A couple of places weren't up to speed for this. By sheer good
luck, we didn't fail but just selected a non-memoized join plan,
at least in the test case we have. Nonetheless, it's a bug,
and I'm not quite sure that it couldn't have worse consequences
in other examples. So back-patch to v14 where Memoize came in.
Richard Guo
Discussion: https://postgr.es/m/CAMbWs48GkNom272sfp0-WeD6_0HSR19BJ4H1c9ZKSfbVnJsvRg@mail.gmail.com
It neglected to recurse to the subpath, meaning you'd get back
a path identical to the input. This could produce wrong query
results if the omission meant that the subpath fails to enforce
some join clause it should be enforcing. We don't have a test
case for this at the moment, but the code is obviously broken
and the fix is equally obvious. Back-patch to v14 where
Memoize was introduced.
Richard Guo
Discussion: https://postgr.es/m/CAMbWs4_R=ORpz=Lkn2q3ebPC5EuWyfZF+tmfCPVLBVK5W39mHA@mail.gmail.com
We might fail to generate a partitionwise join, because
reparameterize_path_by_child() does not support all path types.
This should not be a hard failure condition: we should just fall back
to a non-partitioned join. However, generate_partitionwise_join_paths
did not consider this possibility and would emit the (misleading)
error "could not devise a query plan for the given query" if we'd
failed to make any paths for a child join. Fix it to give up on
partitionwise joining if so. (The accepted technique for giving up
appears to be to set rel->nparts = 0, which I find pretty bizarre,
but there you have it.)
I have not added a test case because there'd be little point:
any omissions of this sort that we identify would soon get fixed
by extending reparameterize_path_by_child(), so the test would stop
proving anything. However, right now there is a known test case based
on failure to cover MaterialPath, and with that I've found that this
is broken in all supported versions. Hence, patch all the way back.
Original report and patch by me; thanks to Richard Guo for
identifying a test case that works against committed versions.
Discussion: https://postgr.es/m/1854233.1669949723@sss.pgh.pa.us
When updating a relation with a rule whose action performed an INSERT
from a multi-row VALUES list, the rewriter might skip processing the
VALUES list, and therefore fail to replace any DEFAULTs in it. This
would lead to an "unrecognized node type" error.
The reason was that RewriteQuery() assumed that a query doing an
INSERT from a multi-row VALUES list would necessarily only have one
item in its fromlist, pointing to the VALUES RTE to read from. That
assumption is correct for the original query, but not for product
queries produced for rule actions. In such cases, there may be
multiple items in the fromlist, possibly including multiple VALUES
RTEs.
What is required instead is for RewriteQuery() to skip any RTEs from
the product query's originating query, which might include one or more
already-processed VALUES RTEs. What's left should then include at most
one VALUES RTE (from the rule action) to be processed.
Patch by me. Thanks to Tom Lane for reviewing.
Back-patch to all supported branches.
Discussion: https://postgr.es/m/CAEZATCV39OOW7LAR_Xq4i%2BLc1Byux%3DeK3Q%3DHD_pF1o9LBt%3DphA%40mail.gmail.com
When the relkind of a relache entry changes, because a table is converted into
a view, pgstats can get confused in 15+, leading to crashes or assertion
failures.
For HEAD, Tom fixed this in b23cd185fd, by removing support for converting a
table to a view, removing the source of the inconsistency. This commit just
adds an assertion that a relcache entry's relkind does not change, just in
case we end up with another case of that in the future. As there's no cases of
changing relkind anymore, we can't add a test that that's handled correctly.
For 15, fix the problem by not maintaining the association with the old pgstat
entry when the relkind changes during a relcache invalidation processing. In
that case the pgstat entry needs to be unlinked first, to avoid
PgStat_TableStatus->relation getting out of sync. Also add a test reproducing
the issues.
No known problem exists in 11-14, so just add the test there.
Reported-by: vignesh C <vignesh21@gmail.com>
Author: Andres Freund <andres@anarazel.de>
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://postgr.es/m/CALDaNm2yXz+zOtv7y5zBd5WKT8O0Ld3YxikuU3dcyCvxF7gypA@mail.gmail.com
Discussion: https://postgr.es/m/CALDaNm3oZA-8Wbps2Jd1g5_Gjrr-x3YWrJPek-mF5Asrrvz2Dg@mail.gmail.com
Backpatch: 15-
Some options of these commands need to be able to identify the start
of the function body within the output of pg_get_functiondef().
It used to be that that always began with "AS", but since the
introduction of new-style SQL functions, it might also start with
"BEGIN" or "RETURN". Fix that on the psql side, and add some
regression tests.
Noted by me awhile ago, but I didn't do anything about it.
Thanks to David Johnston for a nag.
Discussion: https://postgr.es/m/AM9PR01MB8268D5CDABDF044EE9F42173FE8C9@AM9PR01MB8268.eurprd01.prod.exchangelabs.com
This has always worked, but you'd be unlikely to guess it
from the documentation. Add an example showing it.
Lack of docs noted by David Johnston. Back-patch to v13;
the documentation layout we used before that was not very
amenable to squeezing in multiple examples.
Discussion: https://postgr.es/m/CAKFQuwZ4Vy1Xty0G5Ok+ot=NDrU3C6hzF1JwUk-FEkwe3V9_RA@mail.gmail.com
In commit 40c24bfef, I forgot to use get_rule_expr_paren() for the
arguments of AT TIME ZONE, resulting in possibly not printing parens
for expressions that need it. But get_rule_expr_paren() wouldn't have
gotten it right anyway, because isSimpleNode() hadn't been taught that
COERCE_SQL_SYNTAX parent nodes don't guarantee sufficient parentheses.
Improve all that. Also use this methodology for F_IS_NORMALIZED, so
that we don't print useless parens for that.
In passing, remove a comment that was obsoleted later.
Per report from Duncan Sands. Back-patch to v14 where this code
came in. (Before that, we didn't try to print AT TIME ZONE that way,
so there was no bug just ugliness.)
Discussion: https://postgr.es/m/f41566aa-a057-6628-4b7c-b48770ecb84a@deepbluecap.com
Writing "pg_regress --dbname= ..." led to a crash, because
we weren't expecting there to be no database name supplied.
It doesn't seem like a great idea to run regression tests
in whatever is the user's default database; so rather than
supporting this case let's explicitly reject it.
Per report from Xing Guo. Back-patch to all supported
branches.
Discussion: https://postgr.es/m/CACpMh+A8cRvtvtOWVAZsCM1DU81GK4DL26R83y6ugZ1osV=ifA@mail.gmail.com
This also adds references to this new chapter at relevant sections of
our documentation. Previously much of these internal details were
exposed to users, but not explained. This also updates RELEASE
SAVEPOINT.
Discussion: https://postgr.es/m/CANbhV-E_iy9fmrErxrCh8TZTyenpfo72Hf_XD2HLDppva4dUNA@mail.gmail.com
Author: Simon Riggs, Laurenz Albe
Reviewed-by: Bruce Momjian
Backpatch-through: 11
The frontend-side routine in charge of building a SCRAM verifier
mentioned that the restrictions applying to SASLprep on the password
with the encoding are described at the top of fe-auth-scram.c, but this
information is in auth-scram.c.
This is wrong since 8f8b9be, so backpatch all the way down as this is an
important documentation bit.
Spotted while reviewing a different patch.
Backpatch-through: 11
Previously, we'd compress only when the active range of array entries
reached Max(4 * PROCARRAY_MAXPROCS, 2 * pArray->numKnownAssignedXids).
If max_connections is large, the first term could result in not
compressing for a long time, resulting in much wastage of cycles in
hot-standby backends scanning the array to take snapshots. Get rid
of that term, and just bound it to 2 * pArray->numKnownAssignedXids.
That however creates the opposite risk, that we might spend too much
effort compressing. Hence, consider compressing only once every 128
commit records. (This frequency was chosen by benchmarking. While
we only tried one benchmark scenario, the results seem stable over
a fairly wide range of frequencies.)
Also, force compression when processing RecoveryInfo WAL records
(which should be infrequent); the old code could perform compression
then, but would do so only after the same array-range check as for
the transaction-commit path.
Also, opportunistically run compression if the startup process is about
to wait for WAL, though not oftener than once a second. This should
prevent cases where we waste lots of time by leaving the array
not-compressed for long intervals due to low WAL traffic.
Lastly, add a simple check to keep us from uselessly compressing
when the array storage is already compact.
Back-patch, as the performance problem is worse in pre-v14 branches
than in HEAD.
Simon Riggs and Michail Nikolaev, with help from Tom Lane and
Andres Freund.
Discussion: https://postgr.es/m/CALdSSPgahNUD_=pB_j=1zSnDBaiOtqVfzo8Ejt5J_k7qZiU1Tw@mail.gmail.com
This is an oversight in commit 7c337b6b5: I apparently didn't think
about the possibility of a SQL function being executed multiple
times within a query. In that case, functions.c's primitive caching
mechanism allows the same utility parse tree to be presented for
execution more than once. We have to tell ProcessUtility to make
a working copy of the parse tree, or bad things happen.
Normally I'd add a regression test, but I think the reported crasher
is dependent on some rather random implementation choices that are
nowhere near functions.c, so its usefulness as a long-lived test
feels questionable. In any case, this fix is clearly correct given
the design choices of 7c337b6b5.
Per bug #17702 from Xin Wen. Thanks to Daniel Gustafsson for
analysis. Back-patch to v14 where the faulty commit came in
(before that, the responsibility for copying scribble-able
utility parse trees lay elsewhere).
Discussion: https://postgr.es/m/17702-ad24fdcdd1e9047a@postgresql.org
The JOIN_SEMI case Assert'ed that there are no PlaceHolderVars that
need to be evaluated at the semijoin's RHS, which is wrong because
there could be some in the semijoin's qual condition. However, there
could not be any references further up than that, and within the qual
there is not any way that such a PHV could have gone to null yet, so
we don't really need the PHV and there is no need to avoid making the
RHS-removal optimization. The upshot is that there's no actual bug
in production code, and we ought to just remove this misguided Assert.
While we're here, also drop the JOIN_RIGHT case, which is dead code
because reduce_outer_joins() already got rid of JOIN_RIGHT.
Per bug #17700 from Xin Wen. Uselessness of the JOIN_RIGHT case
pointed out by Richard Guo. Back-patch to v12 where this code
was added.
Discussion: https://postgr.es/m/17700-2b5c10d917c30687@postgresql.org
When loading plperl built against Strawberry perl or the msys2 ucrt perl
that have been built with gcc, a binary mismatch has been encountered
which looks like this:
loadable library and perl binaries are mismatched (got handshake key 0000000012800080, needed 0000000012900080)
To cure this we bring the handshake keys into sync by adding
NO_THREAD_SAFE_LOCALE to the defines used to build plperl.
Discussion: https://postgr.es/m/20211005004334.tgjmro4kuachwiuc@alap3.anarazel.de
Discussion: https://postgr.es/m/c2da86a0-2906-744c-923d-16da6047875e@dunslane.net
Backpatch to all live branches.
Commit 9b4eafcaf4 added creattion of a directory to reserve TAP test
ports at the top of the build tree. In a non-vpath build this means at
the top of the source tree, so it needs to be added to .gitignore.
As suggested by Michael Paquier
Backpatch to all live branches.
Strawberry uses __builtin_expect which Visual C doesn't have. For this
case define it as a noop. Solution taken from vim sources.
Backpatch to all live branches
Commit b663a4136, which allowed FDWs to INSERT rows in bulk, added to
nodeModifyTable.c code to flush pending inserts to the foreign-table
result relation(s) before completing processing of the ModifyTable node,
but the code failed to take into account the case where the INSERT query
has modifying CTEs, leading to incorrect results.
Also, that commit failed to flush pending inserts before firing BEFORE
ROW triggers so that rows are visible to such triggers.
In that commit we scanned through EState's
es_tuple_routing_result_relations or es_opened_result_relations list to
find the foreign-table result relations to which pending inserts are
flushed, but that would be inefficient in some cases. So to fix, 1) add
a List member to EState to record the insert-pending result relations,
and 2) modify nodeModifyTable.c so that it adds the foreign-table result
relation to the list in ExecInsert() if appropriate, and flushes pending
inserts properly using the list where needed.
While here, fix a copy-and-pasteo in a comment in ExecBatchInsert(),
which was added by that commit.
Back-patch to v14 where that commit appeared.
Discussion: https://postgr.es/m/CAPmGK16qutyCmyJJzgQOhfBq%3DNoGDqTB6O0QBZTihrbqre%2BoxA%40mail.gmail.com
In commit 272248a0c, we introduced an InitialRunningXacts array to
remember transactions and subtransactions that were running when the
xl_running_xacts record that we decoded was written. This array was
allocated in the snapshot builder memory context after we restore
serialized snapshot but we forgot to reset the array while freeing the
builder memory context. So, the next time when we start decoding in the
same session where we don't restore any serialized snapshot, we ended up
using the uninitialized array and that can lead to unpredictable behavior.
This problem doesn't exist in HEAD as instead of using
InitialRunningXacts, we added the list of transaction IDs and
sub-transaction IDs, that have modified catalogs and are running during
snapshot serialization, to the serialized snapshot (see commit 7f13ac8123).
Reported-by: Maxim Orlov
Author: Masahiko Sawada
Reviewed-by: Amit Kapila, Maxim Orlov
Backpatch-through: 11
Discussion: https://postgr.es/m/CACG=ezZoz_KG+Ryh9MrU_g5e0HiVoHocEvqFF=NRrhrwKmEQJQ@mail.gmail.com
We've made multiple attempts at preventing get_actual_variable_range
from taking an unreasonable amount of time (3ca930fc3, fccebe421).
But there's still an issue for the very first planning attempt after
deletion of a large number of extremal-valued tuples. While that
planning attempt will set "killed" bits on the tuples it visits and
thereby reduce effort for next time, there's still a lot of work it
has to do to visit the heap and then set those bits. It's (usually?)
not worth it to do that much work at plan time to have a slightly
better estimate, especially in a context like this where the table
contents are known to be mutating rapidly.
Therefore, let's bound the amount of work to be done by giving up
after we've visited 100 heap pages. Giving up just means we'll
fall back on the extremal value recorded in pg_statistic, so it
shouldn't mean that planner estimates suddenly become worthless.
Note that this means we'll still gradually whittle down the problem
by setting a few more index "killed" bits in each planning attempt;
so eventually we'll reach a good state (barring further deletions),
even in the absence of VACUUM.
Simon Riggs, per a complaint from Jakub Wartak (with cosmetic
adjustments by me). Back-patch to all supported branches.
Discussion: https://postgr.es/m/CAKZiRmznOwi0oaV=4PHOCM4ygcH4MgSvt8=5cu_vNCfc8FSUug@mail.gmail.com
Currently there is a race condition where if concurrent TAP tests both
test that they can open a port they will assume that it is free and use
it, causing one of them to fail. To prevent this we record a reservation
using an exclusive lock, and any TAP test that discovers a reservation
checks to see if the reserving process is still alive, and looks for
another free port if it is.
Ports are reserved in a directory set by the environment setting
PG_TEST_PORT_DIR, or if that doesn't exist a subdirectory of the top
build directory as set by Makefile.global, or its own
tmp_check directory.
The prove_check recipe in Makefile.global.in is extended to export
top_builddir to the TAP tests. This was already exported by the
prove_installcheck recipes.
Per complaint from Andres Freund
Backpatched from 9b4eafcaf4 to all live branches
Discussion: https://postgr.es/m/20221002164931.d57hlutrcz4d2zi7@awork3.anarazel.de
Once a logical slot has acquired a catalog_xmin, it doesn't let go of
it, even when invalidated by exceeding the max_slot_wal_keep_size, which
means that dead catalog tuples are not removed by vacuum anymore since
the point is invalidated, until the slot is dropped. This could be
catastrophic if catalog churn is high.
Change the computation of Xmin to ignore invalidated slots,
to prevent dead rows from accumulating.
Backpatch to 13, where slot invalidation appeared.
Author: Sirisha Chamarthi <sirichamarthi22@gmail.com>
Reviewed-by: Ashutosh Bapat <ashutosh.bapat.oss@gmail.com>
Discussion: https://postgr.es/m/CAKrAKeUEDeqquN9vwzNeG-CN8wuVsfRYbeOUV9qKO_RHok=j+g@mail.gmail.com
The Hunspell project moved from Sourceforge to Github sometime
in 2016, so update our links to match the new URL. Backpatch
the doc changes to all supported versions.
Discussion: https://postgr.es/m/DC9A662A-360D-4125-A453-5A6CB9C6C4B4@yesql.se
Backpatch-through: v11
I just spent an annoying amount of time reverse-engineering the
100%-undocumented API between ts_headline and the text search
parser's prsheadline function. Add some commentary about that
while it's fresh in mind. Also remove some unused macros in
wparser_def.c.
While at it, I noticed that when commit 78e73e875 added a
CHECK_FOR_INTERRUPTS call in TS_execute_recurse, it missed
doing so in the parallel function TS_phrase_execute, which
surely needs one just as much.
Back-patch because of the missing CHECK_FOR_INTERRUPTS.
Might as well back-patch the rest of this too.
This reverts commit 5cda142bb9
(in v14 only).
It turns out that that fails under force_parallel_mode = regress,
because pageinspect's disk-access functions are marked parallel
safe, which they are not if you try to use them on a temp table.
The cost of fixing that pre-v15 seems to exceed the value of
making this test case fully stable, so we will just leave things
as-is in v14.
pageinspect has occasionally failed on slow buildfarm members,
with symptoms indicating that the expected effects of VACUUM
FREEZE didn't happen. This is presumably because a background
transaction such as auto-analyze was holding back global xmin.
We can work around that by using a temp table in the test.
Since commit a7212be8b, that will use an up-to-date cutoff xmin
regardless of other processes. And pageinspect itself shouldn't
really care whether the table is temp.
Back-patch to v14. There would be no point in older branches
without back-patching a7212be8b, which seems like more trouble
than the problem is worth.
Discussion: https://postgr.es/m/2892135.1668976646@sss.pgh.pa.us
ProcSleep() used a PGPROC* variable to point to PROC_QUEUE->links.next,
because that does "the right thing" with SHMQueueInsertBefore(). While that
largely works, it's certainly not correct and unnecessary - we can just use
SHM_QUEUE* to point to the insertion point.
Noticed when testing a 32bit of postgres with undefined behavior
sanitizer. UBSan noticed that sometimes the supposed PGPROC wasn't
sufficiently aligned (required since 46d6e5f567, ensured indirectly, via
ShmemAllocRaw() guaranteeing cacheline alignment).
For now fix this by using a SHM_QUEUE* for the insertion point. Subsequently
we should replace all the use of PROC_QUEUE and SHM_QUEUE with ilist.h, but
that's a larger change that we don't want to backpatch.
Backpatch to all supported versions - it's useful to be able to run postgres
under UBSan.
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://postgr.es/m/20221117014230.op5kmgypdv2dtqsf@awork3.anarazel.de
Backpatch: 11-
basics.source is supposed to be pretty closely in step with
the examples in chapter 2 of the tutorial, but I forgot to
update it in commit f05a5e000. Fix that, and adjust a couple
of other discrepancies that had crept in over time.
(I notice that advanced.source is nowhere near being in sync
with chapter 3, but I lack the ambition to do something
about that right now.)
Version strings with unequal numbers of parts were being compared
incorrectly. We cure this by treating a missing part in the shorter
version as 0.
per complaint from Jehan-Guillaume de Rorthais, but the fix is mine, not
his.
Discussion: https://postgr.es/m/20220628225325.53d97b8d@karst
Backpatch to release 14 where this code was introduced.
This is a back-patch of the v15-era commit f10f0ae42 into older
supported branches. The idea is to design out bugs in which an
ill-timed relcache flush clears rel->rd_smgr partway through
some code sequence that wasn't expecting that. We had another
report today of a corner case that reliably crashes v14 under
debug_discard_caches (nee CLOBBER_CACHE_ALWAYS), and therefore
would crash once in a blue moon in the field. We're unlikely
to get rid of all such code paths unless we adopt the more
rigorous coding rules instituted by f10f0ae42. Therefore,
even though this is a bit invasive, it's time to back-patch.
Some comfort can be taken in the fact that f10f0ae42 has been
in v15 for 16 months without problems.
I left the RelationOpenSmgr macro present in the back branches,
even though no core code should use it anymore, in order to not break
third-party extensions in minor releases. Such extensions might opt
to start using RelationGetSmgr instead, to reduce their code
differential between v15 and earlier branches. This carries a hazard
of failing to compile against headers from existing minor releases.
However, once compiled the extension should work fine even with such
releases, because RelationGetSmgr is a "static inline" function so
it creates no link-time dependency. So depending on distribution
practices, that might be an OK tradeoff.
Per report from Spyridon Dimitrios Agathos. Original patch
by Amul Sul.
Discussion: https://postgr.es/m/CAFM5RaqdgyusQvmWkyPYaWMwoK5gigdtW-7HcgHgOeAw7mqJ_Q@mail.gmail.com
Discussion: https://postgr.es/m/CANiYTQsU7yMFpQYnv=BrcRVqK_3U3mtAzAsJCaqtzsDHfsUbdQ@mail.gmail.com
During XLOG_HASH_SPLIT_ALLOCATE_PAGE replay, we were checking for a
cleanup lock on the new bucket page after acquiring an exclusive lock on
it and raising a PANIC error on failure. However, it is quite possible
that checkpointer can acquire the pin on the same page before acquiring a
lock on it, and then the replay will lead to an error. So instead, directly
acquire the cleanup lock on the new bucket page during
XLOG_HASH_SPLIT_ALLOCATE_PAGE replay operation.
Reported-by: Andres Freund
Author: Robert Haas
Reviewed-By: Amit Kapila, Andres Freund, Vignesh C
Backpatch-through: 11
Discussion: https://postgr.es/m/20220810022617.fvjkjiauaykwrbse@awork3.anarazel.de
The original report was concerned with a possible inconsistency
between the heap and the visibility map, which I was unable to
confirm. The concern has been retracted.
However, there did seem to be a torn page hazard when using
checksums. By not setting the heap page LSN during redo, the
protections of minRecoveryPoint were bypassed. Fixed, along with a
misleading comment.
It may have been impossible to hit this problem in practice, because
it would require a page tear between the checksum and the flags, so I
am marking this as a theoretical risk. But, as discussed, it did
violate expectations about the page LSN, so it may have other
consequences.
Backpatch to all supported versions.
Reported-by: Konstantin Knizhnik
Reviewed-by: Konstantin Knizhnik
Discussion: https://postgr.es/m/fed17dac-8cb8-4f5b-d462-1bb4908c029e@garret.ru
Backpatch-through: 11
The stanza "SET STORAGE may need to add a TOAST table" does not
test what it's supposed to, and hasn't done so since we added
the ability to store constant column default values as metadata.
We need to use a non-constant default to get the expected table
rewrite to actually happen.
Fix that, and add the missing checks that would have exposed the
problem to begin with.
Noted while reviewing a patch that made changes in this test case.
Back-patch to v11 where the problem came in.
In commit 450ee7012 I supposed that all platforms we now care about have
snprintf(), since that's required by C99. Turns out that Microsoft did
not get around to adding that until VS2015. We've dropped support for
VS2013 as of HEAD (cf 6203583b7), but not in the back branches, so add
a hack for this in the back branches only.
There's no easy shortcut to an exact emulation of standard snprintf
in VS2013, but fortunately we don't need one: this code was just fine
with using sprintf before 450ee7012, so we can make it do so again
on that platform (and any others where the problem might crop up).
Per bug #17681 from Daisuke Higuchi. Back-patch to v12, like the
previous patch.
Discussion: https://postgr.es/m/17681-485ba2ec13e7f392@postgresql.org
Add a little to the header comments for these functions to make it
clearer what guarantees about commit behavior are provided to callers.
(See commit f92944137 for context.)
Although this is only a comment change, it's really documentation
aimed at authors of extensions, so it seems appropriate to back-patch.
Yugo Nagata and Tom Lane, per further discussion of bug #17434.
Discussion: https://postgr.es/m/17434-d9f7a064ce2a88a3@postgresql.org
Commit fede15417 introduced FILTER by jamming it into the existing
example introducing HAVING, which seems pedagogically poor to me;
and it added no information about what the keyword actually does.
Not to mention that the claimed output didn't match the sample
data being used in this running example.
Revert that and instead make an independent example using FILTER.
To help drive home the point that it's a per-aggregate filter,
we need to use two aggregates not just one; for consistency
expand all the examples in this segment to do that.
Also adjust the example using WHERE ... LIKE so that it'd produce
nonempty output with this sample data, and show that output.
Back-patch, as the previous patch was. (Sadly, v10 is now out
of scope.)
Discussion: https://postgr.es/m/166794307526.652.9073408178177444190@wrigleys.postgresql.org
Tom Lane