* Commit 3048898e dropped -ING from PHJ wait event names. Update the
corresponding barrier phases names to match.
* Rename the "DONE" phases to "FREE". That's symmetrical with
"ALLOCATE", and names the activity that actually happens in that phase
(as we do for the other phases) rather than a state. The bug fixed by
commit 8d578b9b might have been more obvious with this name.
* Rename the batch/bucket growth barriers' "ALLOCATE" phases to
"REALLOCATE", a better description of what they do.
* Update the high level comments about phases to highlight phases
are executed by a single process with an asterisk (mostly memory
management phases).
No behavior change, as this is just improving internal identifiers. The
only user-visible sign of this is that a couple of wait events' display
names change from "...Allocate" to "...Reallocate" in pg_stat_activity,
to stay in sync with the internal names.
Reviewed-by: Melanie Plageman <melanieplageman@gmail.com>
Discussion: https://postgr.es/m/CA%2BhUKG%2BMDpwF2Eo2LAvzd%3DpOh81wUTsrwU1uAwR-v6OGBB6%2B7g%40mail.gmail.com
With unlucky timing and parallel_leader_participation=off (not the
default), PHJ could attempt to access per-batch shared state just as it
was being freed. There was code intended to prevent that by checking
for a cleared pointer, but it was racy. Fix, by introducing an extra
barrier phase. The new phase PHJ_BUILD_RUNNING means that it's safe to
access the per-batch state to find a batch to help with, and
PHJ_BUILD_DONE means that it is too late. The last to detach will free
the array of per-batch state as before, but now it will also atomically
advance the phase, so that late attachers can avoid the hazard. This
mirrors the way per-batch hash tables are freed (see phases
PHJ_BATCH_PROBING and PHJ_BATCH_DONE).
An earlier attempt to fix this (commit 3b8981b6, later reverted) missed
one special case. When the inner side is empty (the "empty inner
optimization), the build barrier would only make it to
PHJ_BUILD_HASHING_INNER phase before workers attempted to detach from
the hashtable. In that case, fast-forward the build barrier to
PHJ_BUILD_RUNNING before proceeding, so that our later assertions hold
and we can still negotiate who is cleaning up.
Revealed by build farm failures, where BarrierAttach() failed a sanity
check assertion, because the memory had been clobbered by dsa_free().
In non-assert builds, the result could be a segmentation fault.
Back-patch to all supported releases.
Author: Thomas Munro <thomas.munro@gmail.com>
Author: Melanie Plageman <melanieplageman@gmail.com>
Reported-by: Michael Paquier <michael@paquier.xyz>
Reported-by: David Geier <geidav.pg@gmail.com>
Tested-by: David Geier <geidav.pg@gmail.com>
Discussion: https://postgr.es/m/20200929061142.GA29096%40paquier.xyz
Commit 3048898e dropped -ING from some wait event names that correspond
to barrier phases. Update the phases' names to match.
While we're here making cosmetic changes, also rename "DONE" to "FREE".
That pairs better with "ALLOCATE", and describes the activity that
actually happens in that phase (as we do for the other phases) rather
than describing a state. The distinction is clearer after bugfix commit
3b8981b6 split the phase into two. As for the growth barriers, rename
their "ALLOCATE" phase to "REALLOCATE", which is probably a better
description of what happens then. Also improve the comments about
the phases a bit.
Discussion: https://postgr.es/m/CA%2BhUKG%2BMDpwF2Eo2LAvzd%3DpOh81wUTsrwU1uAwR-v6OGBB6%2B7g%40mail.gmail.com
With very unlucky timing and parallel_leader_participation off, PHJ
could attempt to access per-batch state just as it was being freed.
There was code intended to prevent that by checking for a cleared
pointer, but it was buggy.
Fix, by introducing an extra barrier phase. The new phase
PHJ_BUILD_RUNNING means that it's safe to access the per-batch state to
find a batch to help with, and PHJ_BUILD_DONE means that it is too late.
The last to detach will free the array of per-batch state as before, but
now it will also atomically advance the phase at the same time, so that
late attachers can avoid the hazard, without the data race. This
mirrors the way per-batch hash tables are freed (see phases
PHJ_BATCH_PROBING and PHJ_BATCH_DONE).
Revealed by a one-off build farm failure, where BarrierAttach() failed a
sanity check assertion, because the memory had been clobbered by
dsa_free().
Back-patch to 11, where the code arrived.
Reported-by: Michael Paquier <michael@paquier.xyz>
Discussion: https://postgr.es/m/20200929061142.GA29096%40paquier.xyz
Add a GUC that acts as a multiplier on work_mem. It gets applied when
sizing executor node hash tables that were previously size constrained
using work_mem alone.
The new GUC can be used to preferentially give hash-based nodes more
memory than the generic work_mem limit. It is intended to enable admin
tuning of the executor's memory usage. Overall system throughput and
system responsiveness can be improved by giving hash-based executor
nodes more memory (especially over sort-based alternatives, which are
often much less sensitive to being memory constrained).
The default value for hash_mem_multiplier is 1.0, which is also the
minimum valid value. This means that hash-based nodes continue to apply
work_mem in the traditional way by default.
hash_mem_multiplier is generally useful. However, it is being added now
due to concerns about hash aggregate performance stability for users
that upgrade to Postgres 13 (which added disk-based hash aggregation in
commit 1f39bce0). While the old hash aggregate behavior risked
out-of-memory errors, it is nevertheless likely that many users actually
benefited. Hash agg's previous indifference to work_mem during query
execution was not just faster; it also accidentally made aggregation
resilient to grouping estimate problems (at least in cases where this
didn't create destabilizing memory pressure).
hash_mem_multiplier can provide a certain kind of continuity with the
behavior of Postgres 12 hash aggregates in cases where the planner
incorrectly estimates that all groups (plus related allocations) will
fit in work_mem/hash_mem. This seems necessary because hash-based
aggregation is usually much slower when only a small fraction of all
groups can fit. Even when it isn't possible to totally avoid hash
aggregates that spill, giving hash aggregation more memory will reliably
improve performance (the same cannot be said for external sort
operations, which appear to be almost unaffected by memory availability
provided it's at least possible to get a single merge pass).
The PostgreSQL 13 release notes should advise users that increasing
hash_mem_multiplier can help with performance regressions associated
with hash aggregation. That can be taken care of by a later commit.
Author: Peter Geoghegan
Reviewed-By: Álvaro Herrera, Jeff Davis
Discussion: https://postgr.es/m/20200625203629.7m6yvut7eqblgmfo@alap3.anarazel.de
Discussion: https://postgr.es/m/CAH2-WzmD%2Bi1pG6rc1%2BCjc4V6EaFJ_qSuKCCHVnH%3DoruqD-zqow%40mail.gmail.com
Backpatch: 13-, where disk-based hash aggregation was introduced.
This adds a flag "deterministic" to collations. If that is false,
such a collation disables various optimizations that assume that
strings are equal only if they are byte-wise equal. That then allows
use cases such as case-insensitive or accent-insensitive comparisons
or handling of strings with different Unicode normal forms.
This functionality is only supported with the ICU provider. At least
glibc doesn't appear to have any locales that work in a
nondeterministic way, so it's not worth supporting this for the libc
provider.
The term "deterministic comparison" in this context is from Unicode
Technical Standard #10
(https://unicode.org/reports/tr10/#Deterministic_Comparison).
This patch makes changes in three areas:
- CREATE COLLATION DDL changes and system catalog changes to support
this new flag.
- Many executor nodes and auxiliary code are extended to track
collations. Previously, this code would just throw away collation
information, because the eventually-called user-defined functions
didn't use it since they only cared about equality, which didn't
need collation information.
- String data type functions that do equality comparisons and hashing
are changed to take the (non-)deterministic flag into account. For
comparison, this just means skipping various shortcuts and tie
breakers that use byte-wise comparison. For hashing, we first need
to convert the input string to a canonical "sort key" using the ICU
analogue of strxfrm().
Reviewed-by: Daniel Verite <daniel@manitou-mail.org>
Reviewed-by: Peter Geoghegan <pg@bowt.ie>
Discussion: https://www.postgresql.org/message-id/flat/1ccc668f-4cbc-0bef-af67-450b47cdfee7@2ndquadrant.com
The previous coding relied (without any documentation) on the data[]
member of HashMemoryChunkData being at a MAXALIGN'ed offset. If it
was not, the tuples would not be maxaligned either, leading to failures
on alignment-picky machines. While there seems to be no live bug on any
platform we support, this is clearly pretty fragile: any addition to or
rearrangement of the fields in HashMemoryChunkData could break it.
Let's remove the hazard by getting rid of the data[] member and instead
using pointer arithmetic with an explicitly maxalign'ed offset.
Discussion: https://postgr.es/m/14483.1514938129@sss.pgh.pa.us
Introduce parallel-aware hash joins that appear in EXPLAIN plans as Parallel
Hash Join with Parallel Hash. While hash joins could already appear in
parallel queries, they were previously always parallel-oblivious and had a
partial subplan only on the outer side, meaning that the work of the inner
subplan was duplicated in every worker.
After this commit, the planner will consider using a partial subplan on the
inner side too, using the Parallel Hash node to divide the work over the
available CPU cores and combine its results in shared memory. If the join
needs to be split into multiple batches in order to respect work_mem, then
workers process different batches as much as possible and then work together
on the remaining batches.
The advantages of a parallel-aware hash join over a parallel-oblivious hash
join used in a parallel query are that it:
* avoids wasting memory on duplicated hash tables
* avoids wasting disk space on duplicated batch files
* divides the work of building the hash table over the CPUs
One disadvantage is that there is some communication between the participating
CPUs which might outweigh the benefits of parallelism in the case of small
hash tables. This is avoided by the planner's existing reluctance to supply
partial plans for small scans, but it may be necessary to estimate
synchronization costs in future if that situation changes. Another is that
outer batch 0 must be written to disk if multiple batches are required.
A potential future advantage of parallel-aware hash joins is that right and
full outer joins could be supported, since there is a single set of matched
bits for each hashtable, but that is not yet implemented.
A new GUC enable_parallel_hash is defined to control the feature, defaulting
to on.
Author: Thomas Munro
Reviewed-By: Andres Freund, Robert Haas
Tested-By: Rafia Sabih, Prabhat Sahu
Discussion:
https://postgr.es/m/CAEepm=2W=cOkiZxcg6qiFQP-dHUe09aqTrEMM7yJDrHMhDv_RA@mail.gmail.comhttps://postgr.es/m/CAEepm=37HKyJ4U6XOLi=JgfSHM3o6B-GaeO-6hkOmneTDkH+Uw@mail.gmail.com
Change pg_bsd_indent to follow upstream rules for placement of comments
to the right of code, and remove pgindent hack that caused comments
following #endif to not obey the general rule.
Commit e3860ffa4d wasn't actually using
the published version of pg_bsd_indent, but a hacked-up version that
tried to minimize the amount of movement of comments to the right of
code. The situation of interest is where such a comment has to be
moved to the right of its default placement at column 33 because there's
code there. BSD indent has always moved right in units of tab stops
in such cases --- but in the previous incarnation, indent was working
in 8-space tab stops, while now it knows we use 4-space tabs. So the
net result is that in about half the cases, such comments are placed
one tab stop left of before. This is better all around: it leaves
more room on the line for comment text, and it means that in such
cases the comment uniformly starts at the next 4-space tab stop after
the code, rather than sometimes one and sometimes two tabs after.
Also, ensure that comments following #endif are indented the same
as comments following other preprocessor commands such as #else.
That inconsistency turns out to have been self-inflicted damage
from a poorly-thought-through post-indent "fixup" in pgindent.
This patch is much less interesting than the first round of indent
changes, but also bulkier, so I thought it best to separate the effects.
Discussion: https://postgr.es/m/E1dAmxK-0006EE-1r@gemulon.postgresql.org
Discussion: https://postgr.es/m/30527.1495162840@sss.pgh.pa.us
The new indent version includes numerous fixes thanks to Piotr Stefaniak.
The main changes visible in this commit are:
* Nicer formatting of function-pointer declarations.
* No longer unexpectedly removes spaces in expressions using casts,
sizeof, or offsetof.
* No longer wants to add a space in "struct structname *varname", as
well as some similar cases for const- or volatile-qualified pointers.
* Declarations using PG_USED_FOR_ASSERTS_ONLY are formatted more nicely.
* Fixes bug where comments following declarations were sometimes placed
with no space separating them from the code.
* Fixes some odd decisions for comments following case labels.
* Fixes some cases where comments following code were indented to less
than the expected column 33.
On the less good side, it now tends to put more whitespace around typedef
names that are not listed in typedefs.list. This might encourage us to
put more effort into typedef name collection; it's not really a bug in
indent itself.
There are more changes coming after this round, having to do with comment
indentation and alignment of lines appearing within parentheses. I wanted
to limit the size of the diffs to something that could be reviewed without
one's eyes completely glazing over, so it seemed better to split up the
changes as much as practical.
Discussion: https://postgr.es/m/E1dAmxK-0006EE-1r@gemulon.postgresql.org
Discussion: https://postgr.es/m/30527.1495162840@sss.pgh.pa.us
Limit the size of the hashtable pointer array to not more than
MaxAllocSize, per reports from Kouhei Kaigai and others of "invalid memory
alloc request size" failures. There was discussion of allowing the array
to get larger than that by using the "huge" palloc API, but so far no proof
that that is actually a good idea, and at this point in the 9.5 cycle major
changes from old behavior don't seem like the way to go.
Fix a rather serious secondary bug in the new code, which was that it
didn't ensure nbuckets remained a power of 2 when recomputing it for the
multiple-batch case.
Clean up sloppy division of labor between ExecHashIncreaseNumBuckets and
its sole call site.
Replace some bogus "x[1]" declarations with "x[FLEXIBLE_ARRAY_MEMBER]".
Aside from being more self-documenting, this should help prevent bogus
warnings from static code analyzers and perhaps compiler misoptimizations.
This patch is just a down payment on eliminating the whole problem, but
it gets rid of a lot of easy-to-fix cases.
Note that the main problem with doing this is that one must no longer rely
on computing sizeof(the containing struct), since the result would be
compiler-dependent. Instead use offsetof(struct, lastfield). Autoconf
also warns against spelling that offsetof(struct, lastfield[0]).
Michael Paquier, review and additional fixes by me.
If we expect batching at the very beginning, we size nbuckets for
"full work_mem" (see how many tuples we can get into work_mem,
while not breaking NTUP_PER_BUCKET threshold).
If we expect to be fine without batching, we start with the 'right'
nbuckets and track the optimal nbuckets as we go (without actually
resizing the hash table). Once we hit work_mem (considering the
optimal nbuckets value), we keep the value.
At the end of the first batch, we check whether (nbuckets !=
nbuckets_optimal) and resize the hash table if needed. Also, we
keep this value for all batches (it's OK because it assumes full
work_mem, and it makes the batchno evaluation trivial). So the
resize happens only once.
There could be cases where it would improve performance to allow
the NTUP_PER_BUCKET threshold to be exceeded to keep everything in
one batch rather than spilling to a second batch, but attempts to
generate such a case have so far been unsuccessful; that issue may
be addressed with a follow-on patch after further investigation.
Tomas Vondra with minor format and comment cleanup by me
Reviewed by Robert Haas, Heikki Linnakangas, and Kevin Grittner
Instead of palloc'ing each HashJoinTuple individually, allocate 32kB chunks
and pack the tuples densely in the chunks. This avoids the AllocChunk
header overhead, and the space wasted by standard allocator's habit of
rounding sizes up to the nearest power of two.
This doesn't contain any planner changes, because the planner's estimate of
memory usage ignores the palloc overhead. Now that the overhead is smaller,
the planner's estimates are in fact more accurate.
Tomas Vondra, reviewed by Robert Haas.
This is advantageous first because it allows us to hash the smaller table
regardless of the outer-join type, and second because hash join can be more
flexible than merge join in dealing with arbitrary join quals in a FULL
join. For merge join all the join quals have to be mergejoinable, but hash
join will work so long as there's at least one hashjoinable qual --- the
others can be any condition. (This is true essentially because we don't
keep per-inner-tuple match flags in merge join, while hash join can do so.)
To do this, we need a has-it-been-matched flag for each tuple in the
hashtable, not just one for the current outer tuple. The key idea that
makes this practical is that we can store the match flag in the tuple's
infomask, since there are lots of bits there that are of no interest for a
MinimalTuple. So we aren't increasing the size of the hashtable at all for
the feature.
To write this without turning the hash code into even more of a pile of
spaghetti than it already was, I rewrote ExecHashJoin in a state-machine
style, similar to ExecMergeJoin. Other than that decision, it was pretty
straightforward.
We show the number of buckets, the number of batches (and also the original
number if it has changed), and the peak space used by the hash table. Minor
executor changes to track peak space used.
distribution, by creating a special fast path for the (first few) most common
values of the outer relation. Tuples having hashvalues matching the MCVs
are effectively forced to be in the first batch, so that we never write
them out to the batch temp files.
Bryce Cutt and Ramon Lawrence, with some editorialization by me.
for each temp file, rather than once per sort or hashjoin; this allows
spreading the data of a large sort or join across multiple tablespaces.
(I remain dubious that this will make any difference in practice, but certain
people insisted.) Arrange to cache the results of parsing the GUC variable
instead of recomputing from scratch on every demand, and push usage of the
cache down to the bottommost fd.c level.
tablespace(s) in which to store temp tables and temporary files. This is a
list to allow spreading the load across multiple tablespaces (a random list
element is chosen each time a temp object is to be created). Temp files are
not stored in per-database pgsql_tmp/ directories anymore, but per-tablespace
directories.
Jaime Casanova and Albert Cervera, with review by Bernd Helmle and Tom Lane.
selecting power-of-2, rather than prime, numbers of buckets in hash joins.
If the hash functions are doing their jobs properly by making all hash bits
equally random, this is good enough, and it saves expensive integer division
and modulus operations.
Hashing for aggregation purposes still needs work, so it's not time to
mark any cross-type operators as hashable for general use, but these cases
work if the operators are so marked by hand in the system catalogs.
match because they contain a null join key (and the join operator is
known strict). Improves performance significantly when the inner
relation contains a lot of nulls, as per bug #2930.
Thomas Munro
Bruce Momjian
Peter Geoghegan
Peter Eisentraut
Tom Lane
Andres Freund
Kevin Grittner
Heikki Linnakangas
Magnus Hagander
Robert Haas