Instead of an exact cost comparison, use a fuzzy comparison with 1e-10
delta after all other path metrics have proved equal. This is to avoid
having platform-specific roundoff behaviors determine the choice when
two paths are really the same to our cost estimators. Adjust the
recently-added test case that made it obvious we had a problem here.
This patch adjusts the treatment of parameterized paths so that all paths
with the same parameterization (same set of required outer rels) for the
same relation will have the same rowcount estimate. We cache the rowcount
estimates to ensure that property, and hopefully save a few cycles too.
Doing this makes it practical for add_path_precheck to operate without
a rowcount estimate: it need only assume that paths with different
parameterizations never dominate each other, which is close enough to
true anyway for coarse filtering, because normally a more-parameterized
path should yield fewer rows thanks to having more join clauses to apply.
In add_path, we do the full nine yards of comparing rowcount estimates
along with everything else, so that we can discard parameterized paths that
don't actually have an advantage. This fixes some issues I'd found with
add_path rejecting parameterized paths on the grounds that they were more
expensive than not-parameterized ones, even though they yielded many fewer
rows and hence would be cheaper once subsequent joining was considered.
To make the same-rowcounts assumption valid, we have to require that any
parameterized path enforce *all* join clauses that could be obtained from
the particular set of outer rels, even if not all of them are useful for
indexing. This is required at both base scans and joins. It's a good
thing anyway since the net impact is that join quals are checked at the
lowest practical level in the join tree. Hence, discard the original
rather ad-hoc mechanism for choosing parameterization joinquals, and build
a better one that has a more principled rule for when clauses can be moved.
The original rule was actually buggy anyway for lack of knowledge about
which relations are part of an outer join's outer side; getting this right
requires adding an outer_relids field to RestrictInfo.
Further reflection shows that a single callback isn't very workable if we
desire to let FDWs generate multiple Paths, because that forces the FDW to
do all work necessary to generate a valid Plan node for each Path. Instead
split the former PlanForeignScan API into three steps: GetForeignRelSize,
GetForeignPaths, GetForeignPlan. We had already bit the bullet of breaking
the 9.1 FDW API for 9.2, so this shouldn't cause very much additional pain,
and it's substantially more flexible for complex FDWs.
Add an fdw_private field to RelOptInfo so that the new functions can save
state there rather than possibly having to recalculate information two or
three times.
In addition, we'd not thought through what would be needed to allow an FDW
to set up subexpressions of its choice for runtime execution. We could
treat ForeignScan.fdw_private as an executable expression but that seems
likely to break existing FDWs unnecessarily (in particular, it would
restrict the set of node types allowable in fdw_private to those supported
by expression_tree_walker). Instead, invent a separate field fdw_exprs
which will receive the postprocessing appropriate for expression trees.
(One field is enough since it can be a list of expressions; also, we assume
the corresponding expression state tree(s) will be held within fdw_state,
so we don't need to add anything to ForeignScanState.)
Per review of Hanada Shigeru's pgsql_fdw patch. We may need to tweak this
further as we continue to work on that patch, but to me it feels a lot
closer to being right now.
The original API specification only allowed an FDW to create a single
access path, which doesn't seem like a terribly good idea in hindsight.
Instead, move the responsibility for building the Path node and calling
add_path() into the FDW's PlanForeignScan function. Now, it can do that
more than once if appropriate. There is no longer any need for the
transient FdwPlan struct, so get rid of that.
Etsuro Fujita, Shigeru Hanada, Tom Lane
This patch fixes the planner so that it can generate nestloop-with-
inner-indexscan plans even with one or more levels of joining between
the indexscan and the nestloop join that is supplying the parameter.
The executor was fixed to handle such cases some time ago, but the
planner was not ready. This should improve our plans in many situations
where join ordering restrictions formerly forced complete table scans.
There is probably a fair amount of tuning work yet to be done, because
of various heuristics that have been added to limit the number of
parameterized paths considered. However, we are not going to find out
what needs to be adjusted until the code gets some real-world use, so
it's time to get it in there where it can be tested easily.
Note API change for index AM amcostestimate functions. I'm not aware of
any non-core index AMs, but if there are any, they will need minor
adjustments.
In commit e2c2c2e8b1 I made use of nested
list structures to show which clauses went with which index columns, but
on reflection that's a data structure that only an old-line Lisp hacker
could love. Worse, it adds unnecessary complication to the many places
that don't much care which clauses go with which index columns. Revert
to the previous arrangement of flat lists of clauses, and instead add a
parallel integer list of column numbers. The places that care about the
pairing can chase both lists with forboth(), while the places that don't
care just examine one list the same as before.
The only real downside to this is that there are now two more lists that
need to be passed to amcostestimate functions in case they care about
column matching (which btcostestimate does, so not passing the info is not
an option). Rather than deal with 11-argument amcostestimate functions,
pass just the IndexPath and expect the functions to extract fields from it.
That gets us down to 7 arguments which is better than 11, and it seems
more future-proof against likely additions to the information we keep
about an index path.
It's potentially useful for an index to repeat the same indexable column
or expression in multiple index columns, if the columns have different
opclasses. (If they share opclasses too, the duplicate column is pretty
useless, but nonetheless we've allowed such cases since 9.0.) However,
the planner failed to cope with this, because createplan.c was relying on
simple equal() matching to figure out which index column each index qual
is intended for. We do have that information available upstream in
indxpath.c, though, so the fix is to not flatten the multi-level indexquals
list when putting it into an IndexPath. Then we can rely on the sublist
structure to identify target index columns in createplan.c. There's a
similar issue for index ORDER BYs (the KNNGIST feature), so introduce a
multi-level-list representation for that too. This adds a bit more
representational overhead, but we might more or less buy that back by not
having to search for matching index columns anymore in createplan.c;
likewise btcostestimate saves some cycles.
Per bug #6351 from Christian Rudolph. Likely symptoms include the "btree
index keys must be ordered by attribute" failure shown there, as well as
"operator MMMM is not a member of opfamily NNNN".
Although this is a pre-existing problem that can be demonstrated in 9.0 and
9.1, I'm not going to back-patch it, because the API changes in the planner
seem likely to break things such as index plugins. The corner cases where
this matters seem too narrow to justify possibly breaking things in a minor
release.
If the right-hand side of a semijoin is unique, then we can treat it like a
normal join (or another way to say that is: we don't need to explicitly
unique-ify the data before doing it as a normal join). We were recognizing
such cases when the RHS was a sub-query with appropriate DISTINCT or GROUP
BY decoration, but there's another way: if the RHS is a plain relation with
unique indexes, we can check if any of the indexes prove the output is
unique. Most of the infrastructure for that was there already in the join
removal code, though I had to rearrange it a bit. Per reflection about a
recent example in pgsql-performance.
This commit changes index-only scans so that data is read directly from the
index tuple without first generating a faux heap tuple. The only immediate
benefit is that indexes on system columns (such as OID) can be used in
index-only scans, but this is necessary infrastructure if we are ever to
support index-only scans on expression indexes. The executor is now ready
for that, though the planner still needs substantial work to recognize
the possibility.
To do this, Vars in index-only plan nodes have to refer to index columns
not heap columns. I introduced a new special varno, INDEX_VAR, to mark
such Vars to avoid confusion. (In passing, this commit renames the two
existing special varnos to OUTER_VAR and INNER_VAR.) This allows
ruleutils.c to handle them with logic similar to what we use for subplan
reference Vars.
Since index-only scans are now fundamentally different from regular
indexscans so far as their expression subtrees are concerned, I also chose
to change them to have their own plan node type (and hence, their own
executor source file).
When a btree index contains all columns required by the query, and the
visibility map shows that all tuples on a target heap page are
visible-to-all, we don't need to fetch that heap page. This patch depends
on the previous patches that made the visibility map reliable.
There's a fair amount left to do here, notably trying to figure out a less
chintzy way of estimating the cost of an index-only scan, but the core
functionality seems ready to commit.
Robert Haas and Ibrar Ahmed, with some previous work by Heikki Linnakangas.
The previous coding failed to account properly for the costs of evaluating
the input expressions of aggregates and window functions, as seen in a
recent gripe from Claudio Freire. (I said at the time that it wasn't
counting these costs at all; but on closer inspection, it was effectively
charging these costs once per output tuple. That is completely wrong for
aggregates, and not exactly right for window functions either.)
There was also a hard-wired assumption that aggregates and window functions
had procost 1.0, which is now fixed to respect the actual cataloged costs.
The costing of WindowAgg is still pretty bogus, since it doesn't try to
estimate the effects of spilling data to disk, but that seems like a
separate issue.
This commit provides the core code and documentation needed. A contrib
module test case will follow shortly.
Shigeru Hanada, Jan Urbanski, Heikki Linnakangas
This is a heavily revised version of builtin_knngist_core-0.9. The
ordering operators are no longer mixed in with actual quals, which would
have confused not only humans but significant parts of the planner.
Instead, ordering operators are carried separately throughout planning and
execution.
Since the API for ambeginscan and amrescan functions had to be changed
anyway, this commit takes the opportunity to rationalize that a bit.
RelationGetIndexScan no longer forces a premature index_rescan call;
instead, callers of index_beginscan must call index_rescan too. Aside from
making the AM-side initialization logic a bit less peculiar, this has the
advantage that we do not make a useless extra am_rescan call when there are
runtime key values. AMs formerly could not assume that the key values
passed to amrescan were actually valid; now they can.
Teodor Sigaev and Tom Lane
Fix things so that top-N sorting can be used in child Sort nodes of a
MergeAppend node, when there is a LIMIT and no intervening joins or
grouping. Actually doing this on the executor side isn't too bad,
but it's a bit messier to get the planner to cost it properly.
Per gripe from Robert Haas.
In passing, fix an oversight in the original top-N-sorting patch:
query_planner should not assume that a LIMIT can be used to make an
explicit sort cheaper when there will be grouping or aggregation in
between. Possibly this should be back-patched, but I'm not sure the
mistake is serious enough to be a real problem in practice.
The core of this patch is hash_array() and associated typcache
infrastructure, which works just about exactly like the existing support
for array comparison.
In addition I did some work to ensure that the planner won't think that an
array type is hashable unless its element type is hashable, and similarly
for sorting. This includes adding a datatype parameter to op_hashjoinable
and op_mergejoinable, and adding an explicit "hashable" flag to
SortGroupClause. The lack of a cross-check on the element type was a
pre-existing bug in mergejoin support --- but it didn't matter so much
before, because if you couldn't sort the element type there wasn't any good
alternative to failing anyhow. Now that we have the alternative of hashing
the array type, there are cases where we can avoid a failure by being picky
at the planner stage, so it's time to be picky.
The issue of exactly how to combine the per-element hash values to produce
an array hash is still open for discussion, but the rest of this is pretty
solid, so I'll commit it as-is.
This patch eliminates the former need to sort the output of an Append scan
when an ordered scan of an inheritance tree is wanted. This should be
particularly useful for fast-start cases such as queries with LIMIT.
Original patch by Greg Stark, with further hacking by Hans-Jurgen Schonig,
Robert Haas, and Tom Lane.
mergejoin to shield it from doing mark/restore and refetches. Put an explicit
flag in MergePath so we can centralize the logic that knows about this,
and add costing logic that considers using Materialize even when it's not
forced by the previously-existing considerations. This is in response to
a discussion back in August that suggested that materializing an inner
indexscan can be helpful when the refetch percentage is high enough.
is unique and is not referenced above the join. In this case the inner
side doesn't affect the query result and can be thrown away entirely.
Although perhaps nobody would ever write such a thing by hand, it's
a reasonably common case in machine-generated SQL.
The current implementation only recognizes the case where the inner side
is a simple relation with a unique index matching the query conditions.
This is enough for the use-cases that have been shown so far, but we
might want to try to handle other cases later.
Robert Haas, somewhat rewritten by Tom
an explicit model of rescan costs being different from first-time costs.
The costing of Material nodes in particular now has some visible relationship
to the actual runtime behavior, where before it was essentially fantasy.
This also fixes up a couple of places where different materialized plan types
were treated differently for no very good reason (probably just oversights).
A couple of the regression tests are affected, because the planner now chooses
to put the other relation on the inside of a nestloop-with-materialize.
So far as I can see both changes are sane, and the planner is now more
consistently following the expectation that it should prefer to materialize
the smaller of two relations.
Per a recent discussion with Robert Haas.
"physical tlist" optimization on the outer relation (ie, force a projection
step to occur in its scan). This avoids storing useless column values when
the outer relation's tuples are written to temporary batch files.
Modified version of a patch by Michael Henderson and Ramon Lawrence.
to be syntactically part of a semijoin clause. For example given
WHERE EXISTS(SELECT ... WHERE upper.var = lower.var AND some-condition)
where some-condition is just a restriction on the lower relation, we can
use unique-ification on lower.var after having applied some-condition within
the scan on lower.
There are some unimplemented aspects: recursive queries must use UNION ALL
(should allow UNION too), and we don't have SEARCH or CYCLE clauses.
These might or might not get done for 8.4, but even without them it's a
pretty useful feature.
There are also a couple of small loose ends and definitional quibbles,
which I'll send a memo about to pgsql-hackers shortly. But let's land
the patch now so we can get on with other development.
Yoshiyuki Asaba, with lots of help from Tatsuo Ishii and Tom Lane
inserting a materialize node above an inner-side sort node, when the sort is
expected to spill to disk. (The materialize protects the sort from having
to support mark/restore, allowing it to do its final merge pass on-the-fly.)
We neglected to teach cost_mergejoin about that hack, so it was failing to
include the materialize's costs in the estimated cost of the mergejoin.
The materialize's costs are generally going to be pretty negligible in
comparison to the sort's, so this is only a small error and probably not
worth back-patching; but it's still wrong.
In the similar case where a materialize is inserted to protect an inner-side
node that can't do mark/restore at all, it's still true that the materialize
should not spill to disk, and so we should cost it cheaply rather than
expensively.
Noted while thinking about a question from Tom Raney.
the old JOIN_IN code, but antijoins are new functionality.) Teach the planner
to convert appropriate EXISTS and NOT EXISTS subqueries into semi and anti
joins respectively. Also, LEFT JOINs with suitable upper-level IS NULL
filters are recognized as being anti joins. Unify the InClauseInfo and
OuterJoinInfo infrastructure into "SpecialJoinInfo". With that change,
it becomes possible to associate a SpecialJoinInfo with every join attempt,
which permits some cleanup of join selectivity estimation. That needs to be
taken much further than this patch does, but the next step is to change the
API for oprjoin selectivity functions, which seems like material for a
separate patch. So for the moment the output size estimates for semi and
especially anti joins are quite bogus.
but seem like a separate patch since most of the remaining work is on the
executor side.) I took the opportunity to push selection of the grouping
operators for set operations into the parser where it belongs. Otherwise this
is just a small exercise in making prepunion.c consider both alternatives.
As with the recent DISTINCT patch, this means we can UNION on datatypes that
can hash but not sort, and it means that UNION without ORDER BY is no longer
certain to produce sorted output.
as per my recent proposal:
1. Fold SortClause and GroupClause into a single node type SortGroupClause.
We were already relying on them to be struct-equivalent, so using two node
tags wasn't accomplishing much except to get in the way of comparing items
with equal().
2. Add an "eqop" field to SortGroupClause to carry the associated equality
operator. This is cheap for the parser to get at the same time it's looking
up the sort operator, and storing it eliminates the need for repeated
not-so-cheap lookups during planning. In future this will also let us
represent GROUP/DISTINCT operations on datatypes that have hash opclasses
but no btree opclasses (ie, they have equality but no natural sort order).
The previous representation simply didn't work for that, since its only
indicator of comparison semantics was a sort operator.
3. Add a hasDistinctOn boolean to struct Query to explicitly record whether
the distinctClause came from DISTINCT or DISTINCT ON. This allows removing
some complicated and not 100% bulletproof code that attempted to figure
that out from the distinctClause alone.
This patch doesn't in itself create any new capability, but it's necessary
infrastructure for future attempts to use hash-based grouping for DISTINCT
and UNION/INTERSECT/EXCEPT.
output is not of the same type that's needed for the IN comparison (ie,
where the parser inserted an implicit coercion above the subselect result).
We should record the coerced expression, not just a raw Var referencing
the subselect output, as the quantity that needs to be unique-ified if
we choose to implement the IN as Unique followed by a plain join.
As of 8.3 this error was causing crashes, as seen in bug #4113 from Javier
Hernandez, because the executor was being told to hash or sort the raw
subselect output column using operators appropriate to the coerced type.
In prior versions there was no crash because the executor chose the
hash or sort operators for itself based on the column type it saw.
However, that's still not really right, because what's unique for one data
type might not be unique for another. In corner cases we could get multiple
outputs of a row that should appear only once, as demonstrated by the
regression test case included in this commit.
However, this patch doesn't apply cleanly to 8.2 or before, and the code
involved has shifted enough over time that I'm hesitant to try to back-patch.
Given the lack of complaints from the field about such corner cases, I think
the bug may not be important enough to risk breaking other things with a
back-patch.
need be returned. We keep a heap of the current best N tuples and sift-up
new tuples into it as we scan the input. For M input tuples this means
only about M*log(N) comparisons instead of M*log(M), not to mention a lot
less workspace when N is small --- avoiding spill-to-disk for large M
is actually the most attractive thing about it. Patch includes planner
and executor support for invoking this facility in ORDER BY ... LIMIT
queries. Greg Stark, with some editorialization by moi.
are mostly excluded by constraints: do the CE test a bit earlier to save
some adjust_appendrel_attrs() work on excluded children, and arrange to
use array indexing rather than rt_fetch() to fetch RTEs in the main body
of the planner. The latter is something I'd wanted to do for awhile anyway,
but seeing list_nth_cell() as 35% of the runtime gets one's attention.
that aren't turned into true joins). Since this is the last missing bit of
infrastructure, go ahead and fill out the hash integer_ops and float_ops
opfamilies with cross-type operators. The operator family project is now
DONE ... er, except for documentation ...
representation of equivalence classes of variables. This is an extensive
rewrite, but it brings a number of benefits:
* planner no longer fails in the presence of "incomplete" operator families
that don't offer operators for every possible combination of datatypes.
* avoid generating and then discarding redundant equality clauses.
* remove bogus assumption that derived equalities always use operators
named "=".
* mergejoins can work with a variety of sort orders (e.g., descending) now,
instead of tying each mergejoinable operator to exactly one sort order.
* better recognition of redundant sort columns.
* can make use of equalities appearing underneath an outer join.
which comparison operators to use for plan nodes involving tuple comparison
(Agg, Group, Unique, SetOp). Formerly the executor looked up the default
equality operator for the datatype, which was really pretty shaky, since it's
possible that the data being fed to the node is sorted according to some
nondefault operator class that could have an incompatible idea of equality.
The planner knows what it has sorted by and therefore can provide the right
equality operator to use. Also, this change moves a couple of catalog lookups
out of the executor and into the planner, which should help startup time for
pre-planned queries by some small amount. Modify the planner to remove some
other cavalier assumptions about always being able to use the default
operators. Also add "nulls first/last" info to the Plan node for a mergejoin
--- neither the executor nor the planner can cope yet, but at least the API is
in place.
cases. Operator classes now exist within "operator families". While most
families are equivalent to a single class, related classes can be grouped
into one family to represent the fact that they are semantically compatible.
Cross-type operators are now naturally adjunct parts of a family, without
having to wedge them into a particular opclass as we had done originally.
This commit restructures the catalogs and cleans up enough of the fallout so
that everything still works at least as well as before, but most of the work
needed to actually improve the planner's behavior will come later. Also,
there are not yet CREATE/DROP/ALTER OPERATOR FAMILY commands; the only way
to create a new family right now is to allow CREATE OPERATOR CLASS to make
one by default. I owe some more documentation work, too. But that can all
be done in smaller pieces once this infrastructure is in place.
Bruce Momjian