptor,
not the root table's. To that end, a dedicated TupleTableSlot for
tuple routing is now allocated in EState.
Amit Langote
When the input value to a CoerceToDomain expression node is a read-write
expanded datum, we should pass a read-only pointer to any domain CHECK
expressions and then return the original read-write pointer as the
expression result. Previously we were blindly passing the same pointer to
all the consumers of the value, making it possible for a function in CHECK
to modify or even delete the expanded value. (Since a plpgsql function
will absorb a passed-in read-write expanded array as a local variable
value, it will in fact delete the value on exit.)
A similar hazard of passing the same read-write pointer to multiple
consumers exists in domain_check() and in ExecEvalCase, so fix those too.
The fix requires adding MakeExpandedObjectReadOnly calls at the appropriate
places, which is simple enough except that we need to get the data type's
typlen from somewhere. For the domain cases, solve this by redefining
DomainConstraintRef.tcache as okay for callers to access; there wasn't any
reason for the original convention against that, other than not wanting the
API of typcache.c to be any wider than it had to be. For CASE, there's
no good solution except to add a syscache lookup during executor start.
Per bug #14472 from Marcos Castedo. Back-patch to 9.5 where expanded
values were introduced.
Discussion: https://postgr.es/m/15225.1482431619@sss.pgh.pa.us
aggstate->evalproj is always set up by ExecInitAgg, so there's no
need to test. Doing so led Coverity to think that we might be
intending "slot" to be possibly NULL here, and it quite properly
complained that the rest of combine_aggregates() wasn't prepared
for that.
Also fix a couple of obvious thinkos in Asserts checking that
"inputoff" isn't past the end of the slot.
Errors introduced in commit 8ed3f11bb, so no need for back-patch.
If a query contained two aggregates that could share the transition value,
we would correctly collect the input into a tuplesort only once, but
incorrectly run the transition function over the accumulated input twice,
in finalize_aggregates(). That caused a crash, when we tried to call
tuplesort_performsort() on an already-freed NULL tuplestore.
Backport to 9.6, where sharing of transition state and this bug were
introduced.
Analysis by Tom Lane.
Discussion: https://www.postgresql.org/message-id/ac5b0b69-744c-9114-6218-8300ac920e61@iki.fi
Commit 5dfc198146 introduced the use
of a new type of hash table with linear reprobing for hash aggregates.
Such a hash table behaves very poorly if keys are inserted in hash
order, which does in fact happen in the case where a query use a
Finalize HashAggregate node fed (via Gather) by a Partial
HashAggregate node. In fact, queries with this type of plan tend
to run effectively forever.
Fix that by seeding the hash value differently in each worker
(and in the leader, if it participates).
Andres Freund and Robert Haas
Commit 4212cb7326 rendered a comment
in execMain.c incorrect. Per complaint from Tom Lane, repair.
Patch from Amit Kapila, per wording suggested by Tom Lane and me.
Table partitioning is like table inheritance and reuses much of the
existing infrastructure, but there are some important differences.
The parent is called a partitioned table and is always empty; it may
not have indexes or non-inherited constraints, since those make no
sense for a relation with no data of its own. The children are called
partitions and contain all of the actual data. Each partition has an
implicit partitioning constraint. Multiple inheritance is not
allowed, and partitioning and inheritance can't be mixed. Partitions
can't have extra columns and may not allow nulls unless the parent
does. Tuples inserted into the parent are automatically routed to the
correct partition, so tuple-routing ON INSERT triggers are not needed.
Tuple routing isn't yet supported for partitions which are foreign
tables, and it doesn't handle updates that cross partition boundaries.
Currently, tables can be range-partitioned or list-partitioned. List
partitioning is limited to a single column, but range partitioning can
involve multiple columns. A partitioning "column" can be an
expression.
Because table partitioning is less general than table inheritance, it
is hoped that it will be easier to reason about properties of
partitions, and therefore that this will serve as a better foundation
for a variety of possible optimizations, including query planner
optimizations. The tuple routing based which this patch does based on
the implicit partitioning constraints is an example of this, but it
seems likely that many other useful optimizations are also possible.
Amit Langote, reviewed and tested by Robert Haas, Ashutosh Bapat,
Amit Kapila, Rajkumar Raghuwanshi, Corey Huinker, Jaime Casanova,
Rushabh Lathia, Erik Rijkers, among others. Minor revisions by me.
Previously, a prepared statement created via a Parse message could get
a parallel plan, but one created with a PREPARE statement could not.
This state of affairs was due to confusion on my (rhaas) part: I
erroneously believed that a CREATE TABLE .. AS EXECUTE statement could
only be performed with a prepared statement by PREPARE, but in fact
one created by a Prepare message works just as well. Therefore, it
makes no sense to allow parallel query in one case but not the other.
To fix, allow parallel query with all prepared statements, but run
the parallel plan serially (i.e. without workers) in the case of
CREATE TABLE .. AS EXECUTE. Also, document this.
Amit Kapila and Tobias Bussman, plus an extra sentence of
documentation by me.
The previously code worked OK as long as a Gather node was never
rescanned, or if it was rescanned, as long as it got at least as
many workers on rescan as it had originally. But if the number
of workers ever decreased on a rescan, then it could crash.
Andreas Seltenreich
So far the hashtable stored representative tuples in the form of its
input slot, with all columns in the hashtable that are not
needed (i.e. not grouped upon or functionally dependent) set to NULL.
Thats good for saving memory, but it turns out that having tuples full
of NULL isn't free. slot_deform_tuple is faster if there's no NULL
bitmap even if no NULLs are encountered, and skipping over leading NULLs
isn't free.
So compute a separate tuple descriptor that only contains the needed
columns. As columns have already been moved in/out the slot for the
hashtable that does not imply additional per-row overhead.
Author: Andres Freund
Reviewed-By: Heikki Linnakangas
Discussion: https://postgr.es/m/20161103110721.h5i5t5saxfk5eeik@alap3.anarazel.de
Previously we did a ExecProject() for each individual aggregate
argument. That turned out to be a performance bottleneck in queries with
multiple aggregates.
Doing all the argument computations in one ExecProject() is quite a bit
cheaper because ExecProject's fastpath can do the work at once in a
relatively tight loop, and because it can get all the required columns
with a single slot_getsomeattr and save some other redundant setup
costs.
Author: Andres Freund
Reviewed-By: Heikki Linnakangas
Discussion: https://postgr.es/m/20161103110721.h5i5t5saxfk5eeik@alap3.anarazel.de
The idea behind SPI_push was to allow transitioning back into an
"unconnected" state when a SPI-using procedure calls unrelated code that
might or might not invoke SPI. That sounds good, but in practice the only
thing it does for us is to catch cases where a called SPI-using function
forgets to call SPI_connect --- which is a highly improbable failure mode,
since it would be exposed immediately by direct testing of said function.
As against that, we've had multiple bugs induced by forgetting to call
SPI_push/SPI_pop around code that might invoke SPI-using functions; these
are much harder to catch and indeed have gone undetected for years in some
cases. And we've had to band-aid around some problems of this ilk by
introducing conditional push/pop pairs in some places, which really kind
of defeats the purpose altogether; if we can't draw bright lines between
connected and unconnected code, what's the point?
Hence, get rid of SPI_push[_conditional], SPI_pop[_conditional], and the
underlying state variable _SPI_curid. It turns out SPI_restore_connection
can go away too, which is a nice side benefit since it was never more than
a kluge. Provide no-op macros for the deleted functions so as to avoid an
API break for external modules.
A side effect of this removal is that SPI_palloc and allied functions no
longer permit being called when unconnected; they'll throw an error
instead. The apparent usefulness of the previous behavior was a mirage
as well, because it was depended on by only a few places (which I fixed in
preceding commits), and it posed a risk of allocations being unexpectedly
long-lived if someone forgot a SPI_push call.
Discussion: <20808.1478481403@sss.pgh.pa.us>
There's basically no scenario where it's sensible for this to match
dropped columns, so put a test for dropped-ness into SPI_fnumber()
itself, and excise the test from the small number of callers that
were paying attention to the case. (Most weren't :-(.)
In passing, normalize tests at call sites: always reject attnum <= 0
if we're disallowing system columns. Previously there was a mixture
of "< 0" and "<= 0" tests. This makes no practical difference since
SPI_fnumber() never returns 0, but I'm feeling pedantic today.
Also, in the places that are actually live user-facing code and not
legacy cruft, distinguish "column not found" from "can't handle
system column".
Per discussion with Jim Nasby; thi supersedes his original patch
that just changed the behavior at one call site.
Discussion: <b2de8258-c4c0-1cb8-7b97-e8538e5c975c@BlueTreble.com>
In the previous coding, if an aggregate's transition function returned an
expanded array, nodeAgg.c and nodeWindowAgg.c would always copy it and thus
force it into the flat representation. This led to ping-ponging between
flat and expanded formats, which costs a lot. For an aggregate using
array_append as transition function, I measured about a 15X slowdown
compared to the pre-9.5 code, when working on simple int[] arrays.
Of course, the old code was already O(N^2) in this usage due to copying
flat arrays all the time, but it wasn't quite this inefficient.
To fix, teach nodeAgg.c and nodeWindowAgg.c to allow expanded transition
values without copying, so long as the transition function takes care to
return the transition value already properly parented under the aggcontext.
That puts a bit of extra responsibility on the transition function, but
doing it this way allows us to not need any extra logic in the fast path
of advance_transition_function (ie, with a pass-by-value transition value,
or with a modified-in-place pass-by-reference value). We already know
that that's a hot spot so I'm loath to add any cycles at all there. Also,
while only array_append currently knows how to follow this convention,
this solution allows other transition functions to opt-in without needing
to have a whitelist in the core aggregation code.
(The reason we would need a whitelist is that currently, if you pass a
R/W expanded-object pointer to an arbitrary function, it's allowed to do
anything with it including deleting it; that breaks the core agg code's
assumption that it should free discarded values. Returning a value under
aggcontext is the transition function's signal that it knows it is an
aggregate transition function and will play nice. Possibly the API rules
for expanded objects should be refined, but that would not be a
back-patchable change.)
With this fix, an aggregate using array_append is no longer O(N^2), so it's
much faster than pre-9.5 code rather than much slower. It's still a bit
slower than the bespoke infrastructure for array_agg, but the differential
seems to be only about 10%-20% rather than orders of magnitude.
Discussion: <6315.1477677885@sss.pgh.pa.us>
INSERT ... ON CONFLICT (specifically ExecCheckHeapTupleVisible) contains
another example of this unsafe coding practice. It is much harder to get
a failure out of it than the case fixed in commit 6292c2339, because in
most scenarios any hint bits that could be set would have already been set
earlier in the command. However, Konstantin Knizhnik reported a failure
with a custom transaction manager, and it's clearly possible to get a
failure via a race condition in async-commit mode.
For lack of a reproducible example, no regression test case in this
commit.
I did some testing with Asserts added to tqual.c's functions, and can say
that running "make check-world" exposed these two bugs and no others.
The Asserts are messy enough that I've not added them to the code for now.
Report: <57EE93C8.8080504@postgrespro.ru>
Related-Discussion: <CAO3NbwOycQjt2Oqy2VW-eLTq2M5uGMyHnGm=RNga4mjqcYD7gQ@mail.gmail.com>
A transaction that conflicts against itself, for example
INSERT INTO t(pk) VALUES (1),(1) ON CONFLICT DO NOTHING;
should behave the same regardless of isolation level. It certainly
shouldn't throw a serialization error, as retrying will not help.
We got this wrong due to the ON CONFLICT logic not considering the case,
as reported by Jason Dusek.
Core of this patch is by Peter Geoghegan (based on an earlier patch by
Thomas Munro), though I didn't take his proposed code refactoring for fear
that it might have unexpected side-effects. Test cases by Thomas Munro
and myself.
Report: <CAO3NbwOycQjt2Oqy2VW-eLTq2M5uGMyHnGm=RNga4mjqcYD7gQ@mail.gmail.com>
Related-Discussion: <57EE93C8.8080504@postgrespro.ru>
Tom Lane