This reverts commit 81069a9efc.
Buildfarm results suggest that some platforms have versions of pselect(2)
that are not merely non-atomic, but flat out non-functional. Revert the
use-pselect patch to confirm this diagnosis (and exclude the no-SA_RESTART
patch as the source of trouble). If it's so, we should probably look into
blacklisting specific platforms that have broken pselect.
Discussion: https://postgr.es/m/9696.1493072081@sss.pgh.pa.us
Traditionally we've unblocked signals, called select(2), and then blocked
signals again. The code expects that the select() will be cancelled with
EINTR if an interrupt occurs; but there's a race condition, which is that
an already-pending signal will be delivered as soon as we unblock, and then
when we reach select() there will be nothing preventing it from waiting.
This can result in a long delay before we perform any action that
ServerLoop was supposed to have taken in response to the signal. As with
the somewhat-similar symptoms fixed by commit 893902085, the main practical
problem is slow launching of parallel workers. The window for trouble is
usually pretty short, corresponding to one iteration of ServerLoop; but
it's not negligible.
To fix, use pselect(2) in place of select(2) where available, as that's
designed to solve exactly this problem. Where not available, we continue
to use the old way, and are no worse off than before.
pselect(2) has been required by POSIX since about 2001, so most modern
platforms should have it. A bigger portability issue is that some
implementations are said to be non-atomic, ie pselect() isn't really
any different from unblock/select/reblock. Still, we're no worse off
than before on such a platform.
There is talk of rewriting the postmaster to use a WaitEventSet and
not do signal response work in signal handlers, at which point this
could be reverted, since we'd be using a self-pipe to solve the race
condition. But that's not happening before v11 at the earliest.
Back-patch to 9.6. The problem exists much further back, but the
worst symptom arises only in connection with parallel query, so it
does not seem worth taking any portability risks in older branches.
Discussion: https://postgr.es/m/9205.1492833041@sss.pgh.pa.us
The postmaster keeps signals blocked everywhere except while waiting
for something to happen in ServerLoop(). The code expects that the
select(2) will be cancelled with EINTR if an interrupt occurs; without
that, followup actions that should be performed by ServerLoop() itself
will be delayed. However, some platforms interpret the SA_RESTART
signal flag as meaning that they should restart rather than cancel
the select(2). Worse yet, some of them restart it with the original
timeout delay, meaning that a steady stream of signal interrupts can
prevent ServerLoop() from iterating at all if there are no incoming
connection requests.
Observable symptoms of this, on an affected platform such as HPUX 10,
include extremely slow parallel query startup (possibly as much as
30 seconds) and failure to update timestamps on the postmaster's sockets
and lockfiles when no new connections arrive for a long time.
We can fix this by running the postmaster's signal handlers without
SA_RESTART. That would be quite a scary change if the range of code
where signals are accepted weren't so tiny, but as it is, it seems
safe enough. (Note that postmaster children do, and must, reset all
the handlers before unblocking signals; so this change should not
affect any child process.)
There is talk of rewriting the postmaster to use a WaitEventSet and
not do signal response work in signal handlers, at which point it might
be appropriate to revert this patch. But that's not happening before
v11 at the earliest.
Back-patch to 9.6. The problem exists much further back, but the
worst symptom arises only in connection with parallel query, so it
does not seem worth taking any portability risks in older branches.
Discussion: https://postgr.es/m/9205.1492833041@sss.pgh.pa.us
poll.h is mandated by Single Unix Spec v2, the usual baseline for
postgres on unix. None of the unixoid buildfarms animals has
sys/poll.h but not poll.h. Therefore there's not much point to test
for sys/poll.h's existence and include it optionally.
Author: Andres Freund, per suggestion from Tom Lane
Discussion: https://postgr.es/m/20505.1492723662@sss.pgh.pa.us
It doesn't make any immediate difference to PostgreSQL, but might as well
follow the standard, since one exists. (I looked at RFC 5803 earlier, but
didn't fully understand it back then.)
The new format uses Base64 instead of hex to encode StoredKey and
ServerKey, which makes the verifiers slightly smaller. Using the same
encoding for the salt and the keys also means that you only need one
encoder/decoder instead of two. Although we have code in the backend to
do both, we are talking about teaching libpq how to create SCRAM verifiers
for PQencodePassword(), and libpq doesn't currently have any code for hex
encoding.
Bump catversion, because this renders any existing SCRAM verifiers in
pg_authid invalid.
Discussion: https://www.postgresql.org/message-id/351ba574-85ea-d9b8-9689-8c928dd0955d@iki.fi
In pg_subscription, subconninfo, subslotname, subsynccommit and
subpublications are expected not to be NULL. Therefore this patch
adds BKI_FORCE_NOT_NULL markings to them.
This patch is basically unnecessary unless the code has a bug which
wrongly sets either of those columns to NULL. But it's good to have
this as a safeguard.
Author: Masahiko Sawada
Reviewed-by: Kyotaro Horiguchi
Reported-by: Fujii Masao
Discussion: http://postgr.es/m/CAHGQGwFDWh_Qr-q_GEMpD+qH=vYPMdVqw=ZOSY3kX_Pna9R9SA@mail.gmail.com
is_parallel_safe() supposed that the only relevant property of a SubPlan
was the parallel safety of the referenced subplan tree. This is wrong:
the testexpr or args subtrees might contain parallel-unsafe stuff, as
demonstrated by the test case added here. However, just recursing into the
subtrees fails in a different way: we'll typically find PARAM_EXEC Params
representing the subplan's output columns in the testexpr. The previous
coding supposed that any Param must be treated as parallel-restricted, so
that a naive attempt at fixing this disabled parallel pushdown of SubPlans
altogether. We must instead determine, for any visited Param, whether it
is one that would be computed by a surrounding SubPlan node; if so, it's
safe to push down along with the SubPlan node.
We might later be able to extend this logic to cope with Params used for
correlated subplans and other cases; but that's a task for v11 or beyond.
Tom Lane and Amit Kapila
Discussion: https://postgr.es/m/7064.1492022469@sss.pgh.pa.us
Since it appears that v10 is going to move the goalposts by some amount
in terms of where you can and can't invoke set-returning functions,
arrange for the executor's "set-valued function called in context that
cannot accept a set" errors to include a syntax position if possible,
pointing to the specific SRF that can't be called where it's located.
The main bit of infrastructure needed for this is to make the query source
text accessible in the executor; but it turns out that commit 4c728f382
already did that. We just need a new function executor_errposition()
modeled on parser_errposition(), and we're ready to rock.
While experimenting with this, I noted that the error position wasn't
properly reported if it occurred in a plpgsql FOR-over-query loop,
which turned out to be because SPI_cursor_open_internal wasn't providing
an error context callback during PortalStart. Fix that.
There's a whole lot more that could be done with this infrastructure
now that it's there, but this is not the right time in the development
cycle for that sort of work. Hence, resist the temptation to plaster
executor_errposition() calls everywhere ... for the moment.
Discussion: https://postgr.es/m/5263.1492471571@sss.pgh.pa.us
* Be sure to reset the launcher's pid (LogicalRepCtx->launcher_pid) to 0
even when the launcher emits an error.
* Declare ApplyLauncherWakeup() as a static function because it's called
only in launcher.c.
* Previously IsBackendPId() was used to check whether the launcher's pid
was valid. IsBackendPid() was necessary because there was the bug where
the launcher's pid was not reset to 0. But now it's fixed, so IsBackendPid()
is not necessary and this patch removes it.
Author: Masahiko Sawada
Reviewed-by: Kyotaro Horiguchi
Reported-by: Fujii Masao
Discussion: http://postgr.es/m/CAHGQGwFDWh_Qr-q_GEMpD+qH=vYPMdVqw=ZOSY3kX_Pna9R9SA@mail.gmail.com
Per discussion, plain "scram" is confusing because we actually implement
SCRAM-SHA-256 rather than the original SCRAM that uses SHA-1 as the hash
algorithm. If we add support for SCRAM-SHA-512 or some other mechanism in
the SCRAM family in the future, that would become even more confusing.
Most of the internal files and functions still use just "scram" as a
shorthand for SCRMA-SHA-256, but I did change PASSWORD_TYPE_SCRAM to
PASSWORD_TYPE_SCRAM_SHA_256, as that could potentially be used by 3rd
party extensions that hook into the password-check hook.
Michael Paquier did this in an earlier version of the SCRAM patch set
already, but I didn't include that in the version that was committed.
Discussion: https://www.postgresql.org/message-id/fde71ff1-5858-90c8-99a9-1c2427e7bafb@iki.fi
Seems to have been introduced in commit c219d9b0a. I think there indeed
was a "tupbasics.h" in some early drafts of that refactoring, but it
didn't survive into the committed version.
Amit Kapila
We'd already recognized that we can't pass function pointers across process
boundaries for functions in loadable modules, since a shared library could
get loaded at different addresses in different processes. But actually the
practice doesn't work for functions in the core backend either, if we're
using EXEC_BACKEND. This is the cause of recent failures on buildfarm
member culicidae. Switch to passing a string function name in all cases.
Something like this needs to be back-patched into 9.6, but let's see
if the buildfarm likes it first.
Petr Jelinek, with a bunch of basically-cosmetic adjustments by me
Discussion: https://postgr.es/m/548f9c1d-eafa-e3fa-9da8-f0cc2f654e60@2ndquadrant.com
Standardize on testing a hash index page's type by doing
(opaque->hasho_flag & LH_PAGE_TYPE) == LH_xxx_PAGE
Various places were taking shortcuts like
opaque->hasho_flag & LH_BUCKET_PAGE
which while not actually wrong, is still bad practice because
it encourages use of
opaque->hasho_flag & LH_UNUSED_PAGE
which *is* wrong (LH_UNUSED_PAGE == 0, so the above is constant false).
hash_xlog.c's hash_mask() contained such an incorrect test.
This also ensures that we mask out the additional flag bits that
hasho_flag has accreted since 9.6. pgstattuple's pgstat_hash_page(),
for one, was failing to do that and was thus actively broken.
Also fix assorted comments that hadn't been updated to reflect the
extended usage of hasho_flag, and fix some macros that were testing
just "(hasho_flag & bit)" to use the less dangerous, project-approved
form "((hasho_flag & bit) != 0)".
Coverity found the bug in hash_mask(); I noted the one in
pgstat_hash_page() through code reading.
This contains some protocol changes to SASL authentiation (which is new
in v10):
* For future-proofing, in the AuthenticationSASL message that begins SASL
authentication, provide a list of SASL mechanisms that the server
supports, for the client to choose from. Currently, it's always just
SCRAM-SHA-256.
* Add a separate authentication message type for the final server->client
SASL message, which the client doesn't need to respond to. This makes
it unambiguous whether the client is supposed to send a response or not.
The SASL mechanism should know that anyway, but better to be explicit.
Also, in the server, support clients that don't send an Initial Client
response in the first SASLInitialResponse message. The server is supposed
to first send an empty request in that case, to which the client will
respond with the data that usually comes in the Initial Client Response.
libpq uses the Initial Client Response field and doesn't need this, and I
would assume any other sensible implementation to use Initial Client
Response, too, but let's follow the SASL spec.
Improve the documentation on SASL authentication in protocol. Add a
section describing the SASL message flow, and some details on our
SCRAM-SHA-256 implementation.
Document the different kinds of PasswordMessages that the frontend sends
in different phases of SASL authentication, as well as GSS/SSPI
authentication as separate message formats. Even though they're all 'p'
messages, and the exact format depends on the context, describing them as
separate message formats makes the documentation more clear.
Reviewed by Michael Paquier and Álvaro Hernández Tortosa.
Discussion: https://www.postgresql.org/message-id/CAB7nPqS-aFg0iM3AQOJwKDv_0WkAedRjs1W2X8EixSz+sKBXCQ@mail.gmail.com
Commit 5e6d8d2bb allowed parallel workers to execute parallel-safe
subplans, but it transmitted the query's entire list of subplans to
the worker(s). Since execMain.c blindly does ExecInitNode and later
ExecEndNode on every list element, this resulted in parallel-unsafe plan
nodes nonetheless getting started up and shut down in parallel workers.
That seems mostly harmless as far as core plan node types go (but
maybe not so much for Gather?). But it resulted in postgres_fdw
opening and then closing extra remote connections, and it's likely
that other non-parallel-safe FDWs or custom scan providers would have
worse reactions.
To fix, just make ExecSerializePlan replace parallel-unsafe subplans
with NULLs in the cut-down plan tree that it transmits to workers.
This relies on ExecInitNode and ExecEndNode to do nothing on NULL
input, but they do anyway. If anything else is touching the dropped
subplans in a parallel worker, that would be a bug to be fixed.
(This thus provides a strong guarantee that we won't try to do
something with a parallel-unsafe subplan in a worker.)
This is, I think, the last fix directly occasioned by Andreas Seltenreich's
bug report of a few days ago.
Tom Lane and Amit Kapila
Discussion: https://postgr.es/m/87tw5x4vcu.fsf@credativ.de
We'd managed to avoid doing this so far, but it seems pretty obvious
that it would be forced on us some day, and this is much the cleanest
way of approaching the open problem that parallel-unsafe subplans are
being transmitted to parallel workers. Anyway there's no space cost
due to alignment considerations, and the time cost is pretty minimal
since we're just copying the flag from the corresponding Path node.
(At least in most cases ... some of the klugier spots in createplan.c
have to work a bit harder.)
In principle we could perhaps get rid of SubPlan.parallel_safe,
but I thought it better to keep that in case there are reasons to
consider a SubPlan unsafe even when its child plan is parallel-safe.
This patch doesn't actually do anything with the new flags, but
I thought I'd commit it separately anyway.
Note: although this touches outfuncs/readfuncs, there's no need for
a catversion bump because Plan trees aren't stored on disk.
Discussion: https://postgr.es/m/87tw5x4vcu.fsf@credativ.de
To prevent future bugs along the lines of the one corrected by commit
8ff518699f, or find any that remain
in the current code, add an Assert() that the difference between
parallel_register_count and parallel_terminate_count is in a sane
range.
Kuntal Ghosh, with considerable tidying-up by me, per a suggestion
from Neha Khatri. Reviewed by Tomas Vondra.
Discussion: http://postgr.es/m/CAFO0U+-E8yzchwVnvn5BeRDPgX2z9vZUxQ8dxx9c0XFGBC7N1Q@mail.gmail.com
This commit also does
- add REPLICATION_SUBSCRIBERS into config_group
- mark max_logical_replication_workers and max_sync_workers_per_subscription
as REPLICATION_SUBSCRIBERS parameters
- move those parameters into "Subscribers" section in postgresql.conf.sample
Author: Masahiko Sawada, Petr Jelinek and me
Reported-by: Masahiko Sawada
Discussion: http://postgr.es/m/CAD21AoAonSCoa=v=87ZO3vhfUZA1k_E2XRNHTt=xioWGUa+0ug@mail.gmail.com
This used to mean "Visual C++ except in those parts where Borland C++
was supported where it meant one of those". Now that we don't support
Borland C++ anymore, simplify by using _MSC_VER which is the normal way
to detect Visual C++.
This removes the support for building just libpq using Borland C++ or
Visual C++. This has not worked properly for years, and given the number
of complaints it's clearly not worth the maintenance burden.
Building libpq using the standard MSVC build system is of course still
supported, along with mingw.
This extends the castNode() notation introduced by commit 5bcab1114 to
provide, in one step, extraction of a list cell's pointer and coercion to
a concrete node type. For example, "lfirst_node(Foo, lc)" is the same
as "castNode(Foo, lfirst(lc))". Almost half of the uses of castNode
that have appeared so far include a list extraction call, so this is
pretty widely useful, and it saves a few more keystrokes compared to the
old way.
As with the previous patch, back-patch the addition of these macros to
pg_list.h, so that the notation will be available when back-patching.
Patch by me, after an idea of Andrew Gierth's.
Discussion: https://postgr.es/m/14197.1491841216@sss.pgh.pa.us
We decided in f1b4c771ea to pass the
original slot to ExecConstraints(), but that breaks when there are
BEFORE ROW triggers involved. So we need to do reverse-map the tuples
back to the original descriptor instead, as Amit originally proposed.
Amit Langote, reviewed by Ashutosh Bapat. One overlooked comment
fixed by me.
Discussion: http://postgr.es/m/b3a17254-6849-e542-2353-bde4e880b6a4@lab.ntt.co.jp
Commit 4deb41381 modified isolationtester's query to see whether a
session is blocked to also check for waits occurring in GetSafeSnapshot.
However, it did that in a way that enormously increased the query's
runtime under CLOBBER_CACHE_ALWAYS, causing the buildfarm members
that use that to run about four times slower than before, and in some
cases fail entirely. To fix, push the entire logic into a dedicated
backend function. This should actually reduce the CLOBBER_CACHE_ALWAYS
runtime from what it was previously, though I've not checked that.
In passing, expose a SQL function to check for safe-snapshot blockage,
comparable to pg_blocking_pids. This is more or less free given the
infrastructure built to solve the other problem, so we might as well.
Thomas Munro
Discussion: https://postgr.es/m/20170407165749.pstcakbc637opkax@alap3.anarazel.de
Defaults match the fixed behavior of prior releases, but now DBAs
have better options to tune serializable workloads.
It might be nice to be able to set this per relation, but that part
will need to wait for another release.
Author: Dagfinn Ilmari Mannsåker
If there can certainly be no more than one matching inner row for a given
outer row, then the executor can move on to the next outer row as soon as
it's found one match; there's no need to continue scanning the inner
relation for this outer row. This saves useless scanning in nestloop
and hash joins. In merge joins, it offers the opportunity to skip
mark/restore processing, because we know we have not advanced past the
first possible match for the next outer row.
Of course, the devil is in the details: the proof of uniqueness must
depend only on joinquals (not otherquals), and if we want to skip
mergejoin mark/restore then it must depend only on merge clauses.
To avoid adding more planning overhead than absolutely necessary,
the present patch errs in the conservative direction: there are cases
where inner_unique or skip_mark_restore processing could be used, but
it will not do so because it's not sure that the uniqueness proof
depended only on "safe" clauses. This could be improved later.
David Rowley, reviewed and rather heavily editorialized on by me
Discussion: https://postgr.es/m/CAApHDvqF6Sw-TK98bW48TdtFJ+3a7D2mFyZ7++=D-RyPsL76gw@mail.gmail.com
When the 64bit atomics simulation is in use, we can't necessarily
guarantee the correct alignment of the atomics due to lack of compiler
support for doing so- that's fine from a safety perspective, because
everything is protected by a lock, but we asserted the alignment in
all cases. Weaken them. Per complaint from Alvaro Herrera.
My #ifdefery for PG_HAVE_8BYTE_SINGLE_COPY_ATOMICITY wasn't
sufficient. Fix that. Per complaint from Alexander Korotkov.
Instead of allocating memory in brin_deform_tuple and brin_copy_tuple
over and over during a scan, allow reuse of previously allocated memory.
This is said to make for a measurable performance improvement.
Author: Jinyu Zhang, Álvaro Herrera
Reviewed by: Tomas Vondra
Discussion: https://postgr.es/m/495deb78.4186.1500dacaa63.Coremail.beijing_pg@163.com
When adding atomics back in b64d92f1a, I added 64bit support as
optional; there wasn't yet a direct user in sight. That turned out to
be a bit short-sighted, it'd already have been useful a number of times.
Add a fallback implementation of 64bit atomics, just like the one we
have for 32bit atomics.
Additionally optimize reads/writes to 64bit on a number of platforms
where aligned writes of that size are atomic. This can now be tested
with PG_HAVE_8BYTE_SINGLE_COPY_ATOMICITY.
Author: Andres Freund
Reviewed-By: Amit Kapila
Discussion: https://postgr.es/m/20160330230914.GH13305@awork2.anarazel.de
The WAL-writing piece was forgetting to set the pages-per-range value.
Also, fix the declared type of struct member heapBlk, which I mistakenly
set as OffsetNumber rather than BlockNumber.
Problem was introduced by commit c655899ba9 (April 1st). Any system
that tries to replay the new WAL record written before this fix is
likely to die on replay and require pg_resetwal.
Reported by Tom Lane.
Discussion: https://postgr.es/m/20191.1491524824@sss.pgh.pa.us
An important step of SASLprep normalization, is to convert the string to
Unicode normalization form NFKC. Unicode normalization requires a fairly
large table of character decompositions, which is generated from data
published by the Unicode consortium. The script to generate the table is
put in src/common/unicode, as well test code for the normalization.
A pre-generated version of the tables is included in src/include/common,
so you don't need the code in src/common/unicode to build PostgreSQL, only
if you wish to modify the normalization tables.
The SASLprep implementation depends on the UTF-8 functions from
src/backend/utils/mb/wchar.c. So to use it, you must also compile and link
that. That doesn't change anything for the current users of these
functions, the backend and libpq, as they both already link with wchar.o.
It would be good to move those functions into a separate file in
src/commmon, but I'll leave that for another day.
No documentation changes included, because there is no details on the
SCRAM mechanism in the docs anyway. An overview on that in the protocol
specification would probably be good, even though SCRAM is documented in
detail in RFC5802. I'll write that as a separate patch. An important thing
to mention there is that we apply SASLprep even on invalid UTF-8 strings,
to support other encodings.
Patch by Michael Paquier and me.
Discussion: https://www.postgresql.org/message-id/CAB7nPqSByyEmAVLtEf1KxTRh=PWNKiWKEKQR=e1yGehz=wbymQ@mail.gmail.com
Add a "copy" argument to make it optional to receive a copy of caller
tuple that is safe to use following a subsequent manipulating of
tuplesort's state. This is a performance optimization. Most existing
tuplesort_gettupleslot() callers are made to opt out of copying.
Existing callers that happen to rely on the validity of tuple memory
beyond subsequent manipulations of the tuplesort request their own
copy.
This brings tuplesort_gettupleslot() in line with
tuplestore_gettupleslot(). In the future, a "copy"
tuplesort_getdatum() argument may be added, that similarly allows
callers to opt out of receiving their own copy of tuple.
In passing, clarify assumptions that callers of other tuplesort fetch
routines may make about tuple memory validity, per gripe from Tom
Lane.
Author: Peter Geoghegan
Discussion: CAM3SWZQWZZ_N=DmmL7tKy_OUjGH_5mN=N=A6h7kHyyDvEhg2DA@mail.gmail.com
The original code was overly optimistic about the cost of scanning a
BRIN index, leading to BRIN indexes being selected when they'd be a
worse choice than some other index. This complete rewrite should be
more accurate.
Author: David Rowley, based on an earlier patch by Emre Hasegeli
Reviewed-by: Emre Hasegeli
Discussion: https://postgr.es/m/CAKJS1f9n-Wapop5Xz1dtGdpdqmzeGqQK4sV2MK-zZugfC14Xtw@mail.gmail.com
Also add opr_sanity check that all preloaded immutable functions are
parallel safe. (Per discussion, this does not necessarily have to be
true for all possible such functions, but deviations would be unlikely
enough that maintaining such a test is reasonable.)
Reported-by: David Rowley <david.rowley@2ndquadrant.com>
Reviewed-by: Robert Haas <robertmhaas@gmail.com>
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
It was not used for what the comment claimed, at all. It was actually used
as the 'base' argument to strtol(), when reading the iteration count. We
don't need a constant for base-10, so remove it.
This is the SQL standard-conforming variant of PostgreSQL's serial
columns. It fixes a few usability issues that serial columns have:
- CREATE TABLE / LIKE copies default but refers to same sequence
- cannot add/drop serialness with ALTER TABLE
- dropping default does not drop sequence
- need to grant separate privileges to sequence
- other slight weirdnesses because serial is some kind of special macro
Reviewed-by: Vitaly Burovoy <vitaly.burovoy@gmail.com>
For normal commits and aborts we already reset PgXact->xmin,
so we can simply avoid running SnapshotResetXmin() twice.
During performance tests by Alexander Korotkov, diagnosis
by Andres Freund showed PgXact array as a bottleneck. After
manual analysis by me of the code paths that touch those
memory locations, I was able to identify extraneous code
in the main transaction commit path.
Avoiding touching highly contented shmem improves concurrent
performance slightly on all workloads, confirmed by tests
run by Ashutosh Sharma and Alexander Korotkov.
Simon Riggs
Discussion: CANP8+jJdXE9b+b9F8CQT-LuxxO0PBCB-SZFfMVAdp+akqo4zfg@mail.gmail.com
HandleFunctionRequest() is no longer responsible for reading the protocol
message from the client, since commit 2b3a8b20c2. Fix the outdated
comments.
HandleFunctionRequest() now always returns 0, because the code that used
to return EOF was moved in 2b3a8b20c2. Therefore, the caller no longer
needs to check the return value.
Reported by Andres Freund. Backpatch to all supported versions, even though
this doesn't have any user-visible effect, to make backporting future
patches in this area easier.
Discussion: https://www.postgresql.org/message-id/20170405010525.rt5azbya5fkbhvrx@alap3.anarazel.de
Tom Lane