Hash index depends on estimation of numbers of tuples and pages of relations,
incorrect value could be a reason of significantly growing of index. Vacuum
full recreates heap and reindex all indexes before renewal stats. The patch
fixes that, so indexes will see correct values.
Backpatch to v10 only because earlier versions haven't usable hash index and
growing of hash index is a single user-visible symptom.
Author: Amit Kapila
Reviewed-by: Ashutosh Sharma, me
Discussion: https://www.postgresql.org/message-id/flat/20171115232922.5tomkxnw3iq6jsg7@inml.weebeastie.net
Polygon opclass uses compress method feature of SP-GiST added earlier. For now
it's a single operator class which uses this feature. SP-GiST actually indexes
a bounding boxes of input polygons, so part of supported operations are lossy.
Opclass uses most methods of corresponding opclass over boxes of SP-GiST and
treats bounding boxes as point in 4D-space.
Bump catalog version.
Authors: Nikita Glukhov, Alexander Korotkov with minor editorization by me
Reviewed-By: all authors + Darafei Praliaskouski
Discussion: https://www.postgresql.org/message-id/flat/54907069.1030506@sigaev.ru
Since 9.4, we've allowed the syntax "select union select" and variants
of that. However, the planner wasn't expecting a no-column set operation
and ended up treating the set operation as if it were UNION ALL.
Turns out it's trivial to fix in v10 and later; we just need to be careful
about not generating a Sort node with no sort keys. However, since a weird
corner case like this is never going to be exercised by developers, we'd
better have thorough regression tests if we want to consider it supported.
Per report from Victor Yegorov.
Discussion: https://postgr.es/m/CAGnEbojGJrRSOgJwNGM7JSJZpVAf8xXcVPbVrGdhbVEHZ-BUMw@mail.gmail.com
currentCommandIdUsed is only used to skip redundant increments of the
command counter, and CommandCounterIncrement() is categorically denied
under parallelism anyway. Therefore, it's OK for
GetCurrentCommandId() to mark the counter value used, as long as it
happens in the leader, not a worker.
Prior to commit e9baa5e9fa, the slightly
incorrect check didn't matter, but now it does. A test case added by
commit 1804284042 uncovered the problem
by accident; it caused failures with force_parallel_mode=on/regress.
Report and review by Andres Freund. Patch by me.
Discussion: http://postgr.es/m/20171221143106.5lhtygohvmazli3x@alap3.anarazel.de
This patch does three interrelated things:
* Create a new expression execution step type EEOP_PARAM_CALLBACK
and add the infrastructure needed for add-on modules to generate that.
As discussed, the best control mechanism for that seems to be to add
another hook function to ParamListInfo, which will be called by
ExecInitExpr if it's supplied and a PARAM_EXTERN Param is found.
For stand-alone expressions, we add a new entry point to allow the
ParamListInfo to be specified directly, since it can't be retrieved
from the parent plan node's EState.
* Redesign the API for the ParamListInfo paramFetch hook so that the
ParamExternData array can be entirely virtual. This also lets us get rid
of ParamListInfo.paramMask, instead leaving it to the paramFetch hook to
decide which param IDs should be accessible or not. plpgsql_param_fetch
was already doing the identical masking check, so having callers do it too
seemed redundant. While I was at it, I added a "speculative" flag to
paramFetch that the planner can specify as TRUE to avoid unwanted failures.
This solves an ancient problem for plpgsql that it couldn't provide values
of non-DTYPE_VAR variables to the planner for fear of triggering premature
"record not assigned yet" or "field not found" errors during planning.
* Rework plpgsql to get rid of the need for "unshared" parameter lists,
by dint of turning the single ParamListInfo per estate into a nearly
read-only data structure that doesn't instantiate any per-variable data.
Instead, the paramFetch hook controls access to per-variable data and can
make the right decisions on the fly, replacing the cases that we used to
need multiple ParamListInfos for. This might perhaps have been a
performance loss on its own, but by using a paramCompile hook we can
bypass plpgsql_param_fetch entirely during normal query execution.
(It's now only called when, eg, we copy the ParamListInfo into a cursor
portal. copyParamList() or SerializeParamList() effectively instantiate
the virtual parameter array as a simple physical array without a
paramFetch hook, which is what we want in those cases.) This allows
reverting most of commit 6c82d8d1f, though I kept the cosmetic
code-consolidation aspects of that (eg the assign_simple_var function).
Performance testing shows this to be at worst a break-even change,
and it can provide wins ranging up to 20% in test cases involving
accesses to fields of "record" variables. The fact that values of
such variables can now be exposed to the planner might produce wins
in some situations, too, but I've not pursued that angle.
In passing, remove the "parent" pointer from the arguments to
ExecInitExprRec and related functions, instead storing that pointer in a
transient field in ExprState. The ParamListInfo pointer for a stand-alone
expression is handled the same way; we'd otherwise have had to add
yet another recursively-passed-down argument in expression compilation.
Discussion: https://postgr.es/m/32589.1513706441@sss.pgh.pa.us
That function currently exists to avoid leaking memory in
CacheMemoryContext in case of trouble while the partition key is being
built, but there's a better way: allocate everything in a memcxt that
goes away if the current (sub)transaction fails, and once the partition
key is built and no further errors can occur, make the memcxt permanent
by making it a child of CacheMemoryContext.
Reviewed-by: Tom Lane
Discussion: https://postgr.es/m/20171027172730.eh2domlkpn4ja62m@alvherre.pgsql
Various Perl scripts we use to generate files were in the habit of
printing things like "generated by $0" into their output files.
That looks like a fine idea at first glance, but it results in
non-reproducible output, because in VPATH builds $0 won't be just
the name of the script file, but a full path for it. We'd prefer
that you get identical results whether using VPATH or not, so this
is a bad thing.
Some of these places also printed their input file name(s), causing
an additional hazard of the same type.
Hence, establish a policy that thou shalt not print $0, nor input file
pathnames, into output files (they're still allowed in error messages,
though). Instead just write the script name verbatim. While we are at
it, we can make these annotations more useful by giving the script's
full relative path name within the PG source tree, eg instead of
Gen_fmgrtab.pl let's print src/backend/utils/Gen_fmgrtab.pl.
Not all of the changes made here actually affect any files shipped
in finished tarballs today, but it seems best to apply the policy
everyplace so that nobody copies unsafe code into places where it
could matter.
Christoph Berg and Tom Lane
Discussion: https://postgr.es/m/20171215102223.GB31812@msg.df7cb.de
Generally, error recovery paths that need to do things like
LWLockReleaseAll and pgstat_report_wait_end also need to call
ConditionVariableCancelSleep, but AbortSubTransaction was missed.
Since subtransaction abort might destroy up the DSM segment that
contains the ConditionVariable stored in cv_sleep_target, this
can result in a crash for anything using condition variables.
Reported and diagnosed by Andres Freund.
Discussion: http://postgr.es/m/20171221110048.rxk6464azzl5t2fi@alap3.anarazel.de
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
When a Gather or Gather Merge node is started and stopped multiple
times, accumulate instrumentation data only once, at the end, instead
of after each execution, to avoid recording inflated totals.
Commit 778e78ae9f, the previous attempt
at a fix, instead reset the state after every execution, which worked
for the general instrumentation data but had problems for the additional
instrumentation specific to Sort and Hash nodes.
Report by hubert depesz lubaczewski. Analysis and fix by Amit Kapila,
following a design proposal from Thomas Munro, with a comment tweak
by me.
Discussion: http://postgr.es/m/20171127175631.GA405@depesz.com
Commit dc02c7bca4 changed this call
to create_sort_path() to take -1 rather than limit_tuples because,
at that time, there was no way for a Sort beneath a Gather Merge
to become a top-N sort.
Later, commit 3452dc5240 provided
a way for a Sort beneath a Gather Merge to become a top-N sort,
but failed to revert the previous commit in the process. Do that.
Report and analysis by Jeff Janes; patch by Thomas Munro; review by
Amit Kapila and by me.
Discussion: http://postgr.es/m/CAEepm=1BWtC34vUroA0Uqjw02MaqdUrW+d6WD85_k8SLyPiKHQ@mail.gmail.com
SharedTuplestore allows multiple participants to write into it and
then read the tuples back from it in parallel. Each reader receives
partial results.
For now it always uses disk files, but other buffering policies and
other kinds of scans (ie each reader receives complete results) may be
useful in future.
The upcoming parallel hash join feature will use this facility.
Author: Thomas Munro
Reviewed-By: Peter Geoghegan, Andres Freund, Robert Haas
Discussion: https://postgr.es/m/CAEepm=2W=cOkiZxcg6qiFQP-dHUe09aqTrEMM7yJDrHMhDv_RA@mail.gmail.com
Mechanism names for SCRAM and channel binding names have been included
in scram.h by the libpq frontend code, and this header references a set
of routines which are only used by the backend. scram-common.h is on
the contrary usable by both the backend and libpq, so getting those
names from there seems more reasonable.
Author: Michael Paquier <michael.paquier@gmail.com>
Previously an assertion failure occurred when pg_stop_backup() for
non-exclusive backup was aborted while it's waiting for WAL files to
be archived. This assertion failure happened in do_pg_abort_backup()
which was called when a non-exclusive backup was canceled.
do_pg_abort_backup() assumes that there is at least one non-exclusive
backup running when it's called. But pg_stop_backup() can be canceled
even after it marks the end of non-exclusive backup (e.g.,
during waiting for WAL archiving). This broke the assumption that
do_pg_abort_backup() relies on, and which caused an assertion failure.
This commit changes do_pg_abort_backup() so that it does nothing
when non-exclusive backup has been already marked as completed.
That is, the asssumption is also changed, and do_pg_abort_backup()
now can handle even the case where it's called when there is
no running backup.
Backpatch to 9.6 where SQL-callable non-exclusive backup was added.
Author: Masahiko Sawada and Michael Paquier
Reviewed-By: Robert Haas and Fujii Masao
Discussion: https://www.postgresql.org/message-id/CAD21AoD2L1Fu2c==gnVASMyFAAaq3y-AQ2uEVj-zTCGFFjvmDg@mail.gmail.com
es_query_dsa turns out to be broken by design, because it supposes
that there is only one DSA for the whole query, whereas there is
actually one per Gather (Merge) node. For now, work around that
problem by setting and clearing the pointer around the sections of
code that might need it. It's probably a better idea to get rid of
es_query_dsa altogether in favor of having each node keep track
individually of which DSA is relevant, but that seems like more than
we would want to back-patch.
Thomas Munro, reviewed and tested by Andreas Seltenreich, Amit
Kapila, and by me.
Discussion: http://postgr.es/m/CAEepm=1U6as=brnVvMNixEV2tpi8NuyQoTmO8Qef0-VV+=7MDA@mail.gmail.com
Compilers that don't know that ereport(ERROR) doesn't return
complained about the new coding in scanint8() introduced by
commit 101c7ee3e. Tweak coding to avoid the warning.
Per buildfarm.
The previous commit has shown that the sanity checks around freezing
aren't strong enough. Strengthening them seems especially important
because the existance of the bug has caused corruption that we don't
want to make even worse during future vacuum cycles.
The errors are emitted with ereport rather than elog, despite being
"should never happen" messages, so a proper error code is emitted. To
avoid superflous translations, mark messages as internal.
Author: Andres Freund and Alvaro Herrera
Reviewed-By: Alvaro Herrera, Michael Paquier
Discussion: https://postgr.es/m/20171102112019.33wb7g5wp4zpjelu@alap3.anarazel.de
Backpatch: 9.3-
Previously it was possible that a tuple was not pruned during vacuum,
even though its update xmax (i.e. the updating xid in a multixact with
both key share lockers and an updater) was below the cutoff horizon.
As the freezing code assumed, rightly so, that that's not supposed to
happen, xmax would be preserved (as a member of a new multixact or
xmax directly). That causes two problems: For one the tuple is below
the xmin horizon, which can cause problems if the clog is truncated or
once there's an xid wraparound. The bigger problem is that that will
break HOT chains, which in turn can lead two to breakages: First,
failing index lookups, which in turn can e.g lead to constraints being
violated. Second, future hot prunes / vacuums can end up making
invisible tuples visible again. There's other harmful scenarios.
Fix the problem by recognizing that tuples can be DEAD instead of
RECENTLY_DEAD, even if the multixactid has alive members, if the
update_xid is below the xmin horizon. That's safe because newer
versions of the tuple will contain the locking xids.
A followup commit will harden the code somewhat against future similar
bugs and already corrupted data.
Author: Andres Freund, with changes by Alvaro Herrera
Reported-By: Daniel Wood
Analyzed-By: Andres Freund, Alvaro Herrera, Robert Haas, Peter
Geoghegan, Daniel Wood, Yi Wen Wong, Michael Paquier
Reviewed-By: Alvaro Herrera, Robert Haas, Michael Paquier
Discussion:
https://postgr.es/m/E5711E62-8FDF-4DCA-A888-C200BF6B5742@amazon.comhttps://postgr.es/m/20171102112019.33wb7g5wp4zpjelu@alap3.anarazel.de
Backpatch: 9.3-
The logical slots have a fast code path for sending data so as not to
impose too high a per message overhead. The fast path skips checks for
interrupts and timeouts. However, the existing coding failed to consider
the fact that a transaction with a large number of changes may take a
very long time to be processed and sent to the client. This causes the
walsender to ignore interrupts for potentially a long time and more
importantly it will result in the walsender being killed due to
timeout at the end of such a transaction.
This commit changes the fast path to also check for interrupts and only
allows calling the fast path when the last keepalive check happened less
than half the walsender timeout ago. Otherwise the slower code path will
be taken.
Backpatched to 9.4
Petr Jelinek, reviewed by Kyotaro HORIGUCHI, Yura Sokolov, Craig
Ringer and Robert Haas.
Discussion: https://postgr.es/m/e082a56a-fd95-a250-3bae-0fff93832510@2ndquadrant.com
In order for executor nodes to be able to change their ExecProcNode function
after ExecInitNode() has finished, provide ExecSetExecProcNode(). This allows
any wrappers functions that only execProcnode.c knows about to be reinstalled.
The motivation for wanting to change ExecProcNode after ExecInitNode() has
finished is that it is not known until later whether parallel query is
available, so if a parallel variant is to be installed then ExecInitNode()
is too soon to decide.
Author: Thomas Munro
Reviewed-By: Andres Freund
Discussion: https://postgr.es/m/CAEepm=09rr65VN+cAV5FgyM_z=D77Xy8Fuc9CDDDYbq3pQUezg@mail.gmail.com
Crash restarts currently don't clean up temporary files, as a debugging aid.
If a left-over file happens to have the same name as a segment file we're
trying to create, we'll just truncate and reuse it, but there is a problem:
BufFileOpenShared() determines how many segment files exist by trying to open
.0, .1, .2, ... until it finds no more files. It might be confused by a junk
file that has the next segment number. To defend against that, make sure we
always create a gap after the end file by unlinking the following name if it
exists. Also make it an error to try to open a BufFile that doesn't exist
(has no segment 0), so as not to encourage the development of client code
that depends on an interface that we can't reliably provide.
Author: Thomas Munro
Reviewed-By: Andres Freund
Discussion: https://postgr.es/m/CAEepm%3D2jhCbC_GFQJaaDhWxLB4EXtT3vVd5czuRNaqF5CWSTog%40mail.gmail.com
This patch makes a number of interrelated changes to reduce the overhead
involved in creating/deleting memory contexts. The key ideas are:
* Include the AllocSetContext header of an aset.c context in its first
malloc request, rather than allocating it separately in TopMemoryContext.
This means that we now always create an initial or "keeper" block in an
aset, even if it never receives any allocation requests.
* Create freelists in which we can save and recycle recently-destroyed
asets (this idea is due to Robert Haas).
* In the common case where the name of a context is a constant string,
just store a pointer to it in the context header, rather than copying
the string.
The first change eliminates a palloc/pfree cycle per context, and
also avoids bloat in TopMemoryContext, at the price that creating
a context now involves a malloc/free cycle even if the context never
receives any allocations. That would be a loser for some common
usage patterns, but recycling short-lived contexts via the freelist
eliminates that pain.
Avoiding copying constant strings not only saves strlen() and strcpy()
overhead, but is an essential part of the freelist optimization because
it makes the context header size constant. Currently we make no
attempt to use the freelist for contexts with non-constant names.
(Perhaps someday we'll need to think harder about that, but in current
usage, most contexts with custom names are long-lived anyway.)
The freelist management in this initial commit is pretty simplistic,
and we might want to refine it later --- but in common workloads that
will never matter because the freelists will never get full anyway.
To create a context with a non-constant name, one is now required to
call AllocSetContextCreateExtended and specify the MEMCONTEXT_COPY_NAME
option. AllocSetContextCreate becomes a wrapper macro, and it includes
a test that will complain about non-string-literal context name
parameters on gcc and similar compilers.
An unfortunate side effect of making AllocSetContextCreate a macro is
that one is now *required* to use the size parameter abstraction macros
(ALLOCSET_DEFAULT_SIZES and friends) with it; the pre-9.6 habit of
writing out individual size parameters no longer works unless you
switch to AllocSetContextCreateExtended.
Internally to the memory-context-related modules, the context creation
APIs are simplified, removing the rather baroque original design whereby
a context-type module called mcxt.c which then called back into the
context-type module. That saved a bit of code duplication, but not much,
and it prevented context-type modules from exercising control over the
allocation of context headers.
In passing, I converted the test-and-elog validation of aset size
parameters into Asserts to save a few more cycles. The original thought
was that callers might compute size parameters on the fly, but in practice
nobody does that, so it's useless to expend cycles on checking those
numbers in production builds.
Also, mark the memory context method-pointer structs "const",
just for cleanliness.
Discussion: https://postgr.es/m/2264.1512870796@sss.pgh.pa.us
port.h redirects isnan() to _isnan() on windows, which in turn is
provided by float.h rather than math.h. Therefore include the latter
as well.
Per buildfarm.
A previous commit added inline functions that provide fast(er) and
correct overflow checks for signed integer math. Use them in a
significant portion of backend code. There's more to touch in both
backend and frontend code, but these were the easily identifiable
cases.
The old overflow checks are noticeable in integer heavy workloads.
A secondary benefit is that getting rid of overflow checks that rely
on signed integer overflow wrapping around, will allow us to get rid
of -fwrapv in the future. Which in turn slows down other code.
Author: Andres Freund
Discussion: https://postgr.es/m/20171024103954.ztmatprlglz3rwke@alap3.anarazel.de
Commit 8b304b8b72 removed replacement
selection, but left behind this comment text. The optimization to
which the comment refers is not relevant without replacement
selection, because if we had so few tuples as to require only one
tape, we would have just completed the sort in memory.
Peter Geoghegan
Discussion: http://postgr.es/m/CAH2-WznqupLA8CMjp+vqzoe0yXu0DYYbQSNZxmgN76tLnAOZ_w@mail.gmail.com
I noticed that _SPI_execute_plan initially sets spierrcontext.arg = NULL,
and only fills it in some time later. If an error were to happen in
between, _SPI_error_callback would try to dereference the null pointer.
This is unlikely --- there's not much between those points except
push-snapshot calls --- but it's clearly not impossible. Tweak the
callback to do nothing if the pointer isn't set yet.
It's been like this for awhile, so back-patch to all supported branches.
Those cases currently crash and supporting them is more work then
originally thought, so we'll just prohibit these scenarios for now.
Author: Michael Paquier <michael.paquier@gmail.com>
Reviewed-by: Amit Langote <Langote_Amit_f8@lab.ntt.co.jp>
Reported-by: Мансур Галиев <gomer94@yandex.ru>
Bug: #14866
A COPY into a table should apply identity sequence values just like it
does for ordinary defaults. This was previously forgotten, leading to
null values being inserted, which in turn would fail because identity
columns have not-null constraints.
Author: Michael Paquier <michael.paquier@gmail.com>
Reported-by: Steven Winfield <steven.winfield@cantabcapital.com>
Bug: #14952
When a worker is flagged as BGW_NEVER_RESTART and we fail to start it,
or if it is not marked BGW_NEVER_RESTART but is terminated before
startup succeeds, what BgwHandleStatus should be reported? The
previous code really hadn't considered this possibility (as indicated
by the comments which ignore it completely) and would typically return
BGWH_NOT_YET_STARTED, but that's not a good answer, because then
there's no way for code using GetBackgroundWorkerPid() to tell the
difference between a worker that has not started but will start
later and a worker that has not started and will never be started.
So, when this case happens, return BGWH_STOPPED instead. Update the
comments to reflect this.
The preceding fix by itself is insufficient to fix the problem,
because the old code also didn't send a notification to the process
identified in bgw_notify_pid when startup failed. That might've
been technically correct under the theory that the status of the
worker was BGWH_NOT_YET_STARTED, because the status would indeed not
change when the worker failed to start, but now that we're more
usefully reporting BGWH_STOPPED, a notification is needed.
Without these fixes, code which starts background workers and then
uses the recommended APIs to wait for those background workers to
start would hang indefinitely if the postmaster failed to fork a
worker.
Amit Kapila and Robert Haas
Discussion: http://postgr.es/m/CAA4eK1KDfKkvrjxsKJi3WPyceVi3dH1VCkbTJji2fuwKuB=3uw@mail.gmail.com
When we create an Append node, we can spread out the workers over the
subplans instead of piling on to each subplan one at a time, which
should typically be a bit more efficient, both because the startup
cost of any plan executed entirely by one worker is paid only once and
also because of reduced contention. We can also construct Append
plans using a mix of partial and non-partial subplans, which may allow
for parallelism in places that otherwise couldn't support it.
Unfortunately, this patch doesn't handle the important case of
parallelizing UNION ALL by running each branch in a separate worker;
the executor infrastructure is added here, but more planner work is
needed.
Amit Khandekar, Robert Haas, Amul Sul, reviewed and tested by
Ashutosh Bapat, Amit Langote, Rafia Sabih, Amit Kapila, and
Rajkumar Raghuwanshi.
Discussion: http://postgr.es/m/CAJ3gD9dy0K_E8r727heqXoBmWZ83HwLFwdcaSSmBQ1+S+vRuUQ@mail.gmail.com
When a Gather or Gather Merge node is started and stopped multiple
times, the old code wouldn't reset the shared state between executions,
potentially resulting in dramatically inflated instrumentation data
for nodes beneath it. (The per-worker instrumentation ended up OK,
I think, but the overall totals were inflated.)
Report by hubert depesz lubaczewski. Analysis and fix by Amit Kapila,
reviewed and tweaked a bit by me.
Discussion: http://postgr.es/m/20171127175631.GA405@depesz.com
If a hash join appears in a parallel query, there may be no hash table
available for explain.c to inspect even though a hash table may have
been built in other processes. This could happen either because
parallel_leader_participation was set to off or because the leader
happened to hit the end of the outer relation immediately (even though
the complete relation is not empty) and decided not to build the hash
table.
Commit bf11e7ee introduced a way for workers to exchange
instrumentation via the DSM segment for Sort nodes even though they
are not parallel-aware. This commit does the same for Hash nodes, so
that explain.c has a way to find instrumentation data from an
arbitrary participant that actually built the hash table.
Author: Thomas Munro
Reviewed-By: Andres Freund
Discussion: https://postgr.es/m/CAEepm%3D3DUQC2-z252N55eOcZBer6DPdM%3DFzrxH9dZc5vYLsjaA%40mail.gmail.com
Alvaro Herrera