| Commit message (Collapse) | Author | Age |
|---|
| |
|
|
|
|
|
|
|
|
| |
Commit 19d8e2308bc added enum values with the prefix TU_, but a few
comments still referred to TUUI_, which was used in development
versions of the patches committed as 19d8e2308bc.
Author: Yugo Nagata <nagata@sraoss.co.jp>
Discussion: https://postgr.es/m/20250701110216.8ac8a9e4c6f607f1d954f44a@sraoss.co.jp
Backpatch-through: 16
|
| |
|
|
|
|
|
|
|
|
| |
Here we provide a bit more detail on why TidRangeEval() does return false
when trss_mintid is greater than trss_maxtid.
Reported-by: Junwang Zhao <zhjwpku@gmail.com>
Author: David Rowley <dgrowleyml@gmail.com>
Reviewed-by: Junwang Zhao <zhjwpku@gmail.com>
Discussion: https://postgr.es/m/CAEG8a3KUbUUqQgfK5X8Sj-%2BppPtGNTU%2BZiep0Rxr7SLjoR%2BB6w%40mail.gmail.com
|
| |
|
|
|
| |
Reported-by: Peter Smith <smithpb2250@gmail.com>
Discussion: https://postgr.es/m/CAHut+PspbHQmRCBL1c-opoJeTUKUaFFfUQJd2rhDZqwUrWCi7w@mail.gmail.com
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
When a MERGE's target table is the parent of an inheritance tree, any
INSERT actions insert into the parent table using ModifyTableState's
rootResultRelInfo. However, there are two bugs in the way is
initialized:
1. ExecInitMerge() incorrectly uses a different ResultRelInfo entry
from ModifyTableState's resultRelInfo array to build the insert
projection, which may not be compatible with rootResultRelInfo.
2. ExecInitModifyTable() does not fully initialize rootResultRelInfo.
Specifically, ri_WithCheckOptions, ri_WithCheckOptionExprs,
ri_returningList, and ri_projectReturning are not initialized.
This can lead to crashes, or incorrect query results due to failing to
check WCO's or process the RETURNING list for INSERT actions.
Fix both these bugs in ExecInitMerge(), noting that it is only
necessary to fully initialize rootResultRelInfo if the MERGE has
INSERT actions and the target table is a plain inheritance parent.
Backpatch to v15, where MERGE was introduced.
Reported-by: Andres Freund <andres@anarazel.de>
Author: Dean Rasheed <dean.a.rasheed@gmail.com>
Reviewed-by: Jian He <jian.universality@gmail.com>
Reviewed-by: Tender Wang <tndrwang@gmail.com>
Discussion: https://postgr.es/m/4rlmjfniiyffp6b3kv4pfy4jw3pciy6mq72rdgnedsnbsx7qe5@j5hlpiwdguvc
Backpatch-through: 15
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The code carelessly modified mtstate->ps.plan->targetlist,
which it's not supposed to do. Fortunately, there's not
really any need to do that because the planner already
set up a perfectly acceptable targetlist for the plan node.
We just need to remove the erroneous assignments and update some
relevant comments.
As it happens, the erroneous assignments caused the targetlist to
point to a different part of the source plan tree, so that there
isn't really a risk of the pointer becoming dangling after executor
termination. The only visible effect of this change we can find is
that EXPLAIN will show upper references to the ModifyTable's output
expressions using different variables. Formerly it showed Vars from
the first target relation that survived executor-startup pruning.
Now it always shows such references using the first relation appearing
in the planner output, independently of what happens during executor
pruning. On the whole that seems like a good thing.
Also make a small tweak in ExplainPreScanNode to ensure that the first
relation will receive a refname assignment in set_rtable_names, even
if it got pruned at startup. Previously the Vars might be shown
without any table qualification, which is confusing in a multi-table
query.
I considered back-patching this, but since the bug doesn't seem to
have any really terrible consequences in existing branches, it
seems better to not change their EXPLAIN output. It's not too late
for v18 though, especially since v18 already made other changes in
the EXPLAIN output for these cases.
Reported-by: Tom Lane <tgl@sss.pgh.pa.us>
Author: Andres Freund <andres@anarazel.de>
Co-authored-by: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://postgr.es/m/213261.1747611093@sss.pgh.pa.us
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
As pointed out by Tom Lane, the patch introduced fragile and invasive
design around plan invalidation handling when locking of prunable
partitions was deferred from plancache.c to the executor. In
particular, it violated assumptions about CachedPlan immutability and
altered executor APIs in ways that are difficult to justify given the
added complexity and overhead.
This also removes the firstResultRels field added to PlannedStmt in
commit 28317de72, which was intended to support deferred locking of
certain ModifyTable result relations.
Reported-by: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://postgr.es/m/605328.1747710381@sss.pgh.pa.us
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The macros INJECTION_POINT() and INJECTION_POINT_CACHED() are extended
with an optional argument that can be passed down to the callback
attached when an injection point is run, giving to callbacks the
possibility to manipulate a stack state given by the caller. The
existing callbacks in modules injection_points and test_aio have their
declarations adjusted based on that.
da7226993fd4 (core AIO infrastructure) and 93bc3d75d8e1 (test_aio) and
been relying on a set of workarounds where a static variable called
pgaio_inj_cur_handle is used as runtime argument in the injection point
callbacks used by the AIO tests, in combination with a TRY/CATCH block
to reset the argument value. The infrastructure introduced in this
commit will be reused for the AIO tests, simplifying them.
Reviewed-by: Greg Burd <greg@burd.me>
Discussion: https://postgr.es/m/Z_y9TtnXubvYAApS@paquier.xyz
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
We only need a tuplestore if we're actually going to accumulate
multiple result tuples. Obviously then we don't need one for non-set-
returning functions; but even a SRF doesn't need one if we decide to
use "lazyEval" (one row at a time) mode. In these cases, it's
sufficient to use the junkfilter's result slot to hold the single row
that's due to be returned. We just need to "materialize" that slot
to ensure it holds onto the data past shutdown of the sub-executor.
The original intent of this patch was partially to save a few cycles
(by not putting tuples into a tuplestore only to pull them back out
immediately), but mostly to ensure that we don't use a tuplestore
in non-set-returning functions. That's because I had concerns
about whether a tuplestore is safe to keep across queries,
which was possible for functions invoked via long-lived FmgrInfos
such as those kept in the typcache. There are no cases where SRFs
are called that way, so getting rid of the tuplestore in non-SRFs
should make things safer.
However, it emerges that running fmgr_sql in a short-lived context
(as 595d1efed made it do) makes the existing coding unsafe anyway:
we can end up with a long-lived TupleTableSlot holding a freeable
reference to a short-lived tuple, resulting in a double-free crash.
Not trying to pull tuples out of the tuplestore using that slot
dodges the problem, so I'm going to commit this now rather than
invent a band-aid solution for v18.
Reported-by: Alexander Lakhin <exclusion@gmail.com>
Author: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://postgr.es/m/2443532.1744919968@sss.pgh.pa.us
Discussion: https://postgr.es/m/9f975803-1a1c-4f21-b987-f572e110e860@gmail.com
|
| |
|
|
|
|
|
| |
These are all new to v18
Author: David Rowley <dgrowleyml@gmail.com>
Discussion: https://postgr.es/m/CAApHDvrMcr8XD107H3NV=WHgyBcu=sx5+7=WArr-n_cWUqdFXQ@mail.gmail.com
|
| |
|
|
|
|
|
|
| |
The large majority of these have been introduced by recent commits done
in the v18 development cycle.
Author: Alexander Lakhin <exclusion@gmail.com>
Discussion: https://postgr.es/m/9a7763ab-5252-429d-a943-b28941e0e28b@gmail.com
|
| |
|
|
|
|
|
|
|
|
|
| |
This gets rid of repetitive get_typlen calls in postquel_sub_params,
which show up as costing a few percent of the runtime in simple test
cases (more with more parameters).
In combination with the preceding patches, this gets us most of the
way back down to the amount of per-call overhead that functions.c
had before commit 0dca5d68d. There are some more things that could
be done, but this seems like an okay place to stop for v18.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
At this point, the only data structures we allocate directly in
fcontext are the SQLFunctionCache struct itself, the ParamListInfo
struct, and the execution_state array, all of which are small and
perfectly capable of being re-used across executions of the same
FmgrInfo. Hence, let's give them the same lifespan as the FmgrInfo.
This step gets rid of the separate SQLFunctionLink struct and makes
fn_extra point to SQLFunctionCache again. We also get rid of the
separate fcontext memory context and allocate these items directly
in fn_mcxt.
For notational simplicity, SQLFunctionCache still has an fcontext
field, but it's just a copy of fn_mcxt.
The motivation for this is to allow these structures to live as
long as the FmgrInfo and be re-used across calls, restoring the
original design without its propensity for memory leaks. This
gets rid of some per-call overhead that we added in 0dca5d68d.
We also make an effort to re-use the JunkFilter and result slot.
Those might need to change if the function definition changes,
so we compromise by rebuilding them if the cached plan changes.
This also moves the tuplestore into fn_mcxt so that it can be
re-used across calls, again undoing a change made in 0dca5d68d.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Put the JunkFilter and its result slot (and thence also
some subsidiary data such as the result tupledesc) into a
separate subcontext "jfcontext". This doesn't accomplish
a lot at this point, because we make a new JunkFilter each
time through the SQL function. However, the plan is to make
the fcontext long-lived, and that raises the possibility
that we'll need a new JunkFilter because the plan for the
result-generating query changes. A separate context makes
it easy to free the obsoleted data when that happens.
Also, instead of always running the sub-executor in fcontext,
make a separate context for it if we're doing lazy eval of
a SRF, and otherwise just run it inside CurrentMemoryContext.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Previously, much of this code ran with CurrentMemoryContext set
to be the function's fcontext, so that we tended to leak a lot of
stuff there. Commit 0dca5d68d dealt with that by releasing the
fcontext at the completion of each SQL function call, but we'd
like to go back to the previous approach of allowing the fcontext
to be query-lifespan. To control the leakage problem, rearrange
the code so that we mostly run in the memory context that fmgr_sql
is called in (which we expect to be short-lived). Notably, this
means that parsing/planning is all done in the short-lived context
and doesn't leak cruft into fcontext.
This patch also fixes the allocation of execution_state records
so that we don't leak them across executions. I set that up
with a re-usable array that contains at least as many
execution_state structs as we need for the current querytree.
The chain structure is still there, but it's not really doing
much for us, and maybe somebody will be motivated to get rid
of it. I'm not though.
This incidentally also moves the call of BlessTupleDesc to be
with the code that creates the JunkFilter. That doesn't make
much difference now, but a later patch will reduce the number
of times the JunkFilter gets made, and we needn't bless the
results any more often than that.
We still leak a fair amount in fcontext, particularly when
executing utility statements, but that's material for a
separate patch step; the point here is only to get rid of
unintentional allocations in fcontext.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Late in the development of commit 0dca5d68d, I added a step to copy
the result tlist we extract from the cached final query, because
I was afraid that that might not last as long as the JunkFilter that
we're passing it off to. However, that turns out to cost a noticeable
number of cycles, and it's really quite unnecessary because the
JunkFilter will not examine that tlist after it's been created.
(ExecFindJunkAttribute would use it, but we don't use that function
on this JunkFilter.) Hence, remove the copy step. For safety,
reset the might-become-dangling jf_targetList pointer to NIL.
In passing, remove DR_sqlfunction.cxt, which we don't use anymore;
it's confusing because it's not entirely clear which context it
ought to point at.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
If a GENERATED column is declared to have a domain data type where
the domain's constraints disallow null values, INSERT commands failed
because we built a targetlist that included coercing a null constant
to the domain's type. The failure occurred even when the generated
value would have been perfectly OK. This is adjacent to the issues
fixed in 0da39aa76, but we didn't notice for lack of testing a domain
with such a constraint.
We aren't going to use the result of the targetlist entry for the
generated column --- ExecComputeStoredGenerated will overwrite it.
So it's not really necessary that it have the exact datatype of
the generated column. This patch fixes the problem by changing
the targetlist entry to be a null Const of the domain's base type,
which should be sufficiently legal. (We do have to tweak
ExecCheckPlanOutput to accept the situation, though.)
This has been broken since we implemented generated columns.
However, this patch only applies easily as far back as v14, partly
because I (tgl) only carried 0da39aa76 back that far, but mostly
because v14 significantly refactored the handling of INSERT/UPDATE
targetlists. Given the lack of field complaints and the short
remaining support lifetime of v13, I judge the cost-benefit ratio
not good for devising a version that would work in v13.
Reported-by: jian he <jian.universality@gmail.com>
Author: jian he <jian.universality@gmail.com>
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://postgr.es/m/CACJufxG59tip2+9h=rEv-ykOFjt0cbsPVchhi0RTij8bABBA0Q@mail.gmail.com
Backpatch-through: 14
|
| |
|
|
|
|
|
|
|
|
| |
Make sure that function declarations use names that exactly match the
corresponding names from function definitions in a few places. These
inconsistencies were all introduced during Postgres 18 development.
This commit was written with help from clang-tidy, by mechanically
applying the same rules as similar clean-up commits (the earliest such
commit was commit 035ce1fe).
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The issue happens when building conflict information during apply of
INSERT or UPDATE operations that violate unique constraints on leaf
partitions.
The problem was introduced in commit 9ff68679b5, which removed the
redundant calls to ExecOpenIndices/ExecCloseIndices. The previous code was
relying on the redundant ExecOpenIndices call in
apply_handle_tuple_routing() to build the index information required for
unique key conflict detection.
The fix is to delay building the index information until a conflict is
detected instead of relying on ExecOpenIndices to do the same. The
additional benefit of this approach is that it avoids building index
information when there is no conflict.
Author: Hou Zhijie <houzj.fnst@fujitsu.com>
Reviewed-by:Reviewed-by: Amit Kapila <amit.kapila16@gmail.com>
Discussion: https://postgr.es/m/TYAPR01MB57244ADA33DDA57119B9D26494A62@TYAPR01MB5724.jpnprd01.prod.outlook.com
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
There were several places in ordering-related planning where a
requirement for btree was hardcoded but an amcanorder index could
suffice. This fixes that. We just need to do the necessary mapping
between strategy numbers and compare types and adjust some related
APIs so that this works independent of btree strategy numbers. For
instance, non-btree amcanorder indexes can now be used to support
sorting and merge joins. Also, predtest.c works independent of btree
strategy numbers now.
To avoid performance regressions, some details on btree and other
built-in index types are still hardcoded as shortcuts, but other index
types now have access to the same features by providing the required
flags and callbacks.
Author: Mark Dilger <mark.dilger@enterprisedb.com>
Co-authored-by: Peter Eisentraut <peter@eisentraut.org>
Discussion: https://www.postgresql.org/message-id/flat/E72EAA49-354D-4C2E-8EB9-255197F55330@enterprisedb.com
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This is yet another bit of fallout from the fact that backend/parser
(like other code) feels free to scribble on the parse tree it's
handed. In this case that resulted in modifying the
relatively-short-lived copy in the cached function's source_list.
That would be fine since we only need each source_list tree once
... except that if the parser fails after making some changes,
the function cache entry remains as-is and will still be there
if the user tries to execute the function again. Then we have
problems because we're feeding a non-pristine tree to the parser.
The most expedient fix is a quick copyObject(). I considered
other answers like somehow marking the cache entry invalid
temporarily, but that would add complexity and I'm not sure
it's worth it. In typical scenarios we'd only do this once
per function query per session.
Reported-by: Alexander Lakhin <exclusion@gmail.com>
Author: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://postgr.es/m/6d442183-102c-498a-81d1-eeeb086cdc5a@gmail.com
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
As coded, fmgr_sql() would get an assertion failure for a SQL function
that has an empty body and is declared to return some type other than
VOID. Typically you'd never get that far because fmgr_sql_validator()
would reject such a definition (I suspect that's how come I managed to
miss the bug). But if check_function_bodies is off or the function is
polymorphic, the validation check wouldn't get made.
Reported-by: Alexander Lakhin <exclusion@gmail.com>
Author: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://postgr.es/m/0fde377a-3870-4d18-946a-ce008ee5bb88@gmail.com
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The optimization does not take the removal of TIDs by a concurrent vacuum into
account. The concurrent vacuum can remove dead TIDs and make pages ALL_VISIBLE
while those dead TIDs are referenced in the bitmap. This can lead to a
skip_fetch scan returning too many tuples.
It likely would be possible to implement this optimization safely, but we
don't have the necessary infrastructure in place. Nor is it clear that it's
worth building that infrastructure, given how limited the skip_fetch
optimization is.
In the backbranches we just disable the optimization by always passing
need_tuples=true to table_beginscan_bm(). We can't perform API/ABI changes in
the backbranches and we want to make the change as minimal as possible.
Author: Matthias van de Meent <boekewurm+postgres@gmail.com>
Reported-By: Konstantin Knizhnik <knizhnik@garret.ru>
Discussion: https://postgr.es/m/CAEze2Wg3gXXZTr6_rwC+s4-o2ZVFB5F985uUSgJTsECx6AmGcQ@mail.gmail.com
Backpatch-through: 13
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
In the historical implementation of SQL functions (if they don't get
inlined), we built plans for all the contained queries at first call
within an outer query, and then re-used those plans for the duration
of the outer query, and then forgot everything. This was not ideal,
not least because the plans could not be customized to specific values
of the function's parameters. Our plancache infrastructure seems
mature enough to be used here. That will solve both the problem with
not being able to build custom plans and the problem with not being
able to share work across successive outer queries.
Aside from those performance concerns, this change fixes a
longstanding bugaboo with SQL functions: you could not write DDL that
would affect later statements in the same function. That's mostly
still true with new-style SQL functions, since the results of parse
analysis are baked into the stored query trees (and protected by
dependency records). But for old-style SQL functions, it will now
work much as it does with PL/pgSQL functions, because we delay parse
analysis and planning of each query until we're ready to run it.
Some edge cases that require replanning are now handled better too;
see for example the new rowsecurity test, where we now detect an RLS
context change that was previously missed.
One other edge-case change that might be worthy of a release note
is that we now insist that a SQL function's result be generated
by the physically-last query within it. Previously, if the last
original query was deleted by a DO INSTEAD NOTHING rule, we'd be
willing to take the result from the preceding query instead.
This behavior was undocumented except in source-code comments,
and it seems hard to believe that anyone's relying on it.
Along the way to this feature, we needed a few infrastructure changes:
* The plancache can now take either a raw parse tree or an
analyzed-but-not-rewritten Query as the starting point for a
CachedPlanSource. If given a Query, it is caller's responsibility
that nothing will happen to invalidate that form of the query.
We use this for new-style SQL functions, where what's in pg_proc is
serialized Query(s) and we trust the dependency mechanism to disallow
DDL that would break those.
* The plancache now offers a way to invoke a post-rewrite callback
to examine/modify the rewritten parse tree when it is rebuilding
the parse trees after a cache invalidation. We need this because
SQL functions sometimes adjust the parse tree to make its output
exactly match the declared result type; if the plan gets rebuilt,
that has to be re-done.
* There is a new backend module utils/cache/funccache.c that
abstracts the idea of caching data about a specific function
usage (a particular function and set of input data types).
The code in it is moved almost verbatim from PL/pgSQL, which
has done that for a long time. We use that logic now for
SQL-language functions too, and maybe other PLs will have use
for it in the future.
Author: Alexander Pyhalov <a.pyhalov@postgrespro.ru>
Co-authored-by: Tom Lane <tgl@sss.pgh.pa.us>
Reviewed-by: Pavel Stehule <pavel.stehule@gmail.com>
Discussion: https://postgr.es/m/8216639.NyiUUSuA9g@aivenlaptop
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
ExecInitPartitionInfo() duplicates much of the logic in
ExecInitMerge(), except that it failed to handle DO NOTHING
actions. This would cause an "unknown action in MERGE WHEN clause"
error if a MERGE with any DO NOTHING actions attempted to insert into
a partition not already initialised by ExecInitModifyTable().
Bug: #18871
Reported-by: Alexander Lakhin <exclusion@gmail.com>
Author: Tender Wang <tndrwang@gmail.com>
Reviewed-by: Gurjeet Singh <gurjeet@singh.im>
Discussion: https://postgr.es/m/18871-b44e3c96de3bd2e8%40postgresql.org
Backpatch-through: 15
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This was left out of the original patch for virtual generated columns
(commit 83ea6c54025).
This just involves a bit of extra work in the executor to expand the
generation expressions and run a "IS NOT NULL" test against them.
There is also a bit of work to make sure that not-null constraints are
checked during a table rewrite.
Author: jian he <jian.universality@gmail.com>
Reviewed-by: Xuneng Zhou <xunengzhou@gmail.com>
Reviewed-by: Navneet Kumar <thanit3111@gmail.com>
Reviewed-by: Álvaro Herrera <alvherre@alvh.no-ip.org>
Discussion: https://postgr.es/m/CACJufxHArQysbDkWFmvK+D1TPHQWWTxWN15cMuUaTYX3xhQXgg@mail.gmail.com
|
| |
|
|
|
|
|
|
|
|
|
| |
Modernize code in ExecRelCheck() and ExecConstraints() a bit,
preparing the way for some new code.
Co-authored-by: jian he <jian.universality@gmail.com>
Reviewed-by: Xuneng Zhou <xunengzhou@gmail.com>
Reviewed-by: Navneet Kumar <thanit3111@gmail.com>
Reviewed-by: Álvaro Herrera <alvherre@alvh.no-ip.org>
Discussion: https://postgr.es/m/CACJufxHArQysbDkWFmvK+D1TPHQWWTxWN15cMuUaTYX3xhQXgg@mail.gmail.com
|
| |
|
|
|
|
|
|
| |
Rename ResultRelInfo.ri_ConstraintExprs to ri_CheckConstraintExprs.
This reflects its specific purpose better and avoids confusion with
adjacent fields with similar but distinct purposes.
Discussion: https://postgr.es/m/CACJufxHArQysbDkWFmvK+D1TPHQWWTxWN15cMuUaTYX3xhQXgg@mail.gmail.com
|
| |
|
|
|
|
|
|
|
|
| |
Reduces memory required for hash aggregation by avoiding an allocation
and a pointer in the TupleHashEntryData structure. That structure is
used for all buckets, whether occupied or not, so the savings is
substantial.
Discussion: https://postgr.es/m/AApHDvpN4v3t_sdz4dvrv1Fx_ZPw=twSnxuTEytRYP7LFz5K9A@mail.gmail.com
Reviewed-by: David Rowley <dgrowleyml@gmail.com>
|
| |
|
|
|
|
|
|
|
| |
Allows an "extra" argument that allocates extra memory at the end of
the MinimalTuple. This is important for callers that need to store
additional data, but do not want to perform an additional allocation.
Suggested-by: David Rowley <dgrowleyml@gmail.com>
Discussion: https://postgr.es/m/CAApHDvppeqw2pNM-+ahBOJwq2QmC0hOAGsmCpC89QVmEoOvsdg@mail.gmail.com
|
| |
|
|
|
|
|
| |
Refactor for upcoming optimizations.
Reviewed-by: David Rowley <dgrowleyml@gmail.com>
Discussion: https://postgr.es/m/1cc3b400a0e8eead18ff967436fa9e42c0c14cfb.camel@j-davis.com
|
| |
|
|
|
|
|
|
|
|
| |
The entries aren't freed until the entire hash table is destroyed, so
use the Bump allocator to improve allocation speed, avoid wasting
space on the chunk header, and avoid wasting space due to the
power-of-two allocations.
Discussion: https://postgr.es/m/CAApHDvqv1aNB4cM36FzRwivXrEvBO_LsG_eQ3nqDXTjECaatOQ@mail.gmail.com
Reviewed-by: David Rowley
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Introduce a new conflict type, multiple_unique_conflicts, to handle cases
where an incoming row during logical replication violates multiple UNIQUE
constraints.
Previously, the apply worker detected and reported only the first
encountered key conflict (insert_exists/update_exists), causing repeated
failures as each constraint violation needs to be handled one by one
making the process slow and error-prone.
With this patch, the apply worker checks all unique constraints upfront
once the first key conflict is detected and reports
multiple_unique_conflicts if multiple violations exist. This allows users
to resolve all conflicts at once by deleting all conflicting tuples rather
than dealing with them individually or skipping the transaction.
In the future, this will also allow us to specify different resolution
handlers for such a conflict type.
Add the stats for this conflict type in pg_stat_subscription_stats.
Author: Nisha Moond <nisha.moond412@gmail.com>
Author: Zhijie Hou <houzj.fnst@fujitsu.com>
Reviewed-by: Amit Kapila <amit.kapila16@gmail.com>
Reviewed-by: Peter Smith <smithpb2250@gmail.com>
Reviewed-by: Dilip Kumar <dilipbalaut@gmail.com>
Discussion: https://postgr.es/m/CABdArM7FW-_dnthGkg2s0fy1HhUB8C3ELA0gZX1kkbs1ZZoV3Q@mail.gmail.com
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This field can be optionally set in a PlannedStmt through the planner
hook, giving extensions the possibility to assign an identifier related
to a computed plan. The backend is changed to report it in the backend
entry of a process running (including the extended query protocol), with
semantics and APIs to set or get it similar to what is used for the
existing query ID (introduced in the backend via 4f0b0966c8). The plan
ID is reset at the same timing as the query ID. Currently, this
information is not added to the system view pg_stat_activity; extensions
can access it through PgBackendStatus.
Some patches have been proposed to provide some features in the planning
area, where a plan identifier is used as a key to know the plan involved
(for statistics, plan storage and manipulations, etc.), and the point of
this commit is to provide an anchor in the backend that extensions can
rely on for future work. The reset of the plan identifier is
controlled by core and follows the same pattern as the query identifier
added in 4f0b0966c8.
The contents of this commit are extracted from a larger set proposed
originally by Lukas Fittl, that Sami Imseih has proposed as an
independent change, with a few tweaks sprinkled by me.
Author: Lukas Fittl <lukas@fittl.com>
Author: Sami Imseih <samimseih@gmail.com>
Reviewed-by: Bertrand Drouvot <bertranddrouvot.pg@gmail.com>
Reviewed-by: Michael Paquier <michael@paquier.xyz>
Discussion: https://postgr.es/m/CAP53Pkyow59ajFMHGpmb1BK9WHDypaWtUsS_5DoYUEfsa_Hktg@mail.gmail.com
Discussion: https://postgr.es/m/CAA5RZ0vyWd4r35uUBUmhngv8XqeiJUkJDDKkLf5LCoWxv-t_pw@mail.gmail.com
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Commit cbc127917e introduced tracking of unpruned relids to avoid
processing pruned relations, and changed ExecInitModifyTable() to
initialize only unpruned result relations. As a result, MERGE
statements that prune all target partitions can now lead to crashes
or incorrect behavior during execution.
The crash occurs because some executor code paths rely on
ModifyTableState.resultRelInfo[0] being present and initialized,
even when no result relations remain after pruning. For example,
ExecMerge() and ExecMergeNotMatched() use the first resultRelInfo
to determine the appropriate action. Similarly,
ExecInitPartitionInfo() assumes that at least one result relation
exists.
To preserve these assumptions, ExecInitModifyTable() now includes the
first result relation in the initialized result relation list if all
result relations for that ModifyTable were pruned. To enable that,
ExecDoInitialPruning() ensures the first relation is locked if it was
pruned and locking is necessary.
To support this exception to the pruning logic, PlannedStmt now
includes a list of RT indexes identifying the first result relation
of each ModifyTable node in the plan. This allows
ExecDoInitialPruning() to check whether each such relation was
pruned and, if so, lock it if necessary.
Bug: #18830
Reported-by: Robins Tharakan <tharakan@gmail.com>
Diagnozed-by: Tender Wang <tndrwang@gmail.com>
Diagnozed-by: Dean Rasheed <dean.a.rasheed@gmail.com>
Co-authored-by: Dean Rasheed <dean.a.rasheed@gmail.com>
Reviewed-by: Tender Wang <tndrwang@gmail.com>
Reviewed-by: Dean Rasheed <dean.a.rasheed@gmail.com>
Discussion: https://postgr.es/m/18830-1f31ea1dc930d444%40postgresql.org
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Modules can use RegisterExtensionExplainOption to register new
EXPLAIN options, and GetExplainExtensionId, GetExplainExtensionState,
and SetExplainExtensionState to store related state inside the
ExplainState object.
Since this substantially increases the amount of code that needs
to handle ExplainState-related tasks, move a few bits of existing
code to a new file explain_state.c and add the rest of this
infrastructure there.
See the comments at the top of explain_state.c for further
explanation of how this mechanism works.
This does not yet provide a way for such such options to do anything
useful. The intention is that we'll add hooks for that purpose in a
separate commit.
Discussion: http://postgr.es/m/CA+TgmoYSzg58hPuBmei46o8D3SKX+SZoO4K_aGQGwiRzvRApLg@mail.gmail.com
Reviewed-by: Srinath Reddy <srinath2133@gmail.com>
Reviewed-by: Andrei Lepikhov <lepihov@gmail.com>
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
Reviewed-by: Sami Imseih <samimseih@gmail.com>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
After pushing the bitmap iterator into table-AM specific code (as part
of making bitmap heap scan use the read stream API in 2b73a8cd33b7),
scan_bitmap_next_block() no longer returns the current block number.
Since scan_bitmap_next_block() isn't returning any relevant information
to bitmap table scan code, it makes more sense to get rid of it.
Now, bitmap table scan code only calls table_scan_bitmap_next_tuple(),
and the heap AM implementation of scan_bitmap_next_block() is a local
helper in heapam_handler.c.
Reviewed-by: Tomas Vondra <tomas@vondra.me>
Discussion: https://postgr.es/m/flat/CAAKRu_ZwCwWFeL_H3ia26bP2e7HiKLWt0ZmGXPVwPO6uXq0vaA%40mail.gmail.com
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Make Bitmap Heap Scan use the read stream API instead of invoking
ReadBuffer() for each block indicated by the bitmap.
The read stream API handles prefetching, so remove all of the explicit
prefetching from bitmap heap scan code.
Now, heap table AM implements a read stream callback which uses the
bitmap iterator to return the next required block to the read stream
code.
Tomas Vondra conducted extensive regression testing of this feature.
Andres Freund, Thomas Munro, and I analyzed regressions and Thomas Munro
patched the read stream API.
Author: Melanie Plageman <melanieplageman@gmail.com>
Reviewed-by: Tomas Vondra <tomas@vondra.me>
Tested-by: Tomas Vondra <tomas@vondra.me>
Tested-by: Andres Freund <andres@anarazel.de>
Tested-by: Thomas Munro <thomas.munro@gmail.com>
Tested-by: Nazir Bilal Yavuz <byavuz81@gmail.com>
Discussion: https://postgr.es/m/flat/CAAKRu_ZwCwWFeL_H3ia26bP2e7HiKLWt0ZmGXPVwPO6uXq0vaA%40mail.gmail.com
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
| |
Remove the TBMIterateResult member from the TBMPrivateIterator and
TBMSharedIterator and make tbm_[shared|private_]iterate() take a
TBMIterateResult as a parameter.
This allows tidbitmap API users to manage multiple TBMIterateResults per
scan. This is required for bitmap heap scan to use the read stream API,
with which there may be multiple I/Os in flight at once, each one with a
TBMIterateResult.
Reviewed-by: Tomas Vondra <tomas@vondra.me>
Discussion: https://postgr.es/m/d4bb26c9-fe07-439e-ac53-c0e244387e01%40vondra.me
|
| |
|
|
|
|
|
|
|
|
|
| |
We did not wake up on interrupts while waiting on async events on an
async-capable append node. For example, if you tried to cancel the
query, nothing would happen until one of the async subplans becomes
readable. To fix, add WL_LATCH_SET to the WaitEventSet.
Backpatch down to v14 where async Append execution was introduced.
Discussion: https://www.postgresql.org/message-id/37a40570-f558-40d3-b5ea-5c2079b3b30b@iki.fi
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Expose the count of index searches/index descents in EXPLAIN ANALYZE's
output for index scan/index-only scan/bitmap index scan nodes. This
information is particularly useful with scans that use ScalarArrayOp
quals, where the number of index searches can be unpredictable due to
implementation details that interact with physical index characteristics
(at least with nbtree SAOP scans, since Postgres 17 commit 5bf748b8).
The information shown also provides useful context when EXPLAIN ANALYZE
runs a plan with an index scan node that successfully applied the skip
scan optimization (set to be added to nbtree by an upcoming patch).
The instrumentation works by teaching all index AMs to increment a new
nsearches counter whenever a new index search begins. The counter is
incremented at exactly the same point that index AMs already increment
the pg_stat_*_indexes.idx_scan counter (we're counting the same event,
but at the scan level rather than the relation level). Parallel queries
have workers copy their local counter struct into shared memory when an
index scan node ends -- even when it isn't a parallel aware scan node.
An earlier version of this patch that only worked with parallel aware
scans became commit 5ead85fb (though that was quickly reverted by commit
d00107cd following "debug_parallel_query=regress" buildfarm failures).
Our approach doesn't match the approach used when tracking other index
scan related costs (e.g., "Rows Removed by Filter:"). It is comparable
to the approach used in similar cases involving costs that are only
readily accessible inside an access method, not from the executor proper
(e.g., "Heap Blocks:" output for a Bitmap Heap Scan, which was recently
enhanced to show per-worker costs by commit 5a1e6df3, using essentially
the same scheme as the one used here). It is necessary for index AMs to
have direct responsibility for maintaining the new counter, since the
counter might need to be incremented multiple times per amgettuple call
(or per amgetbitmap call). But it is also necessary for the executor
proper to manage the shared memory now used to transfer each worker's
counter struct to the leader.
Author: Peter Geoghegan <pg@bowt.ie>
Reviewed-By: Robert Haas <robertmhaas@gmail.com>
Reviewed-By: Tomas Vondra <tomas@vondra.me>
Reviewed-By: Masahiro Ikeda <ikedamsh@oss.nttdata.com>
Reviewed-By: Matthias van de Meent <boekewurm+postgres@gmail.com>
Discussion: https://postgr.es/m/CAH2-WzkRqvaqR2CTNqTZP0z6FuL4-3ED6eQB0yx38XBNj1v-4Q@mail.gmail.com
Discussion: https://postgr.es/m/CAH2-Wz=PKR6rB7qbx+Vnd7eqeB5VTcrW=iJvAsTsKbdG+kW_UA@mail.gmail.com
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
| |
Many STRICT function calls will have one or two arguments, in which
case we can speed up checking for NULL input by avoiding setting up
a loop over the arguments. This adds EEOP_FUNCEXPR_STRICT_1 and the
corresponding EEOP_FUNCEXPR_STRICT_2 for functions with one and two
arguments respectively.
Author: Andres Freund <andres@anarazel.de>
Co-authored-by: Daniel Gustafsson <daniel@yesql.se>
Reviewed-by: Andreas Karlsson <andreas@proxel.se>
Discussion: https://postgr.es/m/415721CE-7D2E-4B74-B5D9-1950083BA03E@yesql.se
Discussion: https://postgr.es/m/20191023163849.sosqbfs5yenocez3@alap3.anarazel.de
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Knowing when the side-effects of an expression is the intended result
of the execution, rather than the returnvalue, is important for being
able generate more efficient JITed code. This replaces EEOP_DONE with
two new steps: EEOP_DONE_RETURN and EEOP_DONE_NO_RETURN. Expressions
which return a value should use the former step; expressions used for
their side-effects which don't return value should use the latter.
Author: Andres Freund <andres@anarazel.de>
Co-authored-by: Daniel Gustafsson <daniel@yesql.se>
Reviewed-by: Andreas Karlsson <andreas@proxel.se>
Discussion: https://postgr.es/m/415721CE-7D2E-4B74-B5D9-1950083BA03E@yesql.se
Discussion: https://postgr.es/m/20191023163849.sosqbfs5yenocez3@alap3.anarazel.de
|
| |
|
|
|
|
|
|
|
|
| |
Commit 2367503177 fixed this in RelationFindReplTupleByIndex(), but I
missed two other similar cases.
Per report from Ranier Vilela.
Discussion: https://www.postgresql.org/message-id/CAEudQArUT1dE45WN87F-Gb7XMy_hW6x1DFd3sqdhhxP-RMDa0Q@mail.gmail.com
Backpatch-through: 13
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The function calls GetLatestSnapshot() to acquire a fresh snapshot,
makes it active, and was meant to pass it to table_tuple_lock(), but
instead called GetLatestSnapshot() again to acquire yet another
snapshot. It was harmless because the heap AM and all other known
table AMs ignore the 'snapshot' argument anyway, but let's be tidy.
In the long run, this perhaps should be redesigned so that snapshot
was not needed in the first place. The table AM API uses TID +
snapshot as the unique identifier for the row version, which is
questionable when the row came from an index scan with a Dirty
snapshot. You might lock a different row version when you use a
different snapshot in the table_tuple_lock() call (a fresh MVCC
snapshot) than in the index scan (DirtySnapshot). However, in the heap
AM and other AMs where the TID alone identifies the row version, it
doesn't matter. So for now, just fix the obvious albeit harmless bug.
This has been wrong ever since the table AM API was introduced in
commit 5db6df0c01, so backpatch to all supported versions.
Discussion: https://www.postgresql.org/message-id/83d243d6-ad8d-4307-8b51-2ee5844f6230@iki.fi
Backpatch-through: 13
|
| |
|
|
|
|
|
|
|
|
|
|
| |
Stop comparing access method OID values against HASH_AM_OID and
BTREE_AM_OID, and instead check the IndexAmRoutine for an index to see
if it advertises its ability to perform the necessary ordering,
hashing, or cross-type comparing functionality. A field amcanorder
already existed, this uses it more widely. Fields amcanhash and
amcancrosscompare are added for the other purposes.
Author: Mark Dilger <mark.dilger@enterprisedb.com>
Discussion: https://www.postgresql.org/message-id/flat/E72EAA49-354D-4C2E-8EB9-255197F55330@enterprisedb.com
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
A non-leaf partition with a subplan that is an Append node was
omitted from PlannedStmt.unprunableRelids because it was mistakenly
included in PlannerGlobal.prunableRelids due to the way
PartitionedRelPruneInfo.leafpart_rti_map[] is constructed. This
happened when a non-leaf partition used an unflattened Append or
MergeAppend. As a result, ExecGetRangeTableRelation() reported an
error when called from CreatePartitionPruneState() to process the
partition's own PartitionPruneInfo, since it was treated as prunable
when it should not have been.
Reported-by: Alexander Lakhin <exclusion@gmail.com> (via sqlsmith)
Diagnosed-by: Tender Wang <tndrwang@gmail.com>
Reviewed-by: Tender Wang <tndrwang@gmail.com>
Discussion: https://postgr.es/m/74839af6-aadc-4f60-ae77-ae65f94bf607@gmail.com
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The type argument wasn't actually really necessary. It was a remnant
of converting the API of the gist strategy translation from using
opclass to using opfamily+opcintype (commits c09e5a6a016,
622f678c102). For looking up the gist translation function, we used
the convention "amproclefttype = amprocrighttype = opclass's
opcintype" (see pg_amproc.h). But each operator family should only
have one translation function, and getting the right type for the
lookup is sometimes cumbersome and fragile, so this is all
unnecessarily complicated.
To simplify this, change the gist stategy support procedure to take
"any", "any" as argument. (This is arbitrary but seems intuitive.
The alternative of using InvalidOid as argument(s) upsets various DDL
commands, so it's not practical.) Then we don't need opcintype for
the lookup, and we can remove it from all the API layers introduced by
commit c09e5a6a016.
This also adds some more documentation about the correct signature of
the gist support function and adds more checks in gistvalidate().
This was previously underspecified. (It relied implicitly on
convention mentioned above.)
Discussion: https://www.postgresql.org/message-id/flat/E72EAA49-354D-4C2E-8EB9-255197F55330@enterprisedb.com
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Before executing a cached generic plan, AcquireExecutorLocks() in
plancache.c locks all relations in a plan's range table to ensure the
plan is safe for execution. However, this locks runtime-prunable
relations that will later be pruned during "initial" runtime pruning,
introducing unnecessary overhead.
This commit defers locking for such relations to executor startup and
ensures that if the CachedPlan is invalidated due to concurrent DDL
during this window, replanning is triggered. Deferring these locks
avoids unnecessary locking overhead for pruned partitions, resulting
in significant speedup, particularly when many partitions are pruned
during initial runtime pruning.
* Changes to locking when executing generic plans:
AcquireExecutorLocks() now locks only unprunable relations, that is,
those found in PlannedStmt.unprunableRelids (introduced in commit
cbc127917e), to avoid locking runtime-prunable partitions
unnecessarily. The remaining locks are taken by
ExecDoInitialPruning(), which acquires them only for partitions that
survive pruning.
This deferral does not affect the locks required for permission
checking in InitPlan(), which takes place before initial pruning.
ExecCheckPermissions() now includes an Assert to verify that all
relations undergoing permission checks, none of which can be in the
set of runtime-prunable relations, are properly locked.
* Plan invalidation handling:
Deferring locks introduces a window where prunable relations may be
altered by concurrent DDL, invalidating the plan. A new function,
ExecutorStartCachedPlan(), wraps ExecutorStart() to detect and handle
invalidation caused by deferred locking. If invalidation occurs,
ExecutorStartCachedPlan() updates CachedPlan using the new
UpdateCachedPlan() function and retries execution with the updated
plan. To ensure all code paths that may be affected by this handle
invalidation properly, all callers of ExecutorStart that may execute a
PlannedStmt from a CachedPlan have been updated to use
ExecutorStartCachedPlan() instead.
UpdateCachedPlan() replaces stale plans in CachedPlan.stmt_list. A new
CachedPlan.stmt_context, created as a child of CachedPlan.context,
allows freeing old PlannedStmts while preserving the CachedPlan
structure and its statement list. This ensures that loops over
statements in upstream callers of ExecutorStartCachedPlan() remain
intact.
ExecutorStart() and ExecutorStart_hook implementations now return a
boolean value indicating whether plan initialization succeeded with a
valid PlanState tree in QueryDesc.planstate, or false otherwise, in
which case QueryDesc.planstate is NULL. Hook implementations are
required to call standard_ExecutorStart() at the beginning, and if it
returns false, they should do the same without proceeding.
* Testing:
To verify these changes, the delay_execution module tests scenarios
where cached plans become invalid due to changes in prunable relations
after deferred locks.
* Note to extension authors:
ExecutorStart_hook implementations must verify plan validity after
calling standard_ExecutorStart(), as explained earlier. For example:
if (prev_ExecutorStart)
plan_valid = prev_ExecutorStart(queryDesc, eflags);
else
plan_valid = standard_ExecutorStart(queryDesc, eflags);
if (!plan_valid)
return false;
<extension-code>
return true;
Extensions accessing child relations, especially prunable partitions,
via ExecGetRangeTableRelation() must now ensure their RT indexes are
present in es_unpruned_relids (introduced in commit cbc127917e), or
they will encounter an error. This is a strict requirement after this
change, as only relations in that set are locked.
The idea of deferring some locks to executor startup, allowing locks
for prunable partitions to be skipped, was first proposed by Tom Lane.
Reviewed-by: Robert Haas <robertmhaas@gmail.com> (earlier versions)
Reviewed-by: David Rowley <dgrowleyml@gmail.com> (earlier versions)
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us> (earlier versions)
Reviewed-by: Tomas Vondra <tomas@vondra.me>
Reviewed-by: Junwang Zhao <zhjwpku@gmail.com>
Discussion: https://postgr.es/m/CA+HiwqFGkMSge6TgC9KQzde0ohpAycLQuV7ooitEEpbKB0O_mg@mail.gmail.com
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
These functions should be called at most once per ResultRelInfo;
it's wasteful to do otherwise, and certainly the pattern of
opening twice and then closing twice is a bad idea. Moreover,
aminsertcleanup functions might not be prepared to be called twice,
as the just-hardened code in BRIN demonstrates.
This amounts to an API change, since such coding patterns were
safe even if wasteful before v17. Hence, apply to HEAD only.
(Extension code violating this new rule faces some risk in v17,
but we just fixed brininsertcleanup and there are probably few
other aminsertcleanup functions as yet. So the odds of breaking
usable code seem higher than the odds of doing something useful
with a back-patch.)
Bug: #18815
Reported-by: Sergey Belyashov <sergey.belyashov@gmail.com>
Discussion: https://postgr.es/m/18815-2a0407cc7f40b327@postgresql.org
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Until now ExecChooseHashTableSize() considered only the size of the
in-memory hash table, and ignored the memory needed for the batch files.
Which can be a significant amount, because each batch needs two BufFiles
(each with a BLCKSZ buffer). The same issue applies to increasing the
number of batches during execution.
It's also possible to trigger a "batch explosion", e.g. due to duplicate
values or skew. We've seen reports of joins with hundreds of thousands
(or even millions) of batches, consuming gigabytes of memory, triggering
OOM errors. These cases may be fairly rare, but it's clearly possible to
hit them.
These issues can't be prevented during planning. Even if we improve
that, it does not help with execution-time batch explosion. We can
however reduce the impact and use as little memory as possible.
This patch improves the behavior by adjusting how the memory is divided
between the hash table and batch files. It may be better to use fewer
batch files, even if it means the hash table will exceed the limit.
The capacity of the hash node may be increased either by doubling he
number of batches, or doubling the size of the in-memory hash table. The
outcome is the same, but the memory usage may be very different. For low
nbatch values it's better to add batches, for high nbatch values it's
better to allow a larger hash table.
The patch considers both options, both during the initial sizing and
then during execution, to minimize how much the limit gets exceeded.
It might seem this patch is relaxing the memory limit - allowing it to
be exceeded. But that's not really the case. It has always been like
that, except the memory used by batches was ignored.
Allowing the hash table to grow may also prevent the batch explosion.
If there's a large batch that can't be split (due to hash collisions or
duplicate values), at some point the memory limit will increase enough
for the batch to fit into the hash table.
This patch was in the works for a long time. The early versions were
posted in 2019, and revived every year or two when we happened to get
the next report of OOM due to a hashjoin batch explosion. Each of those
patch versions were reviewed by a couple people. I'm mentioning only
Melanie Plageman and Robert Haas, because they reviewed the last
version, and the older patches are very different.
Reviewed-by: Melanie Plageman, Robert Haas
Discussion: https://postgr.es/m/7bed6c08-72a0-4ab9-a79c-e01fcdd0940f@vondra.me
Discussion: https://postgr.es/m/20190504003414.bulcbnge3rhwhcsh%40development
Discussion: https://postgr.es/m/20190428141901.5dsbge2ka3rxmpk6%40development
|
