Until now, when DROP DATABASE got interrupted in the wrong moment, the removal
of the pg_database row would also roll back, even though some irreversible
steps have already been taken. E.g. DropDatabaseBuffers() might have thrown
out dirty buffers, or files could have been unlinked. But we continued to
allow connections to such a corrupted database.
To fix this, mark databases invalid with an in-place update, just before
starting to perform irreversible steps. As we can't add a new column in the
back branches, we use pg_database.datconnlimit = -2 for this purpose.
An invalid database cannot be connected to anymore, but can still be
dropped.
Unfortunately we can't easily add output to psql's \l to indicate that some
database is invalid, it doesn't fit in any of the existing columns.
Add tests verifying that a interrupted DROP DATABASE is handled correctly in
the backend and in various tools.
Reported-by: Evgeny Morozov <postgresql3@realityexists.net>
Author: Andres Freund <andres@anarazel.de>
Reviewed-by: Daniel Gustafsson <daniel@yesql.se>
Reviewed-by: Thomas Munro <thomas.munro@gmail.com>
Discussion: https://postgr.es/m/20230509004637.cgvmfwrbht7xm7p6@awork3.anarazel.de
Discussion: https://postgr.es/m/20230314174521.74jl6ffqsee5mtug@awork3.anarazel.de
Backpatch: 11-, bug present in all supported versions
When vac_truncate_clog() encounters bogus datfrozenxid / datminmxid values, it
returns early. Unfortunately, until now, it did not release
WrapLimitsVacuumLock. If the backend later tries to acquire
WrapLimitsVacuumLock, the session / autovacuum worker hangs in an
uncancellable way. Similarly, other sessions will hang waiting for the
lock. However, if the backend holding the lock exited or errored out for some
reason, the lock was released.
The bug was introduced as a side effect of 566372b3d6.
It is interesting that there are no production reports of this problem. That
is likely due to a mix of bugs leading to bogus values having gotten less
common, process exit releasing locks and instances of hangs being hard to
debug for "normal" users.
Discussion: https://postgr.es/m/20230621221208.vhsqgduwfpzwxnpg@awork3.anarazel.de
GiST compress functions (like all GiST opclass functions) are
supposed to be called in short-lived memory contexts, so that
minor memory leaks in them are not of concern, and indeed
explicit pfree's are likely slightly counterproductive.
But this one in g_intbig_compress() is more than
slightly counterproductive, because it's guarded by
"if (in != DatumGetArrayTypeP(entry->key))" which means
that if this test succeeds, we've detoasted the datum twice.
(And to add insult to injury, the extra detoast result is
leaked.) Let's just drop the whole stanza, relying on the
GiST temporary context mechanism to clean up in good time.
The analogous bit in g_int_compress() is
if (r != (ArrayType *) DatumGetPointer(entry->key))
pfree(r);
which doesn't have the gratuitous-detoast problem so
I left it alone. Perhaps there is a case for removing
unnecessary pfree's more widely, but I'm not sure if it's
worth the code churn.
The potential extra decompress seems expensive enough to
justify calling this a (minor) performance bug and
back-patching.
Konstantin Knizhnik, Matthias van de Meent, Tom Lane
Discussion: https://postgr.es/m/CAEze2Wi86=DxErfvf+SCB2UKmU2amKOF60BKuJOX=w-RojRn0A@mail.gmail.com
Commit 89e46da5e allowed REPLICA IDENTITY FULL tables to use an index
on the subscriber during apply of update/delete. This commit clarifies
in the documentation that the leftmost field of candidate indexes must
be a column (not an expression) that references the published relation
column.
The source code comments are also updated accordingly.
Reviewed-by: Peter Smith, Amit Kapila
Discussion: https://postgr.es/m/CAD21AoDJjffEvUFKXT27Q5U8-UU9JHv4rrJ9Ke8Zkc5UPWHLvA@mail.gmail.com
Backpatch-through: 16
In the event the llvm option is defined to be 'auto', it is possible
that the host machine might not have a C++ compiler. If that is the
case, then we shouldn't continue reaching for the llvm dependency.
To make it easier to understand the case where LLVM support is disabled due to
lacking a C++ compiler, add a message noting that fact.
Author: Tristan Partin <tristan@neon.tech>
Reviewed-by: Andres Freund <andres@anarazel.de>
Discussion: https://postgr.es/m/CSPIJVUDZFKX.3KHMOAVGF94RV@c3po
Backpatch: 16-, where meson support was added
That was already done in a lot of places, but not all.
Backpatch this to keep the the meson files aligned as long as reasonably
possible.
Author: Tristan Partin <tristan@neon.tech>
Reviewed-by: Andres Freund <andres@anarazel.de>
Discussion: https://postgr.es/m/CSPIJVUDZFKX.3KHMOAVGF94RV@c3po
Backpatch: 16-, where meson support was added
Since PostgresMain calls sigsetjmp, any local variables that are not
marked "volatile" have a risk of unspecified behavior. In practice
this means that when control returns via longjmp, such variables might
get reset to their values as of the time of sigsetjmp, depending on
whether the compiler chose to put them in registers or on the stack.
We were careful about this for "send_ready_for_query", but not the
other local variables.
In the case of the timeout_enabled flags, resetting them to
their initial "false" states is actually good, since we do
"disable_all_timeouts()" in the longjmp cleanup code path. If that
does not happen, we risk uselessly calling "disable_timeout()" later,
which is harmless but a little bit expensive. Let's explicitly reset
these flags so that the behavior is correct and platform-independent.
(This change means that we really don't need the new "volatile"
markings after all, but let's install them anyway since any change
in this logic could re-introduce a problem.)
There is no issue for "firstchar" and "input_message" because those
are explicitly reinitialized each time through the query processing
loop. To make that clearer, move them to be declared inside the loop.
That leaves us with all the function-lifespan locals except the
sigjmp_buf itself marked as volatile, which seems like a good policy
to have going forward.
Because of the possibility of extra disable_timeout() calls, this
seems worth back-patching.
Sergey Shinderuk and Tom Lane
Discussion: https://postgr.es/m/2eda015b-7dff-47fd-d5e2-f1a9899b90a6@postgrespro.ru
As coded, the code would use as a base comparison the namespace OID from
the first object scanned in pg_depend when switching its namespace
dependency entry to the new one, and use it as a base of comparison for
any follow-up checks. It would also be used as the old namespace OID to
switch *from* for the extension's pg_depend entry. Hence, if the first
object scanned has a namespace different than the one stored in the
extension, we would finish by:
- Not checking that the extension objects map with the extension's
schema.
- Not switching the extension -> namespace dependency entry to the new
namespace provided by the user, making ALTER EXTENSION ineffective.
This issue exists since this command has been introduced in d9572c4 for
relocatable extension, so backpatch all the way down to 11. The test
case has been provided by Heikki, that I have tweaked a bit to show the
effects on pg_depend for the extension.
Reported-by: Heikki Linnakangas
Author: Michael Paquier, Heikki Linnakangas
Discussion: https://postgr.es/m/20eea594-a05b-4c31-491b-007b6fceef28@iki.fi
Backpatch-through: 11
The DEFAULT parameter seems most similar to the NULL parameter, so
move it next to it, instead of having it at the end of the parameter
list because it was the last one added.
Commit e213de8e78 fixed a problem with path lengths to a tempdir on
Windows, but caused problems on at least some Unix systems where the
system tempdir is on a different file system. To work around this, only
used the system temdir for the destination of pg_replslot on Windows,
and otherwise restore the old behaviour.
Backpatch to relase 14 like the previous patch.
Problem exposed by a myriad of buildfarm animals.
The symlink to a longer location tripped up some Windows limit on
buildfarm animal fairywren when running with meson, which uses slightly
longer paths.
Backpatch to all live branches to keep the script in sync.
Commit 19d8e2308b allowed a weaker check for HOT with summarizing
indexes, but it did not update README.HOT. So do that now.
Patch by Matthias van de Meent, minor changes by me. Backpatch to 16,
where the optimization was introduced.
Author: Matthias van de Meent
Reviewed-by: Tomas Vondra
Backpatch-through: 16
Discussion: https://postgr.es/m/CAEze2WiEOm8V+c9kUeYp2BPhbEc5s473fUf51xNeqvSFGv44Ew@mail.gmail.com
On Windows, it's sometimes difficult to create a file with a path longer
than 255 chars, and if it can be created it might not be seen by the
archiver. This can be triggered by the test for tar backups with
filenames greater than 100 bytes. So we skip that test if the path would
exceed 255.
Backpatch to all live branches.
Reviewed by Daniel Gustafsson
Discussion: https://postgr.es/m/666ac55b-3400-fb2c-2cea-0281bf36a53c@dunslane.net
We create a file, so we better WAL-log it. In practice, all the
built-in index AMs and all extensions that I'm aware of write a
metapage to the init fork, which is WAL-logged, and replay of the
metapage implicitly creates the fork too. But if ambuildempty() didn't
write any page, we would miss it.
This can be seen with dummy_index_am. Set up replication, create a
'dummy_index_am' index on an unlogged table, and look at the files
created in the replica: the init fork is not created on the
replica. Dummy_index_am doesn't do anything with the relation files,
however, so it doesn't lead to any user-visible errors.
Backpatch to all supported versions.
Reviewed-by: Robert Haas
Discussion: https://www.postgresql.org/message-id/6e5bbc08-cdfc-b2b3-9e23-1a914b9850a9%40iki.fi
This commit reverts the work done by commits 3ba59ccc89 and 72e78d831a.
Those commits were incorrect in asserting that we never acquire any other
heavy-weight lock after acquring page lock other than relation extension
lock. We can acquire a lock on catalogs while doing catalog look up after
acquring page lock.
This won't impact any existing feature but we need to think some other way
to achieve this before parallelizing other write operations or even
improving the parallelism in vacuum (like allowing multiple workers
for an index).
Reported-by: Jaime Casanova
Author: Amit Kapila
Backpatch-through: 13
Discussion: https://postgr.es/m/CAJKUy5jffnRKNvRHKQ0LynRb0RJC-o4P8Ku3x9vGAVLwDBWumQ@mail.gmail.com
llvm_release_context() called llvm_enter_fatal_on_oom(), but was missing
the corresponding llvm_leave_fatal_on_oom() call. As a result, if JIT was
used at all, we were almost always in the "fatal-on-oom" state.
It only makes a difference if you use an extension written in C++, and
run out of memory in a C++ 'new' call. In that case, you would get a
PostgreSQL FATAL error, instead of the default behavior of throwing a
C++ exception.
Back-patch to all supported versions.
Reviewed-by: Daniel Gustafsson
Discussion: https://www.postgresql.org/message-id/54b78cca-bc84-dad8-4a7e-5b56f764fab5@iki.fi
Commit 7b64e4b3 taught DropSubscription() to drop stats entry of
subscription that is not associated with a replication slot for apply
worker at DROP SUBSCRIPTION but missed covering the case where the
subscription is not associated with replication slots for both apply
worker and tablesync worker.
Also add a test to verify that the stats for slot-less subscription is
removed at DROP SUBSCRIPTION time.
Backpatch down to 15.
Author: Masahiko Sawada
Reviewed-by: Nathan Bossart, Hayato Kuroda, Melih Mutlu, Amit Kapila
Discussion: https://postgr.es/m/CAD21AoB71zkP7uPT7JDPsZcvp0749ExEQnOJxeNKPDFisHar+w@mail.gmail.com
Backpatch-through: 15
Clear any potential stale next_phase_at value from the snapshot
builder which otherwise may trip an assertion check ensuring
that there is no next_phase_at value.
This can be reproduced by running 80 concurrent sessions like
the below where $c is a loop counter (assumes there has been
1..$c databases created) :
echo "
CREATE TABLE replication_example(id SERIAL PRIMARY KEY,
somedata int,
text varchar(120));
SELECT 'init' FROM
pg_create_logical_replication_slot('regression_slot_$c',
'test_decoding');
SELECT data FROM
pg_logical_slot_get_changes('regression_slot_$c', NULL,
NULL, 'include-xids', '0',
'skip-empty-xacts', '1');
" | psql -d regress_$c >>psql.log &
Backpatch down to v16.
Bug: #17695
Author: Masahiko Sawada <sawada.mshk@gmail.com>
Reviewed-by: Alexander Lakhin <exclusion@gmail.com>
Reported-by: bowenshi <zxwsbg@qq.com>
Discussion: https://postgr.es/m/17695-6be9277c9295985f@postgresql.org
Backpatch-through: v16
If you create a table and don't insert any data into it, the relation file
is never fsync'd. You don't lose data, because an empty table doesn't have
any data to begin with, but if you crash and lose the file, subsequent
operations on the table will fail with "could not open file" error.
To fix, register an fsync request in mdcreate(), like we do for mdwrite().
Per discussion, we probably should also fsync the containing directory
after creating a new file. But that's a separate and much wider issue.
Backpatch to all supported versions.
Reviewed-by: Andres Freund, Thomas Munro
Discussion: https://www.postgresql.org/message-id/d47d8122-415e-425c-d0a2-e0160829702d%40iki.fi
It's OK to be lazy about re-binning memory segments when allocating,
because that can only leave segments in a bin that's too high. We'll
search higher bins if necessary while allocating next time, and
also eventually re-bin, so no memory can become unreachable that way.
However, when freeing memory, the largest contiguous range of free pages
might go up, so we should re-bin eagerly to make sure we don't leave the
segment in a bin that is too low for get_best_segment() to find.
The re-binning code is moved into a function of its own, so it can be
called whenever free pages are returned to the segment's free page map.
Back-patch to all supported releases.
Author: Dongming Liu <ldming101@gmail.com>
Reviewed-by: Robert Haas <robertmhaas@gmail.com> (earlier version)
Reviewed-by: Thomas Munro <thomas.munro@gmail.com>
Discussion: https://postgr.es/m/CAL1p7e8LzB2LSeAXo2pXCW4%2BRya9s0sJ3G_ReKOU%3DAjSUWjHWQ%40mail.gmail.com
The ginfast.c code previously checked for conflicts in before locking
the relevant buffer, leaving a window where a RW conflict could be
missed. Re-order.
There was also a place where buffer ID and block number were confused
while trying to predicate-lock a page, noted by visual inspection.
Back-patch to all supported releases. Fixes one more problem discovered
with the reproducer from bug #17949, in this case when Dmitry tried
other index types.
Reported-by: Artem Anisimov <artem.anisimov.255@gmail.com>
Reported-by: Dmitry Dolgov <9erthalion6@gmail.com>
Reviewed-by: Heikki Linnakangas <hlinnaka@iki.fi>
Discussion: https://postgr.es/m/17949-a0f17035294a55e2%40postgresql.org
When performing a bitmap heap scan, we don't want to miss concurrent
writes that occurred after we observed the heap's rs_nblocks, but before
we took predicate locks on index pages. Therefore, we can't skip
fetching any heap tuples that are referenced by the index, because we
need to test them all with CheckForSerializableConflictOut(). The
old optimization that would ignore any references to blocks >=
rs_nblocks gets in the way of that requirement, because it means that
concurrent writes in that window are ignored.
Removing that optimization shouldn't affect correctness at any isolation
level, because any new tuples shouldn't be visible to an MVCC snapshot.
There also shouldn't be any error-causing references to heap blocks past
the end, because we should have held at least an AccessShareLock on the
table before the index scan. It can't get smaller while our transaction
is running. For now, though, we'll keep the optimization at lower
levels to avoid making unnecessary changes in a bug fix.
Back-patch to all supported releases. In release 11, the code is in a
different place but not fundamentally different. Fixes one aspect of
bug #17949.
Reported-by: Artem Anisimov <artem.anisimov.255@gmail.com>
Reviewed-by: Dmitry Dolgov <9erthalion6@gmail.com>
Reviewed-by: Heikki Linnakangas <hlinnaka@iki.fi>
Discussion: https://postgr.es/m/17949-a0f17035294a55e2%40postgresql.org
When predicate-locking btrees, we have a special case for completely
empty btrees, since there is no page to lock. This was racy, because,
without buffer lock held, a matching key could be inserted between the
_bt_search() and the PredicateLockRelation() calls.
Fix, by rechecking _bt_search() after taking the relation-level SIREAD
lock, if using SERIALIZABLE isolation and an empty btree is discovered.
Back-patch to all supported releases. Fixes one aspect of bug #17949.
Reported-by: Artem Anisimov <artem.anisimov.255@gmail.com>
Reviewed-by: Dmitry Dolgov <9erthalion6@gmail.com>
Reviewed-by: Heikki Linnakangas <hlinnaka@iki.fi>
Discussion: https://postgr.es/m/17949-a0f17035294a55e2%40postgresql.org
The test inserted 70k rows into a foreign table, in order to verify
correct behavior with more than 65535 parameters, and was added in
response to a bug report.
However, this is rather expensive, especially when running the tests
under valgrind, CLOBBER_CACHE_ALWAYS etc. It doesn't seem worth it to
keep running the test, so remove it from all branches (14+).
Backpatch-through: 14
Discussion: https://postgr.es/m/2131017.1623451468@sss.pgh.pa.us
Creation of a file with a very long name can create problems on Windows
due to its file path limits. Work around that by creating the file via a
symlink with a shorter name.
Error displayed by buildfarm animal fairywren.o
Backpatch to all live branches
When set, this environment variable was only effective for data
directories but not for all the other temporary files created by
PostgreSQL::Test::Utils. Keeping the temporary files after a successful
run can be useful for debugging purposes.
The documentation is updated to reflect the new behavior, with contents
available in doc/ since v16 and in src/test/perl/README since v15.
Author: Jacob Champion
Reviewed-by: Daniel Gustafsson
Discussion: https://postgr.es/m/CAAWbhmgHtDH1SGZ+Fw05CsXtE0mzTmjbuUxLB9mY9iPKgM6cUw@mail.gmail.com
Discussion: https://postgr.es/m/YyPd9unV14SX2bLF@paquier.xyz
Backpatch-through: 11
Commit dd38ff28ad added a new error message "missing contrecord" when
we fail to reassemble a record. Unfortunately that caused noisy
messages to be logged by pg_waldump at end of segment, and by walsender
when asked to shut down on a segment boundary.
Remove the new error message, so that this condition signals end-of-
WAL without a message. It's arguably a reportable condition that should
not be silenced while performing crash recovery, but fixing that without
introducing noise in the other cases will require more research.
Back-patch to 15.
Reported-by: Tomas Vondra <tomas.vondra@enterprisedb.com>
Discussion: https://postgr.es/m/6a1df56e-4656-b3ce-4b7a-a9cb41df8189%40enterprisedb.com
Commit f24523672d fixed a memory leak by moving the modifiedCols bitmap
into the per-row memory context. In the case of AFTER UPDATE triggers,
the bitmap is however referenced from an event kept until the end of the
query, resulting in a use-after-free bug.
Fixed by copying the bitmap into the AfterTriggerEvents memory context,
which is the one where we keep the trigger events. There's only one
place that needs to do the copy, but the memory context may not exist
yet. Doing that in a separate function seems more readable.
Report by Alexander Pyhalov, fix by me. Backpatch to 13, where the
bitmap was added to the event by commit 71d60e2aa0.
Reported-by: Alexander Pyhalov
Backpatch-through: 13
Discussion: https://postgr.es/m/acddb17c89b0d6cb940eaeda18c08bbe@postgrespro.ru
The Incremental Sort had a couple issues, resulting in leaking memory
during rescans, possibly triggering OOM. The code had a couple of
related flaws:
1. During rescans, the sort states were reset but then also set to NULL
(despite the comment saying otherwise). ExecIncrementalSort then
sees NULL and initializes a new sort state, leaking the memory used
by the old one.
2. Initializing the sort state also automatically rebuilt the info about
presorted keys, leaking the already initialized info. presorted_keys
was also unnecessarily reset to NULL.
Patch by James Coleman, based on patches by Laurenz Albe and Tom Lane.
Backpatch to 13, where Incremental Sort was introduced.
Author: James Coleman, Laurenz Albe, Tom Lane
Reported-by: Laurenz Albe, Zu-Ming Jiang
Backpatch-through: 13
Discussion: https://postgr.es/m/b2bd02dff61af15e3526293e2771f874cf2a3be7.camel%40cybertec.at
Discussion: https://postgr.es/m/db03c582-086d-e7cd-d4a1-3bc722f81765%40inf.ethz.ch
a316a3bc fixed the code in build_simpl_rel() that propagates
RelOptInfo.userid from parent to child rels so that it works
correctly for the child rels of a UNION ALL subquery rel, though
no tests were added in that commit. So do so here.
As noted in the discussion, coming up with a test case in the core
regression suite for this fix has turned out to be tricky, so the
test case is added to the postgres_fdw's suite instead.
postgresGetForeignRelSize()'s use of user mapping for the user
specified in RelOptInfo.userid makes it relatively easier to craft
a test case around.
Discussion: https://postgr.es/m/CA%2BHiwqH91GaFNXcXbLAM9L%3DzBwUmSyv699Mtv3i1_xtk9Xec_A%40mail.gmail.com
Backpatch-through: 16
The logic that introduced partitioned indexes missed a few things when
invalidating a partitioned index when these are created, still the code
is written to handle recursions:
1) If created from scratch because a mapping index could not be found,
the new index created could be itself invalid, if for example it was a
partitioned index with one of its leaves invalid.
2) A CCI was missing when indisvalid is set for a parent index, leading
to inconsistent trees when recursing across more than one level for a
partitioned index creation if an invalidation of the parent was
required.
This could lead to the creation of a partition index tree where some of
the partitioned indexes are marked as invalid, but some of the parents
are marked valid, which is not something that should happen (as
validatePartitionedIndex() defines, indisvalid is switched to true for a
partitioned index iff all its partitions are themselves valid).
This patch makes sure that indisvalid is set to false on a partitioned
index if at least one of its partition is invalid. The flag is set to
true if *all* its partitions are valid.
The regression test added in this commit abuses of a failed concurrent
index creation, marked as invalid, that maps with an index created on
its partitioned table afterwards.
Reported-by: Alexander Lakhin
Reviewed-by: Alexander Lakhin
Discussion: https://postgr.es/m/14987634-43c0-0cb3-e075-94d423607e08@gmail.com
Backpatch-through: 11
ALTER TABLE .. SET ACCESS METHOD was not registering a dependency to the
new access method with the relation altered in its rewrite phase, making
possible the drop of an access method even if there are relations that
depend on it. During the rewrite, a temporary relation is created to
build the new relation files before swapping the new and old files, and,
while the temporary relation was registering a correct dependency to the
new AM, the old relation did not do that. A dependency on the access
method is added when the relation files are swapped, which is the point
where pg_class is updated.
Materialized views and tables use the same code path, hence both were
impacted.
Backpatch down to 15, where this command has been introduced.
Reported-by: Alexander Lakhin
Reviewed-by: Nathan Bossart, Andres Freund
Discussion: https://postgr.es/m/18000-9145c25b1af475ca@postgresql.org
Backpatch-through: 15
run_command(check: true) already would have errorred out before the check is
reached.
Author: Tristan Partin <tristan@neon.tech>
Discussion: CSPIJVUDZFKX.3KHMOAVGF94RV@c3po
An outer join cannot be formed using an input path that is parameterized
by a value that is supposed to be nulled by the outer join. This is
obviously nonsensical, and it could lead to a bad plan being selected;
although currently it seems that we'll hit various sanity-check
assertions first.
I think that such cases were formerly prevented by the delay_upper_joins
mechanism, but now that that's gone we need an explicit check.
(Perhaps we should avoid generating baserel paths that could
lead to this situation in the first place; but it seems like
having a defense at the join level would be a good idea anyway.)
Richard Guo and Tom Lane, per report from Jaime Casanova
Discussion: https://postgr.es/m/CAJKUy5g2uZRrUDZJ8p-=giwcSHVUn0c9nmdxPSY0jF0Ov8VoEA@mail.gmail.com
If the given composite datum is toasted out-of-line,
DatumGetHeapTupleHeader will perform database accesses to detoast it.
That can invalidate the result of get_cached_rowtype, as documented
(perhaps not plainly enough) in that function's API spec; which leads
to strange errors or crashes when we try to use the TupleDesc to read
the tuple. In short then, trying to update a field of a composite
column could fail intermittently if the overall column value is wide
enough to require toasting.
We can fix the bug at no cost by just changing the order of
operations, since we don't need the TupleDesc until after detoasting.
(Other callers of get_cached_rowtype appear to get this right already,
so there's only one bug.)
Note that the added regression test case reveals this bug reliably
only with debug_discard_caches/CLOBBER_CACHE_ALWAYS.
Per bug #17994 from Alexander Lakhin. Sadly, this patch does not fix
the missing-values issue revealed in the bug discussion; we'll need
some more work to cover that.
Discussion: https://postgr.es/m/17994-5c7100b51b4790e9@postgresql.org
Meson validates 'choice' options for us, so technically this case is
impossible. A better error message helps people reading the code
understand what is going on in that branch.
Author: Tristan Partin <tristan@neon.tech>
Discussion: https://www.postgresql.org/message-id/flat/CSPIJVUDZFKX.3KHMOAVGF94RV%40c3po
Andres Freund