Coverity complained that one caller of getFormattedTypeName() failed
to free the returned string. Which is true, but rather than fixing
that one, let's get rid of this tedious and error-prone requirement.
Now that getFormattedTypeName() caches its result, strdup'ing that
result and expecting the caller to free it accomplishes little except
to waste cycles. We do create a leak in the case where getTypes didn't
make a TypeInfo for the type, but that basically shouldn't ever happen.
Back-patch, as commit 6c450a861 was. This isn't a particularly
interesting bug fix, but the API change seems like a hazard for
future back-patching activity if we don't back-patch it.
Replace fixed-length command buffers with psprintf() calls. We didn't
have anything as convenient as psprintf() when this code was written,
but now that we do, there's little reason for the limitation to
stand. Removing it eliminates some corner cases where (for example)
starting the postmaster with a whole lot of options fails.
Most individual file names that pg_ctl deals with are still restricted
to MAXPGPATH, but we've seldom had complaints about that limitation
so long as it only applies to one filename.
Back-patch to all supported branches.
Phil Krylov
Discussion: https://postgr.es/m/567e199c6b97ee19deee600311515b86@krylov.eu
For no particularly good reason, getPolicies() queried pg_policy
separately for each table. We can collect all the policies in
a single query instead, and attach them to the correct TableInfo
objects using findTableByOid() lookups. On the regression
database, this reduces the number of queries substantially, and
provides a visible savings even when running against a local
server.
Per complaint from Hubert Depesz Lubaczewski. Since this is such
a simple fix and can have a visible performance benefit, back-patch
to all supported branches.
Discussion: https://postgr.es/m/20210826084430.GA26282@depesz.com
There's long been a "TODO: there might be some value in caching
the results" annotation on pg_dump's getFormattedTypeName function;
but we hadn't gotten around to checking what it was costing us to
repetitively look up type names. It turns out that when dumping the
current regression database, about 10% of the total number of queries
issued are duplicative format_type() queries. However, Hubert Depesz
Lubaczewski reported a not-unusual case where these account for over
half of the queries issued by pg_dump. Individually these queries
aren't expensive, but when network lag is a factor, they add up to a
problem. We can very easily add some caching to getFormattedTypeName
to solve it.
Since this is such a simple fix and can have a visible performance
benefit, back-patch to all supported branches.
Discussion: https://postgr.es/m/20210826084430.GA26282@depesz.com
Strictly speaking this isn't a bug, but since all references to catalog
objects are schema-qualified, we might as well be consistent. The
omission first appeared in commit 1c5d9270e339, so backpatch to 12.
Author: Justin Pryzby <pryzbyj@telsasoft.com>
Discussion: https://postgr.es/m/20210827193151.GN26465@telsasoft.com
Add pg_resetxlog -u option to set the oldest xid in pg_control.
Previously -x set this value be -2 billion less than the -x value.
However, this causes the server to immediately scan all relation's
relfrozenxid so it can advance pg_control's oldest xid to be inside the
autovacuum_freeze_max_age range, which is inefficient and might disrupt
diagnostic recovery. pg_upgrade will use this option to better create
the new cluster to match the old cluster.
Reported-by: Jason Harvey, Floris Van Nee
Discussion: https://postgr.es/m/20190615183759.GB239428@rfd.leadboat.com, 87da83168c644fd9aae38f546cc70295@opammb0562.comp.optiver.com
Author: Bertrand Drouvot
Backpatch-through: 9.6
These have been introduced by 7fbe0c8, and could happen for
pg_basebackup and pg_receivewal.
Per report from Coverity for the ones in walmethods.c, I have spotted
the ones in receivelog.c after more review.
Backpatch-through: 10
The logic handling the opening of new WAL segments was fuzzy when using
--compress if a partial, non-compressed, segment with the same base name
existed in the repository storing those files. In this case, using
--compress would cause the code to first check for the existence and the
size of a non-compressed segment, followed by the opening of a new
compressed, partial, segment. The code was accidentally working
correctly on most platforms as the buildfarm has proved, except
bowerbird where gzflush() could fail in this code path. It is wrong
anyway to take the code path used pre-padding when creating a new
partial, non-compressed, segment, so let's fix it.
Note that this issue exists when users mix successive runs of
pg_receivewal with or without compression, as discovered with the tests
introduced by ffc9dda.
While on it, this refactors the code so as code paths that need to know
about the ".gz" suffix are down from four to one in walmethods.c, easing
a bit the introduction of new compression methods. This addresses a
second issue where log messages generated for an unexpected failure
would not show the compressed segment name involved, which was
confusing, printing instead the name of the non-compressed equivalent.
Reported-by: Georgios Kokolatos
Discussion: https://postgr.es/m/YPDLz2x3o1aX2wRh@paquier.xyz
Backpatch-through: 10
pg_dump failed to preserve the 'enabled' flag (which can be not only
disabled, but also REPLICA or ALWAYS) for partitions which had it
changed from their respective parents. Attempt to handle that by
including a definition for such triggers in the dump, but replace the
standard CREATE TRIGGER line with an ALTER TRIGGER line.
Backpatch to 11, where these triggers can exist. In branches 11 and 12,
pick up a few test lines from commit b9b408c48724 to verify that
pg_upgrade is okay with these arrangements.
Co-authored-by: Justin Pryzby <pryzby@telsasoft.com>
Co-authored-by: Álvaro Herrera <alvherre@alvh.no-ip.org>
Discussion: https://postgr.es/m/20200930223450.GA14848@telsasoft.com
Instead use the common/int.h functions to check for integer overflow
in a more C-standard-compliant fashion. This is motivated by recent
failures on buildfarm member moonjelly, where it appears that
development-tip gcc is optimizing without regard to the -fwrapv
switch. Presumably that's a gcc bug that will be fixed soon, but
we might as well install cleaner coding here rather than wait.
(This does not address the question of whether we'll ever be able
to get rid of using -fwrapv. Testing shows that this spot is the
only place where doing so creates visible regression test failures,
but unfortunately that proves very little.)
Back-patch to v12. The common/int.h functions exist in v11, but
that branch doesn't use them in any client-side code. I judge
that this case isn't interesting enough in the real world to take
even a small risk of issues from being the first such use.
Tom Lane and Fabien Coelho
Discussion: https://postgr.es/m/73927.1624815543@sss.pgh.pa.us
002_pgbench_no_server was printing some array pointers instead of the
actual contents of those arrays for the expected outputs of stdout and
stderr for a tested command. This does not add any new information that
can help with debugging as the test names allow to track failure
locations, if any.
This commit simply removes those logs as the rest of the printed
information is redundant with command_checks_all().
Per discussion with Andrew Dunstan and Álvaro Herrera.
Discussion: https://postgr.es/m/YNXNFaG7IgkzZanD@paquier.xyz
Backpatch-through: 11
This fixes a couple of problems within the so-said code of this commit
subject:
- Replace the use of open() with slurp_file(), fixing an issue reported
by buildfarm member fairywren whose perl installation keep around CRLF
characters, causing the matching patterns for the logs to fail.
- Remove the eval block, which is not really necessary.
This set of issues has come into light after fixing a different issue
with c13585fe, and this is wrong since this code has been introduced.
Reported-by: Andrew Dunstan, and buildfarm member fairywren
Author: Michael Paquier
Reviewed-by: Andrew Dunstan
Discussion: https://postgr.es/m/0f49303e-7784-b3ee-200b-cbf67be2eb9e@dunslane.net
Backpatch-through: 11
The logic checking for the format of per-thread logs used grep() with
directly "$re", which would cause the test to consider all the logs as
a match without caring about their format at all. Using "/$re/" makes
grep() perform a regex test, which is what we want here.
While on it, improve some of the tests to be more picky with the
patterns expected and add more comments to describe the tests.
Issue discovered while digging into a separate patch.
Author: Fabien Coelho, Michael Paquier
Discussion: https://postgr.es/m/YNPsPAUoVDCpPOGk@paquier.xyz
Backpatch-through: 11
Recent glibc versions have made mktime() fail if tm_isdst is
inconsistent with the prevailing timezone; in particular it fails for
tm_isdst = 1 when the zone is UTC. (This seems wildly inconsistent
with the POSIX-mandated treatment of "incorrect" values for the other
fields of struct tm, so if you ask me it's a bug, but I bet they'll
say it's intentional.) This has been observed to cause cosmetic
problems when pg_restore'ing an archive created in a different
timezone.
To fix, do mktime() using the field values from the archive, and if
that fails try again with tm_isdst = -1. This will give a result
that's off by the UTC-offset difference from the original zone, but
that was true before, too. It's not terribly critical since we don't
do anything with the result except possibly print it. (Someday we
should flush this entire bit of logic and record a standard-format
timestamp in the archive instead. That's not okay for a back-patched
bug fix, though.)
Also, guard our only other use of mktime() by having initdb's
build_time_t() set tm_isdst = -1 not 0. This case could only have
an issue in zones that are DST year-round; but I think some do exist,
or could in future.
Per report from Wells Oliver. Back-patch to all supported
versions, since any of them might need to run with a newer glibc.
Discussion: https://postgr.es/m/CAOC+FBWDhDHO7G-i1_n_hjRzCnUeFO+H-Czi1y10mFhRWpBrew@mail.gmail.com
The set of subcommands supported by \dAp, \do and \dy was described
incorrectly in psql's --help. The documentation was already consistent
with the code.
Reported-by: inoas, from IRC
Author: Matthijs van der Vleuten
Reviewed-by: Neil Chen
Discussion: https://postgr.es/m/6a984e24-2171-4039-9050-92d55e7b23fe@www.fastmail.com
Backpatch-through: 9.6
An incorrectly-encoded multibyte character near the end of a string
could cause various processing loops to run past the string's
terminating NUL, with results ranging from no detectable issue to
a program crash, depending on what happens to be in the following
memory.
This isn't an issue in the server, because we take care to verify
the encoding of strings before doing any interesting processing
on them. However, that lack of care leaked into client-side code
which shouldn't assume that anyone has validated the encoding of
its input.
Although this is certainly a bug worth fixing, the PG security team
elected not to regard it as a security issue, primarily because
any untrusted text should be sanitized by PQescapeLiteral or
the like before being incorporated into a SQL or psql command.
(If an app fails to do so, the same technique can be used to
cause SQL injection, with probably much more dire consequences
than a mere client-program crash.) Those functions were already
made proof against this class of problem, cf CVE-2006-2313.
To fix, invent PQmblenBounded() which is like PQmblen() except it
won't return more than the number of bytes remaining in the string.
In HEAD we can make this a new libpq function, as PQmblen() is.
It seems imprudent to change libpq's API in stable branches though,
so in the back branches define PQmblenBounded as a macro in the files
that need it. (Note that just changing PQmblen's behavior would not
be a good idea; notably, it would completely break the escaping
functions' defense against this exact problem. So we just want a
version for those callers that don't have any better way of handling
this issue.)
Per private report from houjingyi. Back-patch to all supported branches.
"typename" is a C++ keyword, so pg_upgrade.h fails to compile in C++.
Fortunately, there seems no likely reason for somebody to need to
do that. Nonetheless, it's project policy that all .h files should
pass cpluspluscheck, so rename the argument to fix that.
Oversight in 57c081de0; back-patch as that was. (The policy requiring
pg_upgrade.h to pass cpluspluscheck only goes back to v12, but it
seems best to keep this code looking the same in all branches.)
Commits 29aeda6e4 et al closed up some oversights involving not checking
for non-upgradable types within container types, such as arrays and
ranges. However, I only looked at version.c, failing to notice that
there were substantially-equivalent tests in check.c. (The division
of responsibility between those files is less than clear...)
In addition, because genbki.pl does not guarantee that auto-generated
rowtype OIDs will hold still across versions, we need to consider that
the composite type associated with a system catalog or view is
non-upgradable. It seems unlikely that someone would have a user
column declared that way, but if they did, trying to read it in another
PG version would likely draw "no such pg_type OID" failures, thanks
to the type OID embedded in composite Datums.
To support the composite and reg*-type cases, extend the recursive
query that does the search to allow any base query that returns
a column of pg_type OIDs, rather than limiting it to exactly one
starting type.
As before, back-patch to all supported branches.
Discussion: https://postgr.es/m/2798740.1619622555@sss.pgh.pa.us
pg_checksums uses two counters, total size and current size,
to calculate the progress. Previously the progress that
pg_checksums reported could not reach 100% at the end.
The cause of this issue was that the sizes of only pages excluding
new ones in each file were counted as the current size
while the size of each file is counted as the total size.
That is, the total size of all new pages could be reported
as the difference between the total size and current size.
This commit fixes this issue by making pg_checksums count
the sizes of all pages including new ones in each file as
the current size.
Back-patch to v12 where progress reporting was added to pg_checksums.
Reported-by: Shinya Kato
Author: Shinya Kato
Reviewed-by: Fujii Masao
Discussion: https://postgr.es/m/TYAPR01MB289656B1ACA0A5E7CAD07BE3C47A9@TYAPR01MB2896.jpnprd01.prod.outlook.com
Despite the clear comments pointing out that the duplicative code
segments in ReadHead() and _discoverArchiveFormat() needed to be
in sync, they were not: the latter did not bother to apply any of
the sanity checks in the former. We'd missed noticing this partly
because none of those checks would fail in scenarios we customarily
test, and partly because the oversight would be masked if both
segments execute, which they would in cases other than needing to
autodetect the format of a non-seekable stdin source. However,
in a case meeting all these requirements --- for example, trying
to read a newer-than-supported archive format from non-seekable
stdin --- pg_restore missed applying the version check and would
likely dump core or otherwise misbehave.
The whole thing is silly anyway, because there seems little reason
to duplicate the logic beyond the one-line verification that the
file starts with "PGDMP". There seems to have been an undocumented
assumption that multiple major formats (major enough to require
separate reader modules) would nonetheless share the first half-dozen
fields of the custom-format header. This seems unlikely, so let's
fix it by just nuking the duplicate logic in _discoverArchiveFormat().
Also get rid of the pointless attempt to seek back to the start of
the file after successful autodetection. That wastes cycles and
it means we have four behaviors to verify not two.
Per bug #16951 from Sergey Koposov. This has been broken for
decades, so back-patch to all supported versions.
Discussion: https://postgr.es/m/16951-a4dd68cf0de23048@postgresql.org
Jasen Betts reported yet another unintended side effect of commit
85c54287a: reconnecting with "\c service=whatever" did not have the
expected results. The reason is that starting from the output of
PQconndefaults() effectively allows environment variables (such
as PGPORT) to override entries in the service file, whereas the
normal priority is the other way around.
Not using PQconndefaults at all would require yet a third main code
path in do_connect's parameter setup, so I don't really want to fix
it that way. But we can have the logic effectively ignore all the
default values for just a couple more lines of code.
This patch doesn't change the behavior for "\c -reuse-previous=on
service=whatever". That remains significantly different from before
85c54287a, because many more parameters will be re-used, and thus
not be possible for service entries to replace. But I think this
is (mostly?) intentional. In any case, since libpq does not report
where it got parameter values from, it's hard to do differently.
Per bug #16936 from Jasen Betts. As with the previous patches,
back-patch to all supported branches. (9.5 is unfortunately now
out of support, so this won't get fixed there.)
Discussion: https://postgr.es/m/16936-3f524322a53a29f0@postgresql.org
pg_waldump --stats=record identifies a record by a combination
of the RmgrId and the four bits of the xl_info field of the record.
But XACT records use the first bit of those four bits for an optional
flag variable, and the following three bits for the opcode to
identify a record. So previously the same type of XACT record
could have different four bits (three bits are the same but the
first one bit is different), and which could cause
pg_waldump --stats=record to show two lines of per-record statistics
for the same XACT record. This is a bug.
This commit changes pg_waldump --stats=record so that it processes
only XACT record differently, i.e., filters the opcode out of xl_info
and uses a combination of the RmgrId and those three bits as
the identifier of a record, only for XACT record. For other records,
the four bits of the xl_info field are still used.
Back-patch to all supported branches.
Author: Kyotaro Horiguchi
Reviewed-by: Shinya Kato, Fujii Masao
Discussion: https://postgr.es/m/2020100913412132258847@highgo.ca
psql's editing commands decide whether the user has edited the file
by checking for change of modification timestamp. This is probably
fine for a pre-existing file, but with a temporary file that is
created within the command, it's possible for a fast typist to
save-and-exit in less than the one-second granularity of stat(2)
timestamps. On Windows FAT filesystems the granularity is even
worse, 2 seconds, making the race a bit easier to hit.
To fix, try to set the temp file's mod time to be two seconds ago.
It's unlikely this would fail, but then again the race condition
itself is unlikely, so just ignore any error.
Also, we might as well check the file size as well as its mod time.
While this is a difficult bug to hit, it still seems worth
back-patching, to ensure that users' edits aren't lost.
Laurenz Albe, per gripe from Jacob Champion; based on fix suggestions
from Jacob and myself
Discussion: https://postgr.es/m/0ba3f2a658bac6546d9934ab6ba63a805d46a49b.camel@cybertec.at
The same test for REINDEX (VERBOSE) was done twice, while it is clear
that the second test should use --concurrently. Issue introduced in
5dc92b8, for what looks like a copy-paste mistake.
Reviewed-by: Mark Dilger
Discussion: https://postgr.es/m/A7AE97EA-F4B0-4CAB-8FFF-3FECD31F9D63@enterprisedb.com
Backpatch-through: 12
When ON_ERROR_ROLLBACK is enabled, psql releases a temporary savepoint
if it's idle in a valid transaction block after executing a query. But psql
doesn't do that after RELEASE or ROLLBACK is executed because a temporary
savepoint has already been destroyed in that case.
This commit changes psql's ON_ERROR_ROLLBACK so that it doesn't release
a temporary savepoint also when COMMIT AND CHAIN is executed. A temporary
savepoint doesn't need to be released in that case because
COMMIT AND CHAIN also destroys any savepoints defined within the transaction
to commit. Otherwise psql tries to release the savepoint that
COMMIT AND CHAIN has already destroyed and cause an error
"ERROR: savepoint "pg_psql_temporary_savepoint" does not exist".
Back-patch to v12 where transaction chaining was added.
Reported-by: Arthur Nascimento
Author: Arthur Nascimento
Reviewed-by: Fujii Masao, Vik Fearing
Discussion: https://postgr.es/m/16867-3475744069228158@postgresql.org
Generation expressions of generated columns are always inherited, so
there is no need to set them separately in child tables, and there is
no syntax to do so either. The code previously used the code paths
for the handling of default values, for which different rules apply;
in particular it might want to set a default value explicitly for an
inherited column. This resulted in unrestorable dumps. For generated
columns, just skip them in inherited tables.
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://www.postgresql.org/message-id/flat/15830.1575468847%40sss.pgh.pa.us
When reporting connection errors, we might show a database name in the
message that's not the one we actually tried to connect to, if the
database was taken from libpq defaults instead of from user parameters.
Fix such error messages to use PQdb(), which reports the correct name.
(But, per commit 2930c05634bc, make sure not to try to print NULL.)
Apply to branches 9.5 through 13. Branch master has already been
changed differently by commit 58cd8dca3de0.
Reported-by: Robert Haas <robertmhaas@gmail.com>
Discussion: https://postgr.es/m/CA+TgmobssJ6rS22dspWnu-oDxXevGmhMD8VcRBjmj-b9UDqRjw@mail.gmail.com
The loops to identify word boundaries could access past the end of
the input string. Likely that would never result in an actual
crash, but it makes valgrind unhappy.
The logic to try different numbers of words didn't work when the
input has two words but we only have a match to the first, eg
"\h with select". (We must "continue" the pass loop, not "break".)
The logic to compute nl_count was bizarrely managed, and in at
least two code paths could end up calling PageOutput with
nl_count = 0, resulting in failing to paginate output that should
have been fed to the pager. Also, in v12 and up, the nl_count
calculation hadn't been updated to account for the addition of a URL.
The PQExpBuffer holding the command syntax details wasn't freed,
resulting in a session-lifespan memory leak.
While here, improve some comments, choose a more descriptive name
for a variable, fix inconsistent datatype choice for another variable.
Per bug #16837 from Alexander Lakhin. This code is very old,
so back-patch to all supported branches.
Kyotaro Horiguchi and Tom Lane
Discussion: https://postgr.es/m/16837-479bcd56040c71b3@postgresql.org
The context is an object that no longer bears some aclitem that it bore
initially. (A user issued REVOKE or GRANT statements upon the object.)
pg_dump is forming SQL to reproduce the object ACL. Since initdb
creates no ACL bearing GRANT OPTION, reaching this bug requires an
extension where the creation script establishes such an ACL. No PGXN
extension does that. If an installation did reach the bug, pg_dump
would have omitted a semicolon, causing a REVOKE and the next SQL
statement to fail. Separately, since the affected code exists to
eliminate an entire aclitem, it wants plain REVOKE, not REVOKE GRANT
OPTION FOR. Back-patch to 9.6, where commit
23f34fa4ba358671adab16773e79c17c92cbc870 first appeared.
Discussion: https://postgr.es/m/20210109102423.GA160022@rfd.leadboat.com
This is the same fix as commit 9eabfe300 applied to INDEX ATTACH
entries, but for table-to-publication attachments. As in that
case, even though the backend doesn't record "ownership" of the
attachment, we still ought to label it in the dump archive with
the role name that should run the ALTER PUBLICATION command.
The existing behavior causes the ALTER to be done by the original
role that started the restore; that will usually work fine, but
there may be corner cases where it fails.
The bulk of the patch is concerned with changing struct
PublicationRelInfo to include a pointer to the associated
PublicationInfo object, so that we can get the owner's name
out of that when the time comes. While at it, I rewrote
getPublicationTables() to do just one query of pg_publication_rel,
not one per table.
Back-patch to v10 where this code was introduced.
Discussion: https://postgr.es/m/1165710.1610473242@sss.pgh.pa.us
The point of this restriction is to avoid trying to substitute variables
into timestamp literal values, which may contain strings like '12:34'.
There is a good deal more that should be done to reduce pgbench's
tendency to substitute where it shouldn't. But this is sufficient to
solve the case complained of by Jaime Soler, and it's simple enough
to back-patch.
Back-patch to v11; before commit 9d36a3866, pgbench had a slightly
different definition of what a variable name is, and anyway it seems
unwise to change long-stable branches for this.
Fabien Coelho
Discussion: https://postgr.es/m/alpine.DEB.2.22.394.2006291740420.805678@pseudo
Although a partitioned index's attachment to its parent doesn't
have separate ownership, the ArchiveEntry for it needs to be
marked with an owner anyway, to ensure that the ALTER command
is run by the appropriate role when restoring with
--use-set-session-authorization. Without this, the ALTER will
be run by the role that started the restore session, which will
usually work but it's formally the wrong thing.
Back-patch to v11 where this type of ArchiveEntry was added.
In HEAD, add equivalent commentary to the just-added TABLE ATTACH
case, which I'd made do the right thing already.
Discussion: https://postgr.es/m/1094034.1610418498@sss.pgh.pa.us
We found last February that the error-case tests added by commit
008cf0409 failed on OpenBSD, because that platform doesn't really
check locale names. At the time it seemed that that was only an issue
for LC_CTYPE, but testing on a more recent version of OpenBSD shows
that it's now equally lax about LC_COLLATE.
Rather than dropping the LC_COLLATE test too, put back LC_CTYPE
(reverting c4b0edb07), and adjust these tests to accept the different
error message that we get if setlocale() doesn't reject a bogus locale
name. The point of these tests is not really what the backend does
with the locale name, but to show that createdb quotes funny locale
names safely; so we're not losing test reliability this way.
Back-patch as appropriate.
Discussion: https://postgr.es/m/231373.1610058324@sss.pgh.pa.us
Buildfarm members pogona and petalura have shown a failure when
pg_ctl/t/004_logrotate.pl starts just before local midnight.
The default rotate-at-midnight behavior occurs just before the
Perl script examines current_logfiles, so it figures that the
rotation it's already requested has occurred ... but in reality,
that rotation happens just after it looks, so the expected new
log data goes into a different file than the one it's examining.
In HEAD, src/test/kerberos/t/001_auth.pl has acquired similar code
that evidently has a related failure mode. Besides being quite new,
few buildfarm critters run that test, so it's unsurprising that
we've not yet seen a failure there.
Fix both cases by setting log_rotation_age = 0 so that no time-based
rotation can occur. Also absorb 004_logrotate.pl's decision to
set lc_messages = 'C' into the kerberos test, in hopes that it will
work in non-English prevailing locales.
Report: https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=pogona&dt=2020-12-24%2022%3A10%3A04
Report: https://buildfarm.postgresql.org/cgi-bin/show_log.pl?nm=petalura&dt=2020-02-01%2022%3A20%3A04
pg_rewind looks at the control file to check what timeline a server is on.
But promotion doesn't immediately write a checkpoint, it merely writes
an end-of-recovery WAL record. If pg_rewind runs immediately after
promotion, before the checkpoint has completed, it will think think that
the server is still on the earlier timeline. We ran into this issue a long
time ago already, see commit 484a848a73f.
It's a bit bogus that pg_rewind doesn't determine the timeline correctly
until the end-of-recovery checkpoint has completed. We probably should
fix that. But for now work around it by waiting for the checkpoint
to complete before running pg_rewind, like we did in commit 484a848a73f.
In the passing, tidy up the new test a little bit. Rerder the INSERTs so
that the comments make more sense, remove a spurious CHECKPOINT call after
pg_rewind has already run, and add --debug option, so that if this fails
again, we'll have more data.
Per buildfarm failure at https://buildfarm.postgresql.org/cgi-bin/show_stage_log.pl?nm=rorqual&dt=2020-12-06%2018%3A32%3A19&stg=pg_rewind-check.
Backpatch to all supported versions.
Discussion: https://www.postgresql.org/message-id/1713707e-e318-761c-d287-5b6a4aa807e8@iki.fi
Buildfarm has been failing sporadically on the new test. I was able to
reproduce this by adding a random 0-10 s delay in the walreceiver, just
before it connects to the primary. There's a race condition where node_3
is promoted before it has fully caught up with node_1, leading to diverged
timelines. When node_1 is later reconfigured as standby following node_3,
it fails to catch up:
LOG: primary server contains no more WAL on requested timeline 1
LOG: new timeline 2 forked off current database system timeline 1 before current recovery point 0/30000A0
That's the situation where you'd need to use pg_rewind, but in this case
it happens already when we are just setting up the actual pg_rewind
scenario we want to test, so change the test so that it waits until
node_3 is connected and fully caught up before promoting it, so that you
get a clean, controlled failover.
Also rewrite some of the comments, for clarity. The existing comments
detailed what each step in the test did, but didn't give a good overview
of the situation the steps were trying to create.
For reasons I don't understand, the test setup had to be written slightly
differently in 9.6 and 9.5 than in later versions. The 9.5/9.6 version
needed node 1 to be reinitialized from backup, whereas in later versions
it could be shut down and reconfigured to be a standby. But even 9.5 should
support "clean switchover", where primary makes sure that pending WAL is
replicated to standby on shutdown. It would be nice to figure out what's
going on there, but that's independent of pg_rewind and the scenario that
this test tests.
Discussion: https://www.postgresql.org/message-id/b0a3b95b-82d2-6089-6892-40570f8c5e60%40iki.fi
If the target is a standby server, its WAL doesn't end at the last
checkpoint record, but at minRecoveryPoint. We must scan all the
WAL from the last common checkpoint all the way up to minRecoveryPoint
for modified pages, and also consider that portion when determining
whether the server needs rewinding.
Backpatch to all supported versions.
Author: Ian Barwick and me
Discussion: https://www.postgresql.org/message-id/CABvVfJU-LDWvoz4-Yow3Ay5LZYTuPD7eSjjE4kGyNZpXC6FrVQ%40mail.gmail.com
