| Age | Commit message (Collapse) | Author |
|---|
|
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
|
|
Author: Zhang Mingli
Reviewed-by: Richard Guo
Discussion: https://postgr.es/m/198a8d1e-0792-4e7f-828e-902aa342f36e@Spark
|
|
ExecInitModifyTable() forgot to trim MERGE-related lists to exclude
entries for result relations pruned during initial pruning, so fix
that.
While at it, make the function's use of the pruned resultRelations
list, rather than ModifyTable.resultRelations, more consistent.
Reported-by: Alexander Lakhin <exclusion@gmail.com> (via sqlsmith)
Reviewed-by: Junwang Zhao <zhjwpku@gmail.com>
Discussion: https://postgr.es/m/e72c94d9-e5f9-4753-9bc1-69d72bd54b8a@gmail.com
|
|
wal_buffers_full has been introduced in pg_stat_wal in 8d9a935965f, as
some information providing metrics for the tuning of the GUC
wal_buffers. WalUsage has been introduced before that in df3b181499.
Moving this field is proving to be beneficial for several reasons:
- This information can now be made available in more layers, providing
more granularity than just pg_stat_wal, on a per-query basis: EXPLAIN,
pgss and VACUUM/ANALYZE logs.
- A patch is under discussion to provide statistics for WAL at backend
level, and this move simplifies a bit the handling of pending
statistics. The remaining data in PgStat_PendingWalStats now relates to
write/sync counters and times, with equivalents present in pg_stat_io,
that backend statistics are able to already track. So this should cut
all the dependencies between PgStat_PendingWalStats and WAL stats at
backend level.
As of this change, wal_buffers_full only shows in pg_stat_wal.
Author: Bertrand Drouvot
Reviewed-by: Ilia Evdokimov
Discussion: https://postgr.es/m/Z6SOha5YFFgvpwQY@ip-10-97-1-34.eu-west-3.compute.internal
|
|
Commit 27cc7cd2bc8a accidentally placed the assertion ensuring
that the pointer isn't NULL after it had already been accessed.
Fix by moving the pointer dereferencing to after the assertion.
Backpatch to all supported branches.
Author: Dmitry Koval <d.koval@postgrespro.ru>
Reviewed-by: Daniel Gustafsson <daniel@yesql.se>
Reviewed-by: Michael Paquier <michael@paquier.xyz>
Discussion: https://postgr.es/m/1618848d-cdc7-414b-9c03-08cf4bef4408@postgrespro.ru
Backpatch-through: 13
|
|
Requires adding a guard against shift-by-32. Previously, that was
impossible because the number of partitions was always greater than 1,
but a new injection point can force the number of partitions to 1.
Discussion: https://postgr.es/m/ff4e59305e5d689e03cd256a736348d3e7958f8f.camel@j-davis.com
|
|
This adds a new variant of generated columns that are computed on read
(like a view, unlike the existing stored generated columns, which are
computed on write, like a materialized view).
The syntax for the column definition is
... GENERATED ALWAYS AS (...) VIRTUAL
and VIRTUAL is also optional. VIRTUAL is the default rather than
STORED to match various other SQL products. (The SQL standard makes
no specification about this, but it also doesn't know about VIRTUAL or
STORED.) (Also, virtual views are the default, rather than
materialized views.)
Virtual generated columns are stored in tuples as null values. (A
very early version of this patch had the ambition to not store them at
all. But so much stuff breaks or gets confused if you have tuples
where a column in the middle is completely missing. This is a
compromise, and it still saves space over being forced to use stored
generated columns. If we ever find a way to improve this, a bit of
pg_upgrade cleverness could allow for upgrades to a newer scheme.)
The capabilities and restrictions of virtual generated columns are
mostly the same as for stored generated columns. In some cases, this
patch keeps virtual generated columns more restricted than they might
technically need to be, to keep the two kinds consistent. Some of
that could maybe be relaxed later after separate careful
considerations.
Some functionality that is currently not supported, but could possibly
be added as incremental features, some easier than others:
- index on or using a virtual column
- hence also no unique constraints on virtual columns
- extended statistics on virtual columns
- foreign-key constraints on virtual columns
- not-null constraints on virtual columns (check constraints are supported)
- ALTER TABLE / DROP EXPRESSION
- virtual column cannot have domain type
- virtual columns are not supported in logical replication
The tests in generated_virtual.sql have been copied over from
generated_stored.sql with the keyword replaced. This way we can make
sure the behavior is mostly aligned, and the differences can be
visible. Some tests for currently not supported features are
currently commented out.
Reviewed-by: Jian He <jian.universality@gmail.com>
Reviewed-by: Dean Rasheed <dean.a.rasheed@gmail.com>
Tested-by: Shlok Kyal <shlok.kyal.oss@gmail.com>
Discussion: https://www.postgresql.org/message-id/flat/a368248e-69e4-40be-9c07-6c3b5880b0a6@eisentraut.org
|
|
This commit introduces changes to track unpruned relations explicitly,
making it possible for top-level plan nodes, such as ModifyTable and
LockRows, to avoid processing partitions pruned during initial
pruning. Scan-level nodes, such as Append and MergeAppend, already
avoid the unnecessary processing by accessing partition pruning
results directly via part_prune_index. In contrast, top-level nodes
cannot access pruning results directly and need to determine which
partitions remain unpruned.
To address this, this commit introduces a new bitmapset field,
es_unpruned_relids, which the executor uses to track the set of
unpruned relations. This field is referenced during plan
initialization to skip initializing certain nodes for pruned
partitions. It is initialized with PlannedStmt.unprunableRelids,
a new field that the planner populates with RT indexes of relations
that cannot be pruned during runtime pruning. These include relations
not subject to partition pruning and those required for execution
regardless of pruning.
PlannedStmt.unprunableRelids is computed during set_plan_refs() by
removing the RT indexes of runtime-prunable relations, identified
from PartitionPruneInfos, from the full set of relation RT indexes.
ExecDoInitialPruning() then updates es_unpruned_relids by adding
partitions that survive initial pruning.
To support this, PartitionedRelPruneInfo and PartitionedRelPruningData
now include a leafpart_rti_map[] array that maps partition indexes to
their corresponding RT indexes. The former is used in set_plan_refs()
when constructing unprunableRelids, while the latter is used in
ExecDoInitialPruning() to convert partition indexes returned by
get_matching_partitions() into RT indexes, which are then added to
es_unpruned_relids.
These changes make it possible for ModifyTable and LockRows nodes to
process only relations that remain unpruned after initial pruning.
ExecInitModifyTable() trims lists, such as resultRelations,
withCheckOptionLists, returningLists, and updateColnosLists, to
consider only unpruned partitions. It also creates ResultRelInfo
structs only for these partitions. Similarly, child RowMarks for
pruned relations are skipped.
By avoiding unnecessary initialization of structures for pruned
partitions, these changes improve the performance of updates and
deletes on partitioned tables during initial runtime pruning.
Due to ExecInitModifyTable() changes as described above, EXPLAIN on a
plan for UPDATE and DELETE that uses runtime initial pruning no longer
lists partitions pruned during initial pruning.
Reviewed-by: Robert Haas <robertmhaas@gmail.com> (earlier versions)
Reviewed-by: Tomas Vondra <tomas@vondra.me>
Discussion: https://postgr.es/m/CA+HiwqFGkMSge6TgC9KQzde0ohpAycLQuV7ooitEEpbKB0O_mg@mail.gmail.com
|
|
This turns GistTranslateCompareType() into a callback function of the
gist index AM instead of a standalone function. The existing callers
are changed to use IndexAmTranslateCompareType(). This then makes
that code not hardcoded toward gist.
This means in particular that the temporal keys code is now
independent of gist. Also, this generalizes commit 74edabce7a3, so
other index access methods other than the previously hardcoded ones
could now work as REPLICA IDENTITY in a logical replication
subscriber.
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
|
|
Follow up to commit 630f9a43cec. The previous name had become
confusing, because it doesn't actually translate a strategy number but
a CompareType into a strategy number. We might add the inverse at
some point, which would then probably be called something like
GistTranslateStratnum.
Reviewed-by: Mark Dilger <mark.dilger@enterprisedb.com>
Discussion: https://www.postgresql.org/message-id/flat/E72EAA49-354D-4C2E-8EB9-255197F55330@enterprisedb.com
|
|
Consistently use "Size" (or size_t, or in some places int64 or double)
as the type for variables holding memory allocation sizes. In most
places variables' data types were fine already, but we had an ancient
habit of computing bytes from kilobytes-units GUCs with code like
"work_mem * 1024L". That risks overflow on Win64 where they did not
make "long" as wide as "size_t". We worked around that by restricting
such GUCs' ranges, so you couldn't set work_mem et al higher than 2GB
on Win64. This patch removes that restriction, after replacing such
calculations with "work_mem * (Size) 1024" or variants of that.
It should be noted that this patch was constructed by searching
outwards from the GUCs that have MAX_KILOBYTES as upper limit.
So I can't positively guarantee there are no other places doing
memory-size arithmetic in int or long variables. I do however feel
pretty confident that increasing MAX_KILOBYTES on Win64 is safe now.
Also, nothing in our code should be dealing in multiple-gigabyte
allocations without authorization from a relevant GUC, so it seems
pretty likely that this search caught everything that could be at
risk of overflow.
Author: Vladlen Popolitov <v.popolitov@postgrespro.ru>
Co-authored-by: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://postgr.es/m/1a01f0-66ec2d80-3b-68487680@27595217
|
|
Per buildfarm member koel
|
|
This commit builds on the prior change that moved PartitionPruneInfos
out of individual plan nodes into a list in PlannedStmt, making it
possible to initialize PartitionPruneStates without traversing the
plan tree and perform runtime initial pruning before ExecInitNode()
initializes the plan trees. These tasks are now handled in a new
routine, ExecDoInitialPruning(), which is called by InitPlan()
before calling ExecInitNode() on various plan trees.
ExecDoInitialPruning() performs the initial pruning and saves the
result -- a Bitmapset of indexes for surviving child subnodes -- in
es_part_prune_results, a list in EState.
PartitionPruneStates created for initial pruning are stored in
es_part_prune_states, another list in EState, for later use during
exec pruning. Both lists are parallel to es_part_prune_infos, which
holds the PartitionPruneInfos from PlannedStmt, enabling shared
indexing.
PartitionPruneStates initialized in ExecDoInitialPruning() now
include only the PartitionPruneContexts for initial pruning steps.
Exec pruning contexts are initialized later in
ExecInitPartitionExecPruning() when the parent plan node is
initialized, as the exec pruning step expressions depend on the parent
node's PlanState.
The existing function PartitionPruneFixSubPlanMap() has been
repurposed for this initialization to avoid duplicating a similar
loop structure for finding PartitionedRelPruningData to initialize
exec pruning contexts for. It has been renamed to
InitExecPruningContexts() to reflect its new primary responsibility.
The original logic to "fix subplan maps" remains intact but is now
encapsulated within the renamed function.
This commit removes two obsolete Asserts in partkey_datum_from_expr().
The ExprContext used for pruning expression evaluation is now
independent of the parent PlanState, making these Asserts unnecessary.
By centralizing pruning logic and decoupling it from the plan
initialization step (ExecInitNode()), this change sets the stage for
future patches that will use the result of initial pruning to
save the overhead of redundant processing for pruned partitions.
Reviewed-by: Robert Haas <robertmhaas@gmail.com>
Reviewed-by: Tomas Vondra <tomas@vondra.me>
Discussion: https://postgr.es/m/CA+HiwqFGkMSge6TgC9KQzde0ohpAycLQuV7ooitEEpbKB0O_mg@mail.gmail.com
|
|
Add BitmapTableScanSetup(), a helper which contains all of the code that
must be done on every scan of the table in a bitmap table scan. This
includes scanning the index, building the bitmap, and setting up the
scan descriptors.
Pushing this setup into a helper function makes BitmapHeapNext() more
readable.
Reviewed-by: Nazir Bilal Yavuz <byavuz81@gmail.com>
Discussion: https://postgr.es/m/CAN55FZ1vXu%2BZdT0_MM-i1vbTdfHHf0KR3cK6R5gs6dNNNpyrJw%40mail.gmail.com
|
|
Instead of deciding at runtime whether to read from casetest.value
or caseValue_datum, split EEOP_CASE_TESTVAL into two opcodes and
make the decision during expression compilation. Similarly for
EEOP_DOMAIN_TESTVAL. This actually results in net less code,
mainly because llvmjit_expr.c's code for handling these opcodes
gets shorter. The performance gain is doubtless negligible, but
this seems worth changing anyway on grounds of simplicity and
understandability.
Author: Andreas Karlsson <andreas@proxel.se>
Co-authored-by: Xing Guo <higuoxing@gmail.com>
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://postgr.es/m/CACpMh+AiBYAWn+D1aU7Rsy-V1tox06Cbc0H3qA7rwL5zdJ=anQ@mail.gmail.com
|
|
This moves PartitionPruneInfo from plan nodes to PlannedStmt,
simplifying traversal by centralizing all PartitionPruneInfo
structures in a single list in it, which holds all instances for the
main query and its subqueries. Instead of plan nodes (Append or
MergeAppend) storing PartitionPruneInfo pointers, they now reference
an index in this list.
A bitmapset field is added to PartitionPruneInfo to store the RT
indexes corresponding to the apprelids field in Append or MergeAppend.
This allows execution pruning logic to verify that it operates on the
correct plan node, mainly to facilitate debugging.
Duplicated code in set_append_references() and
set_mergeappend_references() is refactored into a new function,
register_pruneinfo(). This updates RT indexes by applying rtoffet
and adds PartitionPruneInfo to the global list in PlannerGlobal.
By allowing pruning to be performed without traversing the plan tree,
this change lays the groundwork for runtime initial pruning to occur
independently of plan tree initialization.
Reviewed-by: Alvaro Herrera <alvherre@alvh.no-ip.org> (earlier version)
Reviewed-by: Robert Haas <robertmhaas@gmail.com>
Reviewed-by: Tomas Vondra <tomas@vondra.me>
Discussion: https://postgr.es/m/CA+HiwqFGkMSge6TgC9KQzde0ohpAycLQuV7ooitEEpbKB0O_mg@mail.gmail.com
|
|
This commit refactors ExecScan() by moving its tuple-fetching,
filtering, and projection logic into an inline-able function,
ExecScanExtended(), defined in src/include/executor/execScan.h.
ExecScanExtended() accepts parameters for EvalPlanQual state,
qualifiers (ExprState), and projection (ProjectionInfo).
Specialized variants of the execution function of a given Scan node
(for example, ExecSeqScan() for SeqScan) can then pass const-NULL for
unused parameters. This allows the compiler to inline the logic and
eliminate unnecessary branches or checks. Each variant function thus
contains only the necessary code, optimizing execution for scans
where these features are not needed.
The variant function to be used is determined in the ExecInit*()
function of the node and assigned to the ExecProcNode function pointer
in the node's PlanState, effectively turning runtime checks and
conditional branches on the NULLness of epqstate, qual, and projInfo
into static ones, provided the compiler successfully eliminates
unnecessary checks from the inlined code of ExecScanExtended().
Currently, only ExecSeqScan() is modified to take advantage of this
inline-ability. Other Scan nodes might benefit from such specialized
variant functions but that is left as future work.
Benchmarks performed by Junwang Zhao, David Rowley and myself show up
to a 5% reduction in execution time for queries that rely heavily on
Seq Scans. The most significant improvements were observed in
scenarios where EvalPlanQual, qualifiers, and projection were not
required, but other cases also benefit from reduced runtime overhead
due to the inlining and removal of unnecessary code paths.
The idea for this patch first came from Andres Freund in an off-list
discussion. The refactoring approach implemented here is based on a
proposal by David Rowley, significantly improving upon the patch I
(amitlan) initially proposed.
Suggested-by: Andres Freund <andres@anarazel.de>
Co-authored-by: David Rowley <dgrowleyml@gmail.com>
Reviewed-by: David Rowley <dgrowleyml@gmail.com>
Reviewed-by: Junwang Zhao <zhjwpku@gmail.com>
Tested-by: Junwang Zhao <zhjwpku@gmail.com>
Tested-by: David Rowley <dgrowleyml@gmail.com>
Discussion: https://postgr.es/m/CA+HiwqGaH-otvqW_ce-paL=96JvU4j+Xbuk+14esJNDwefdkOg@mail.gmail.com
|
|
This allows the RETURNING list of INSERT/UPDATE/DELETE/MERGE queries
to explicitly return old and new values by using the special aliases
"old" and "new", which are automatically added to the query (if not
already defined) while parsing its RETURNING list, allowing things
like:
RETURNING old.colname, new.colname, ...
RETURNING old.*, new.*
Additionally, a new syntax is supported, allowing the names "old" and
"new" to be changed to user-supplied alias names, e.g.:
RETURNING WITH (OLD AS o, NEW AS n) o.colname, n.colname, ...
This is useful when the names "old" and "new" are already defined,
such as inside trigger functions, allowing backwards compatibility to
be maintained -- the interpretation of any existing queries that
happen to already refer to relations called "old" or "new", or use
those as aliases for other relations, is not changed.
For an INSERT, old values will generally be NULL, and for a DELETE,
new values will generally be NULL, but that may change for an INSERT
with an ON CONFLICT ... DO UPDATE clause, or if a query rewrite rule
changes the command type. Therefore, we put no restrictions on the use
of old and new in any DML queries.
Dean Rasheed, reviewed by Jian He and Jeff Davis.
Discussion: https://postgr.es/m/CAEZATCWx0J0-v=Qjc6gXzR=KtsdvAE7Ow=D=mu50AgOe+pvisQ@mail.gmail.com
|
|
This changes commit 7406ab623fe in that the gist strategy number
mapping support function is changed to use the CompareType enum as
input, instead of the "well-known" RT*StrategyNumber strategy numbers.
This is a bit cleaner, since you are not dealing with two sets of
strategy numbers. Also, this will enable us to subsume this system
into a more general system of using CompareType to define operator
semantics across index methods.
Discussion: https://www.postgresql.org/message-id/flat/E72EAA49-354D-4C2E-8EB9-255197F55330@enterprisedb.com
|
|
RowCompareType served as a way to describe the fundamental meaning of
an operator, notionally independent of an operator class (although so
far this was only really supported for btrees). Its original purpose
was for use inside RowCompareExpr, and it has also found some small
use outside, such as for get_op_btree_interpretation().
We want to expand this now, as a more general way to describe operator
semantics for other index access methods, including gist (to improve
GistTranslateStratnum()) and others not written yet. To avoid future
confusion, we rename the type to CompareType and the symbols from
ROWCOMPARE_XXX to COMPARE_XXX to reflect their more general purpose.
Reviewed-by: Mark Dilger <mark.dilger@enterprisedb.com>
Discussion: https://www.postgresql.org/message-id/flat/E72EAA49-354D-4C2E-8EB9-255197F55330@enterprisedb.com
|
|
This reverts commit e0ece2a981ee9068f50c4423e303836c2585eb02 due to
performance regressions.
Reported-by: David Rowley
|
|
Oversight in commit e0ece2a98.
|
|
This adds support for the NOT ENFORCED/ENFORCED flag for constraints,
with support for check constraints.
The plan is to eventually support this for foreign key constraints,
where it is typically more useful.
Note that CHECK constraints do not currently support ALTER operations,
so changing the enforceability of an existing constraint isn't
possible without dropping and recreating it. This could be added
later.
Author: Amul Sul <amul.sul@enterprisedb.com>
Reviewed-by: Peter Eisentraut <peter@eisentraut.org>
Reviewed-by: jian he <jian.universality@gmail.com>
Tested-by: Triveni N <triveni.n@enterprisedb.com>
Discussion: https://www.postgresql.org/message-id/flat/CAAJ_b962c5AcYW9KUt_R_ER5qs3fUGbe4az-SP-vuwPS-w-AGA@mail.gmail.com
|
|
|
|
Previously, the caller needed to allocate the memory and the
TupleHashTable would store a pointer to it. That wastes space for the
palloc overhead as well as the size of the pointer itself.
Now, the TupleHashTable relies on the caller to correctly specify the
additionalsize, and allocates that amount of space. The caller can
then request a pointer into that space.
Discussion: https://postgr.es/m/b9cbf0219a9859dc8d240311643ff4362fd9602c.camel@j-davis.com
Reviewed-by: Heikki Linnakangas
|
|
Functions hash_agg_entry_size() and build_hash_tables() make use of
those values for memory size estimates.
Because this change only affects memory estimates, don't backpatch.
Discussion: https://postgr.es/m/7530bd8783b1a78d53a3c70383e38d8da0a5ffe5.camel%40j-davis.com
|
|
Provide additionalsize argument, which can affect the calculations for
'nbuckets'. Also, future work for Hash Aggregation will rely on the
correct additionalsize.
Discussion: https://postgr.es/m/7530bd8783b1a78d53a3c70383e38d8da0a5ffe5.camel%40j-davis.com
|
|
Discussion: https://postgr.es/m/7530bd8783b1a78d53a3c70383e38d8da0a5ffe5.camel%40j-davis.com
|
|
CHUNKHDRSZ was defined as 16 bytes, which was true when that code went in,
but since c6e0fe1f2, 8 is a more accurate value. Here we adjust it to use
sizeof(MemoryChunk), which is normally 8, or 16 for cassert builds.
c6e0fe1f2 first appeared in v16, so this is technically wrong in v16 up
to master, but let's apply this only to master as adjusting this does
influence the estimated number of batches in the aggregate costing code
and we don't want to cause plan instability in released versions.
Reviewed-by: Tom Lane
Discussion: https://postgr.es/m/CAApHDvpMpRQvsTqZo3FinXkgytwxwF8sCyZm83xDj-1s_hLe+w@mail.gmail.com
|
|
Backpatch-through: 13
|
|
This adjusts slot_deform_heap_tuple() to add special-case loops to
eliminate much of the branching that was done within the body of the
main deform loop.
Previously, while looping over each attribute to deform,
slot_deform_heap_tuple() would always recheck if the given attribute was
NULL by looking at HeapTupleHasNulls() and if so, went on to check the
tuple's NULL bitmap. Since many tuples won't contain any NULLs, we can
just check HeapTupleHasNulls() once and when there are no NULLs, use a
more compact version of the deforming loop which contains no NULL checking
code at all.
The same is possible for the "slow" mode checking part of the loop. That
variable was checked several times for each attribute, once to determine
if the offset to the attribute value could be taken from the attcacheoff,
and again to check if the offset could be cached for next time.
These "slow" checks can mostly be eliminated by instead having multiple
loops. Initially, we can start in the non-slow loop and break out of
that loop if and only if we must stop caching the offset. This
eliminates branching for both slow and non-slow deforming methods. The
amount of code required for the no nulls / non-slow version is very
small. It's possible to have separate loops like this due to the fact
that once we move into slow mode, we never need to switch back into
non-slow mode for a given tuple.
We have the compiler take care of writing out the multiple required
loops by having a pg_attribute_always_inline function which gets called
various times passing in constant values for the "slow" and "hasnulls"
parameters. This allows the compiler to eliminate const-false branches
and remove comparisons for const-true ones.
This commit has shown overall query performance increases of around 5-20%
in deform-heavy OLAP-type workloads.
Author: David Rowley
Reviewed-by: Victor Yegorov
Discussion: https://postgr.es/m/CAGnEbog92Og2CpC2S8=g_HozGsWtt_3kRS1sXjLz0jKSoCNfLw@mail.gmail.com
Discussion: https://postgr.es/m/CAApHDvo9e0XG71WrefYaRv5n4xNPLK4k8LjD0mSR3c9KR2vi2Q@mail.gmail.com
|
|
Here we convert CompactAttribute.attalign from a char, which is directly
derived from pg_attribute.attalign into a uint8, which stores the number
of bytes to align the column's value by in the tuple.
This allows tuple deformation and tuple size calculations to move away
from using the inefficient att_align_nominal() macro, which manually
checks each TYPALIGN_* char to translate that into the alignment bytes
for the given type. Effectively, this commit changes those to TYPEALIGN
calls, which are branchless and only perform some simple arithmetic with
some bit-twiddling.
The removed branches were often mispredicted by CPUs, especially so in
real-world tables which often contain a mishmash of different types
with different alignment requirements.
Author: David Rowley
Reviewed-by: Andres Freund, Victor Yegorov
Discussion: https://postgr.es/m/CAApHDvrBztXP3yx=NKNmo3xwFAFhEdyPnvrDg3=M0RhDs+4vYw@mail.gmail.com
|
|
The new compact_attrs array stores a few select fields from
FormData_pg_attribute in a more compact way, using only 16 bytes per
column instead of the 104 bytes that FormData_pg_attribute uses. Using
CompactAttribute allows performance-critical operations such as tuple
deformation to be performed without looking at the FormData_pg_attribute
element in TupleDesc which means fewer cacheline accesses.
For some workloads, tuple deformation can be the most CPU intensive part
of processing the query. Some testing with 16 columns on a table
where the first column is variable length showed around a 10% increase in
transactions per second for an OLAP type query performing aggregation on
the 16th column. However, in certain cases, the increases were much
higher, up to ~25% on one AMD Zen4 machine.
This also makes pg_attribute.attcacheoff redundant. A follow-on commit
will remove it, thus shrinking the FormData_pg_attribute struct by 4
bytes.
Author: David Rowley
Reviewed-by: Andres Freund, Victor Yegorov
Discussion: https://postgr.es/m/CAApHDvrBztXP3yx=NKNmo3xwFAFhEdyPnvrDg3=M0RhDs+4vYw@mail.gmail.com
|
|
It was reasonable to preserve the old API of BuildTupleHashTable()
in the back branches, but in HEAD we should actively discourage use
of that version. There are no remaining callers in core, so just
get rid of it. Then rename BuildTupleHashTableExt() back to
BuildTupleHashTable().
While at it, fix up the miserably-poorly-maintained header comment
for BuildTupleHashTable[Ext]. It looks like more than one patch in
this area has had the opinion that updating comments is beneath them.
Discussion: https://postgr.es/m/538343.1734646986@sss.pgh.pa.us
|
|
Append, MergeAppend, and RecursiveUnion can all use the support
functions added in commit 276279295. The first two can report a
fixed result slot type if all their children return the same fixed
slot type. That does nothing for the append step itself, but might
allow optimizations in the parent plan node. RecursiveUnion can
optimize tuple hash table operations in the same way as SetOp now
does.
Patch by me; thanks to Richard Guo and David Rowley for review.
Discussion: https://postgr.es/m/1850138.1731549611@sss.pgh.pa.us
|
|
The original design for set operations involved appending the two
input relations into one and adding a flag column that allows
distinguishing which side each row came from. Then the SetOp node
pries them apart again based on the flag. This is bizarre. The
only apparent reason to do it is that when sorting, we'd only need
one Sort node not two. But since sorting is at least O(N log N),
sorting all the data is actually worse than sorting each side
separately --- plus, we have no chance of taking advantage of
presorted input. On top of that, adding the flag column frequently
requires an additional projection step that adds cycles, and then
the Append node isn't free either. Let's get rid of all of that
and make the SetOp node have two separate children, using the
existing outerPlan/innerPlan infrastructure.
This initial patch re-implements nodeSetop.c and does a bare minimum
of work on the planner side to generate correctly-shaped plans.
In particular, I've tried not to change the cost estimates here,
so that the visible changes in the regression test results will only
involve removal of useless projection steps and not any changes in
whether to use sorted vs hashed mode.
For SORTED mode, we combine successive identical tuples from each
input into groups, and then merge-join the groups. The tuple
comparisons now use SortSupport instead of simple equality, but
the group-formation part should involve roughly the same number of
tuple comparisons as before. The cross-comparisons between left and
right groups probably add to that, but I'm not sure to quantify how
many more comparisons we might need.
For HASHED mode, nodeSetop's logic is almost the same as before,
just refactored into two separate loops instead of one loop that
has an assumption that it will see all the left-hand inputs first.
In both modes, I added early-exit logic to not bother reading the
right-hand relation if the left-hand input is empty, since neither
INTERSECT nor EXCEPT modes can produce any output if the left input
is empty. This could have been done before in the hashed mode, but
not in sorted mode. Sorted mode can also stop as soon as it exhausts
the left input; any remaining right-hand tuples cannot have matches.
Also, this patch adds some infrastructure for detecting whether
child plan nodes all output the same type of tuple table slot.
If they do, the hash table logic can use slightly more efficient
code based on assuming that that's the input slot type it will see.
We'll make use of that infrastructure in other plan node types later.
Patch by me; thanks to Richard Guo and David Rowley for review.
Discussion: https://postgr.es/m/1850138.1731549611@sss.pgh.pa.us
|
|
1a0da347a7ac98db replaced Bitmap Table Scan's separate private and
shared bitmap iterators with a unified iterator. It accidentally set up
the prefetch iterator twice for non-parallel bitmap table scans. Remove
the extra set up call to tbm_begin_iterate().
|
|
1a0da347a7ac98db replaced Bitmap Table Scan's individual private and
shared iterators with a unified iterator. It neglected, however, to
check if the iterator had already been cleaned up before doing so on
rescan. Add this check both on rescan and end scan to be safe.
Reported-by: Richard Guo
Author: Richard Guo
Discussion: https://postgr.es/m/CAMbWs48nrhcLY1kcd-u9oD%2B6yiS631F_8Fx8ZGsO-BYDwH%2Bbyw%40mail.gmail.com
|
|
8f4ee9626 fixed an old Assert failure that could happen when the slot
type used to look up the hash table for BuildTupleHashTableExt() users
wasn't a TTSOpsMinimalTuple slot. The fix for that in the back branches
had to be to pass the TupleTableSlotOps as NULL, however in master,
since we have the inputOps parameter as was added by d96d1d515, we can
pass that down instead.
At least one caller uses a fixed slot that's always TTSOpsVirtual, so
passing down inputOps for these cases allows ExecBuildGroupingEqual() to
skip adding the EEOP_INNER_FETCHSOME ExprEvalStep.
This should increase the performance of hashed subplans very slightly.
Author: Tom Lane, David Rowley
Discussion: https://postgr.es/m/2543667.1734483723@sss.pgh.pa.us
|
|
If the non-recursive part of a recursive CTE ended up using
TTSOpsBufferHeapTuple as the table slot type, then a duplicate value
could cause an Assert failure in CheckOpSlotCompatibility() when
checking the hash table for the duplicate value. The expected slot type
for the deform step was TTSOpsMinimalTuple so the Assert failed when the
TTSOpsBufferHeapTuple slot was used.
This is a long-standing bug which we likely didn't notice because it
seems much more likely that the non-recursive term would have required
projection and used a TTSOpsVirtual slot, which CheckOpSlotCompatibility
is ok with.
There doesn't seem to be any harm done here other than the Assert
failure. Both TTSOpsMinimalTuple and TTSOpsBufferHeapTuple slot types
require tuple deformation, so the EEOP_*_FETCHSOME ExprState step would
have properly existed in the ExprState.
The solution is to pass NULL for the ExecBuildGroupingEqual's 'lops'
parameter. This means the ExprState's EEOP_*_FETCHSOME step won't
expect a fixed slot type. This makes CheckOpSlotCompatibility() happy as
no checking is performed when the ExprEvalStep is not expecting a fixed
slot type.
Reported-by: Richard Guo
Reviewed-by: Tom Lane
Discussion: https://postgr.es/m/CAMbWs4-8U9q2LAtf8+ghV11zeUReA3AmrYkxzBEv0vKnDxwkKA@mail.gmail.com
Backpatch-through: 13, all supported versions
|
|
With the repurposing of TBMIterator as an interface for both parallel
and serial iteration through TIDBitmaps in commit 7f9d4187e7bab10329cc,
bitmap table scans may now use it.
Modify bitmap table scan code to use the TBMIterator. This requires
moving around a bit of code, so a few variables are initialized
elsewhere.
Author: Melanie Plageman
Reviewed-by: Tomas Vondra
Discussion: https://postgr.es/m/c736f6aa-8b35-4e20-9621-62c7c82e2168%40vondra.me
|
|
Add and use TBMPrivateIterator, which replaces the current TBMIterator
for serial use cases, and repurpose TBMIterator to be a unified
interface for both the serial ("private") and parallel ("shared") TID
Bitmap iterator interfaces. This encapsulation simplifies call sites for
callers supporting both parallel and serial TID Bitmap access.
TBMIterator is not yet used in this commit.
Author: Melanie Plageman
Reviewed-by: Tomas Vondra, Heikki Linnakangas
Discussion: https://postgr.es/m/063e4eb4-32d9-439e-a0b1-75565a9835a8%40iki.fi
|
|
0f5738202 adjusted the execGrouping.c code so it made use of ExprStates to
generate hash values. That commit made a wrong assumption that the slot
type to pass to ExecBuildHash32FromAttrs() is always &TTSOpsMinimalTuple.
That's not the case as the slot type depends on the slot type passed to
LookupTupleHashEntry(), which for nodeRecursiveunion.c, could be any of
the current slot types.
Here we fix this by adding a new parameter to BuildTupleHashTableExt()
to allow the slot type to be passed in. In the case of nodeSubplan.c
and nodeAgg.c the slot type is always &TTSOpsVirtual, so for both of
those cases, it's beneficial to pass the known slot type as that allows
ExecBuildHash32FromAttrs() to skip adding the tuple deform step to the
resulting ExprState. Another possible fix would have been to have
ExecBuildHash32FromAttrs() set "fetch.kind" to NULL so that
ExecComputeSlotInfo() always determines the EEOP_INNER_FETCHSOME is
required, however, that option isn't favorable as slows down aggregation
and hashed subplan evaluation due to the extra (needless) deform step.
Thanks to Nathan Bossart for bisecting to find the offending commit
based on Paul's report.
Reported-by: Paul Ramsey <pramsey@cleverelephant.ca>
Discussion: https://postgr.es/m/99F064C1-B3EB-4BE7-97D2-D2A0AA487A71@cleverelephant.ca
|
|
This speeds up obtaining hash values for GROUP BY and hashed SubPlans by
using the ExprState support for hashing, thus allowing JIT compilation for
obtaining hash values for these operations.
This, even without JIT compilation, has been shown to improve Hash
Aggregate performance in some cases by around 15% and hashed NOT IN
queries in one case by over 30%, however, real-world cases are likely to
see smaller gains as the test cases used were purposefully designed to
have high hashing overheads by keeping the hash table small to prevent
additional memory overheads that would be a factor when working with large
hash tables.
In passing, fix a hypothetical bug in ExecBuildHash32Expr() so that the
initial value is stored directly in the ExprState's result field if
there are no expressions to hash. None of the current users of this
function use an initial value, so the bug is only hypothetical.
Reviewed-by: Andrei Lepikhov <lepihov@gmail.com>
Discussion: https://postgr.es/m/CAApHDvpYSO3kc9UryMevWqthTBrxgfd9djiAjKHMPUSQeX9vdQ@mail.gmail.com
|
|
Hashing of a single Var is a very common operation for ExprState to
perform. Here we add dedicated ExecJust* functions which helps speed up
Hash Joins by removing the interpretation overhead in ExecInterpExpr().
This change currently only affects Hash Joins on a single column. Hash
Joins with multiple join keys or an expression still run through
ExecInterpExpr().
Some testing has shown up to 10% query performance increases on recent AMD
hardware and nearly 7% increase on an Apple M2 for a query performing a
hash join with a large number of lookups on a small hash table.
This change was extracted from a larger patch which adjusts GROUP BY /
hashed subplans / hashed set operations to use ExprState hashing.
Discussion: https://postgr.es/m/CAApHDvr8Zc0ZgzVoCZLdHGOFNhiJeQ6vrUcS9V7N23zMWQb-eA@mail.gmail.com
|
