This avoids possibly selecting Python 2.x on systems that have
both Python 2 and Python 3. We used to feel that what "python"
links to is a user choice that we should honor. However, we're
about to cease support for Python 2, so users will no longer have
any choice of that sort. This small change is being made ahead
of the big Python-2-ectomy so that we can see how much of the
buildfarm is not yet prepared for that. Systems with only
Python 2 will continue to build that way, for now.
Discussion: https://postgr.es/m/2872c9a0-4b0a-1354-d5f6-94d6f85ba354@enterprisedb.com
In adverse circumstances, the fsync calls cause this test to run for
quite a long time (multiple minutes) and even suffer timeout failures.
This seems to date from before we made an effort to disable fsync in
all our test cases; there's not a lot of point in using it if there's
not a plan to force an O/S crash during the test.
Discussion: https://postgr.es/m/440239.1642560607@sss.pgh.pa.us
The base Makefile will output help to the user when invoking make in
an unconfigured tree, the help was however always referring to a file
which may not be present as it's only in tarballs. Dynamically check
for the presence of the INSTALL file and fall back on README.git when
it's not available (which is the case of Git checkouts).
Reported-by: Tim McNamara <tim@mcnamara.nz>
Reviewed-by: Magnus Hagander <magnus@hagander.net>
Reviewed-by: Peter Eisentraut <peter.eisentraut@enterprisedb.com>
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://postgr.es/m/730dae39-abaa-4140-893b-95d732fed003@www.fastmail.com
bbsink_copystream wants to store a type byte just before the buffer,
but basebackup.c wants the buffer to be aligned so that it can call
PageIsNew() and PageGetLSN() on it. Therefore, instead of inserting
1 extra byte before the buffer, insert MAXIMUM_ALIGNOF extra bytes
and only use the last one.
On most machines this doesn't cause any problem (except perhaps for
performance) but some buildfarm machines with -fsanitize=alignment
dump core.
Discussion: http://postgr.es/m/CA+TgmoYx5=1A2K9JYV-9zdhyokU4KKTyNQ9q7CiXrX=YBBMWVw@mail.gmail.com
If connectivity to the server has been lost or become flaky, the
user might well try to send a query cancel. It's highly annoying
if PQcancel hangs up in such a case, but that's exactly what's likely
to happen. To ameliorate this problem, apply the PGconn's
tcp_user_timeout and keepalives settings to the TCP connection used
to send the cancel. This should be safe on Unix machines, since POSIX
specifies that setsockopt() is async-signal-safe. We are guessing
that WSAIoctl(SIO_KEEPALIVE_VALS) is similarly safe on Windows.
(Note that at least in psql and our other frontend programs, there's
no safety issue involved anyway, since we run PQcancel in its own
thread rather than in a signal handler.)
Most of the value here comes from the expectation that tcp_user_timeout
will be applied as a connection timeout. That appears to happen on
Linux, even though its tcp(7) man page claims differently. The
keepalive options probably won't help much, but as long as we can
apply them for not much code, we might as well.
Jelte Fennema, reviewed by Fujii Masao and myself
Discussion: https://postgr.es/m/AM5PR83MB017870DE81FC84D5E21E9D1EF7AA9@AM5PR83MB0178.EURPRD83.prod.outlook.com
In the new approach, all files across all tablespaces are sent in a
single COPY OUT operation. The CopyData messages are no longer raw
archive content; rather, each message is prefixed with a type byte
that describes its purpose, e.g. 'n' signifies the start of a new
archive and 'd' signifies archive or manifest data. This protocol
is significantly more extensible than the old approach, since we can
later create more message types, though not without concern for
backward compatibility.
The new protocol sends a few things to the client that the old one
did not. First, it sends the name of each archive explicitly, instead
of letting the client compute it. This is intended to make it easier
to write future patches that might send archives in a format other
that tar (e.g. cpio, pax, tar.gz). Second, it sends explicit progress
messages rather than allowing the client to assume that progress is
defined by the number of bytes received. This will help with future
features where the server compresses the data, or sends it someplace
directly rather than transmitting it to the client.
The old protocol is still supported for compatibility with previous
releases. The new protocol is selected by means of a new
TARGET option to the BASE_BACKUP command. Currently, the
only supported target is 'client'. Support for additional
targets will be added in a later commit.
Patch by me. The patch set of which this is a part has had review
and/or testing from Jeevan Ladhe, Tushar Ahuja, Suraj Kharage,
Dipesh Pandit, and Mark Dilger.
Discussion: http://postgr.es/m/CA+TgmoaYZbz0=Yk797aOJwkGJC-LK3iXn+wzzMx7KdwNpZhS5g@mail.gmail.com
Where we test vacuum_truncate's effects, sometimes this is failing to
truncate as expected on the build farm. That could be explained by page
skipping, so disable it explicitly, with the theory that commit fe246d1c
didn't go far enough.
Back-patch to 12, where the vacuum_truncate tests were added.
Discussion: https://postgr.es/m/CA%2BhUKGLT2UL5_JhmBzUgkdyKfc%3D5J-gJSQJLysMs4rqLUKLAzw%40mail.gmail.com
This reverts commit e0e567a106.
On various platforms, the new approach using the sysconfig module
reported incorrect values for the include directory, and so any
Python-related compilations failed. Revert for now and revisit later.
The tests added by 14d755b000 added a
test case for psql's \set ECHO errors. After the test, it then reset
this to \set ECHO none, which is the default. But the regression
tests are actually run under \set ECHO all (psql -a), so that would
have been the correct way to restore the previous state. Otherwise,
test cases added after that point would not have their input lines
displayed. This was never the intention, so fix this now.
With Python 3.10, configure spits out warnings about the module
distutils.sysconfig being deprecated and scheduled for removal in
Python 3.12. Change the uses in configure to use the module sysconfig
instead. The logic stays the same.
Note that sysconfig exists since Python 2.7, so this moves the minimum
required version up from Python 2.6.
Discussion: https://www.postgresql.org/message-id/flat/c74add3c-09c4-a9dd-1a03-a846e5b2fc52%40enterprisedb.com
The previous coding was correct, but the style and commentary were a bit
vague about which operations had to happen, in what circumstances, and
in what order. Rearrange so that the epilogue does nothing in the DFA END
state. That allows turning some conditional statements in the backtracking
logic into asserts. With that, we can be more explicit about needing
to backtrack at least one byte in non-END states to ensure checking the
current byte sequence in the slow path. No change to the regression tests,
since they should be able catch deficiencies here already.
In passing, improve the comments around DFA states where the first
continuation byte has a restricted range.
The original coding (from c33869cc3) failed with "more than one row
returned by a subquery used as an expression" if there were unrelated
triggers of the same tgname on parent partitioned tables. (That's
possible because statement-level triggers don't get inherited.) Fix
by applying LIMIT 1 after sorting the candidates by inheritance level.
Also, wrap the subquery in a CASE so that we don't have to execute it at
all when the trigger is visibly non-inherited. Aside from saving some
cycles, this avoids the need for a confusing and undocumented NULLIF().
While here, tweak the format of the emitted query to look a bit
nicer for "psql -E", and add some explanation of this subquery,
because it badly needs it.
Report and patch by Justin Pryzby (with some editing by me).
Back-patch to v13 where the faulty code came in.
Discussion: https://postgr.es/m/20211217154356.GJ17618@telsasoft.com
Most tests invoking pg_basebackup themselves did not yet use -cfast, which
makes pg_basebackup take considerably longer. The only reason this didn't
cause the tests to take many minutes is that spread checkpoints only throttle
when writing out a buffer and there aren't that many dirty buffers in the
tests...
Discussion: https://postgr.es/m/20220117195711.xx4qbxutrrlmo2dg@alap3.anarazel.de
BufferGetBlockNumber() is not that cheap and obviously cannot change during
one heap_prune_page(), so only call it once. We might be able to do better and
pass the block number from the caller, but that'd be a larger change...
Discussion: https://postgr.es/m/20211211045710.ljtuu4gfloh754rs@alap3.anarazel.de
Currently, database OIDs, relfilenodes, and tablespace OIDs can all
change when a cluster is upgraded using pg_upgrade. It seems better
to preserve them, because (1) it makes troubleshooting pg_upgrade
easier, since you don't have to do a lot of work to match up files
in the old and new clusters, (2) it allows 'rsync' to save bandwidth
when used to re-sync a cluster after an upgrade, and (3) if we ever
encrypt or sign blocks, we would likely want to use a nonce that
depends on these values.
This patch only arranges to preserve relfilenodes and tablespace
OIDs. The task of preserving database OIDs is left for another patch,
since it involves some complexities that don't exist in these cases.
Database OIDs have a similar issue, but there are some tricky points
in that case that do not apply to these cases, so that problem is left
for another patch.
Shruthi KC, based on an earlier patch from Antonin Houska, reviewed
and with some adjustments by me.
Discussion: http://postgr.es/m/CA+TgmoYgTwYcUmB=e8+hRHOFA0kkS6Kde85+UNdon6q7bt1niQ@mail.gmail.com
It seems highly unlikely that gettext() can be relied on to be
async-signal-safe. psql used to understand that, but someone got
it wrong long ago in the src/bin/scripts/ version of handle_sigint,
and then the bad idea was perpetuated when those two versions were
unified into src/fe_utils/cancel.c.
I'm unsure why there have not been field complaints about this
... maybe gettext() is signal-safe once it's translated at least
one message? But we have no business assuming any such thing.
In cancel.c (v13 and up), I preserved our ability to localize
"Cancel request sent" messages by invoking gettext() before
the signal handler is set up. In earlier branches I just made
src/bin/scripts/ not localize those messages, as psql did then.
(Just for extra unsafety, the src/bin/scripts/ version was
invoking fprintf() from a signal handler. Sigh.)
Noted while fixing signal-safety issues in PQcancel() itself.
Back-patch to all supported branches.
Discussion: https://postgr.es/m/2937814.1641960929@sss.pgh.pa.us
PQcancel() is supposed to be safe to call from a signal handler,
and indeed psql uses it that way. All of the library functions
it uses are specified to be async-signal-safe by POSIX ...
except for strerror. Neither plain strerror nor strerror_r
are considered safe. When this code was written, back in the
dark ages, we probably figured "oh, strerror will just index
into a constant array of strings" ... but in any locale except C,
that's unlikely to be true. Probably the reason we've not heard
complaints is that (a) this error-handling code is unlikely to be
reached in normal use, and (b) in many scenarios, localized error
strings would already have been loaded, after which maybe it's
safe to call strerror here. Still, this is clearly unacceptable.
The best we can do without relying on strerror is to print the
decimal value of errno, so make it do that instead. (This is
probably not much loss of user-friendliness, given that it is
hard to get a failure here.)
Back-patch to all supported branches.
Discussion: https://postgr.es/m/2937814.1641960929@sss.pgh.pa.us
The token in nodeTokenType() is actually the whole rest of the string,
so we need to take into account the length to do the correct
comparison.
Without this, postgres_fdw tests fail under
-DWRITE_READ_PARSE_PLAN_TREES.
Before, SQL-level boolean constants were represented by a string with
a cast, and internal Boolean values in DDL commands were usually
represented by Integer nodes. This takes the place of both of these
uses, making the intent clearer and having some amount of type safety.
Reviewed-by: Pavel Stehule <pavel.stehule@gmail.com>
Discussion: https://www.postgresql.org/message-id/flat/8c1a2e37-c68d-703c-5a83-7a6077f4f997@enterprisedb.com
The logic is similar to default_tablespace in some ways, so as no SET
queries on default_table_access_method are generated before dumping or
restoring an object (table or materialized view support table AMs) when
specifying this new option.
This option is useful to enforce the use of a default access method even
if some tables included in a dump use an AM different than the system's
default.
There are already two cases in the TAP tests of pg_dump with a table and
a materialized view that use a non-default table AM, and these are
extended that the new option does not generate SET clauses on
default_table_access_method.
Author: Justin Pryzby
Discussion: https://postgr.es/m/20211207153930.GR17618@telsasoft.com
"jsonlog" is a new value that can be added to log_destination to provide
logs in the JSON format, with its output written to a file, making it
the third type of destination of this kind, after "stderr" and
"csvlog". The format is convenient to feed logs to other applications.
There is also a plugin external to core that provided this feature using
the hook in elog.c, but this had to overwrite the output of "stderr" to
work, so being able to do both at the same time was not possible. The
files generated by this log format are suffixed with ".json", and use
the same rotation policies as the other two formats depending on the
backend configuration.
This takes advantage of the refactoring work done previously in ac7c807,
bed6ed3, 8b76f89 and 2d77d83 for the backend parts, and 72b76f7 for the
TAP tests, making the addition of any new file-based format rather
straight-forward.
The documentation is updated to list all the keys and the values that
can exist in this new format. pg_current_logfile() also required a
refresh for the new option.
Author: Sehrope Sarkuni, Michael Paquier
Reviewed-by: Nathan Bossart, Justin Pryzby
Discussion: https://postgr.es/m/CAH7T-aqswBM6JWe4pDehi1uOiufqe06DJWaU5=X7dDLyqUExHg@mail.gmail.com
The need for this was foreseen long ago, but when record_eq
actually became hashable (in commit 01e658fa7), we missed updating
this spot.
Per bug #17363 from Elvis Pranskevichus. Back-patch to v14 where
the faulty commit came in.
Discussion: https://postgr.es/m/17363-f6d42fd0d726be02@postgresql.org
Since enum labels have to be single-quoted, this part of the
tab completion machinery got side-swiped by commit cd69ec66c.
A side-effect of that commit is that (at least with some versions
of Readline) the text string passed for completion will omit the
leading quote mark of the enum label literal. Libedit still acts
the same as before, though, so adapt COMPLETE_WITH_ENUM_VALUE so
that it can cope with either convention.
Also, when we fail to find any valid completion, set
rl_completion_suppress_quote = 1. Otherwise readline will
go ahead and append a closing quote, which is unwanted.
Per report from Peter Eisentraut. Back-patch to v13 where
cd69ec66c came in.
Discussion: https://postgr.es/m/8ca82d89-ec3d-8b28-8291-500efaf23b25@enterprisedb.com
By default, wait_for_catchup() waits for the replication connection
to reach the primary's write LSN. That's fine, but in an apparent
attempt to save one query round-trip, it was coded so that we
executed pg_current_wal_lsn() again during each probe query.
Thus, we presented the standby with a moving target to be reached.
(While the test script itself couldn't be causing the write LSN
to advance while it's blocked in wait_for_catchup(), it's plenty
plausible that background activity such as autovacuum is emitting
more WAL.) That could make the test take longer than necessary,
and potentially it could mask bugs by allowing the standby to process
more WAL than a strict interpretation of the test scenario allows.
So, change wait_for_catchup() to do it "by the book", explicitly
collecting the write LSN to wait for at the outset.
Also, various call sites were instructing wait_for_catchup() to
wait for the standby to reach the primary's insert LSN rather than
its write LSN. This also seems like a bad idea. While in most
test scenarios those are the same, if they are different then the
inserted-but-not-yet-written WAL is not presently available to the
standby. The test isn't doing anything to make it become so, so
again we have the potential for unwanted test delay, perhaps even
a test timeout. (Again, background activity would be needed to
make this more than a hypothetical problem.) Hence, change the
callers where necessary so that the wait target is always the
primary's write LSN.
While at it, simplify callers by making use of wait_for_catchup's
default arguments wherever possible (the preceding change makes
this possible in more places than it was before). And rewrite
wait_for_catchup's documentation a bit.
Patch by me; thanks to Julien Rouhaud for review.
Discussion: https://postgr.es/m/2368336.1641843098@sss.pgh.pa.us
Add pg_statistic_ext_data.stxdinherit flag, so that for each extended
statistics definition we can store two versions of data - one for the
relation alone, one for the whole inheritance tree. This is analogous to
pg_statistic.stainherit, but we failed to include such flag in catalogs
for extended statistics, and we had to work around it (see commits
859b3003de, 36c4bc6e72 and 20b9fa308e).
This changes the relationship between the two catalogs storing extended
statistics objects (pg_statistic_ext and pg_statistic_ext_data). Until
now, there was a simple 1:1 mapping - for each definition there was one
pg_statistic_ext_data row, and this row was inserted while creating the
statistics (and then updated during ANALYZE). With the stxdinherit flag,
we don't know how many rows there will be (child relations may be added
after the statistics object is defined), so there may be up to two rows.
We could make CREATE STATISTICS to always create both rows, but that
seems wasteful - without partitioning we only need stxdinherit=false
rows, and declaratively partitioned tables need only stxdinherit=true.
So we no longer initialize pg_statistic_ext_data in CREATE STATISTICS,
and instead make that a responsibility of ANALYZE. Which is what we do
for regular statistics too.
Patch by me, with extensive improvements and fixes by Justin Pryzby.
Author: Tomas Vondra, Justin Pryzby
Reviewed-by: Tomas Vondra, Justin Pryzby
Discussion: https://postgr.es/m/20210923212624.GI831%40telsasoft.com
Since this test schedule is not run by default, it's next door to
unused. Moreover, its test coverage is very thin, and what there is
is just about entirely superseded by the src/test/recovery tests.
Let's drop it instead of carrying obsolete tests.
Discussion: https://postgr.es/m/3911012.1641246643@sss.pgh.pa.us
AFAICS we had no coverage of this point except in the seldom-used,
slated-for-removal standby_schedule test suite. Sequence updates
are enough different from regular table updates that it seems worth
covering them explicitly in src/test/recovery.
Discussion: https://postgr.es/m/999497.1641431891@sss.pgh.pa.us
Commit 859b3003de disabled building of extended stats for inheritance
trees, to prevent updating the same catalog row twice. While that
resolved the issue, it also means there are no extended stats for
declaratively partitioned tables, because there are no data in the
non-leaf relations.
That also means declaratively partitioned tables were not affected by
the issue 859b3003de addressed, which means this is a regression
affecting queries that calculate estimates for the whole inheritance
tree as a whole (which includes e.g. GROUP BY queries).
But because partitioned tables are empty, we can invert the condition
and build statistics only for the case with inheritance, without losing
anything. And we can consider them when calculating estimates.
It may be necessary to run ANALYZE on partitioned tables, to collect
proper statistics. For declarative partitioning there should no prior
statistics, and it might take time before autoanalyze is triggered. For
tables partitioned by inheritance the statistics may include data from
child relations (if built 859b3003de), contradicting the current code.
Report and patch by Justin Pryzby, minor fixes and cleanup by me.
Backpatch all the way back to PostgreSQL 10, where extended statistics
were introduced (same as 859b3003de).
Author: Justin Pryzby
Reported-by: Justin Pryzby
Backpatch-through: 10
Discussion: https://postgr.es/m/20210923212624.GI831%40telsasoft.com
Since commit 859b3003de we only build extended statistics for individual
relations, ignoring the child relations. This resolved the issue with
updating catalog tuple twice, but we still tried to use the statistics
when calculating estimates for the whole inheritance tree. When the
relations contain very distinct data, it may produce bogus estimates.
This is roughly the same issue 427c6b5b9 addressed ~15 years ago, and we
fix it the same way - by ignoring extended statistics when calculating
estimates for the inheritance tree as a whole. We still consider
extended statistics when calculating estimates for individual child
relations, of course.
This may result in plan changes due to different estimates, but if the
old statistics were not describing the inheritance tree particularly
well it's quite likely the new plans is actually better.
Report and patch by Justin Pryzby, minor fixes and cleanup by me.
Backpatch all the way back to PostgreSQL 10, where extended statistics
were introduced (same as 859b3003de).
Author: Justin Pryzby
Reported-by: Justin Pryzby
Backpatch-through: 10
Discussion: https://postgr.es/m/20210923212624.GI831%40telsasoft.com
The log_autovacuum_min_duration instrumentation used its own dedicated
code for logging, which was not reused by VACUUM VERBOSE. This was
highly duplicative, and sometimes led to each code path using slightly
different accounting for essentially the same information.
Clean things up by making VACUUM VERBOSE reuse the same instrumentation
code. This code restructuring changes the structure of the VACUUM
VERBOSE output itself, but that seems like an overall improvement. The
most noticeable change in VACUUM VERBOSE output is that it no longer
outputs a distinct message per index per round of index vacuuming. Most
of the same information (about each index) is now shown in its new
per-operation summary message. This is far more legible.
A few details are no longer displayed by VACUUM VERBOSE, but that's no
real loss in practice, especially in the common case where we don't need
multiple index scans/rounds of vacuuming. This super fine-grained
information is still available via DEBUG2 messages, which might still be
useful in debugging scenarios.
VACUUM VERBOSE now shows new instrumentation, which is typically very
useful: all of the log_autovacuum_min_duration instrumentation that it
missed out on before now. This includes information about WAL overhead,
buffers hit/missed/dirtied information, and I/O timing information.
VACUUM VERBOSE still retains a few INFO messages of its own. This is
limited to output concerning the progress of heap rel truncation, as
well as some basic information about parallel workers. These details
are still potentially quite useful. They aren't a good fit for the log
output, which must summarize the whole operation.
Author: Peter Geoghegan <pg@bowt.ie>
Reviewed-By: Masahiko Sawada <sawada.mshk@gmail.com>
Reviewed-By: Andres Freund <andres@anarazel.de>
Discussion: https://postgr.es/m/CAH2-WzmW4Me7_qR4X4ka7pxP-jGmn7=Npma_-Z-9Y1eD0MQRLw@mail.gmail.com
Add a new TAP test under src/test/recovery to run the standard
regression tests while a streaming replica replays the WAL. This
provides a basic workout for WAL decoding and redo code, and compares
the replicated result.
Optionally, enable (expensive) wal_consistency_checking if listed in
the env variable PG_TEST_EXTRA.
Reviewed-by: 綱川 貴之 (Takayuki Tsunakawa) <tsunakawa.takay@fujitsu.com>
Reviewed-by: Andres Freund <andres@anarazel.de>
Reviewed-by: Andrew Dunstan <andrew@dunslane.net>
Reviewed-by: Tom Lane <tgl@sss.pgh.pa.us>
Reviewed-by: Anastasia Lubennikova <lubennikovaav@gmail.com>
Reviewed-by: Michael Paquier <michael@paquier.xyz>
Discussion: https://postgr.es/m/CA%2BhUKGKpRWQ9SxdxxDmTBCJoR0YnFpMBe7kyzY8SUQk%2BHeskxg%40mail.gmail.com
The tablespace tests in the main regression tests have been changed to
use "in-place" tablespaces, so that they work when streamed to a replica
on the same host. Add a new TAP test that exercises tablespaces with
absolute paths, for coverage.
Reviewed-by: Andres Freund <andres@anarazel.de>
Reviewed-by: Michael Paquier <michael@paquier.xyz>
Discussion: https://postgr.es/m/CA%2BhUKGKpRWQ9SxdxxDmTBCJoR0YnFpMBe7kyzY8SUQk%2BHeskxg%40mail.gmail.com
Remove the machinery from pg_regress that manages the testtablespace
directory. Instead, use "in-place" tablespaces, because they work
correctly when there is a streaming replica running on the same host.
Reviewed-by: Andres Freund <andres@anarazel.de>
Reviewed-by: Michael Paquier <michael@paquier.xyz>
Discussion: https://postgr.es/m/CA%2BhUKGKpRWQ9SxdxxDmTBCJoR0YnFpMBe7kyzY8SUQk%2BHeskxg%40mail.gmail.com
Provide a developer-only GUC allow_in_place_tablespaces, disabled by
default. When enabled, tablespaces can be created with an empty
LOCATION string, meaning that they should be created as a directory
directly beneath pg_tblspc. This can be used for new testing scenarios,
in a follow-up patch. Not intended for end-user usage, since it might
confuse backup tools that expect symlinks.
Reviewed-by: Andres Freund <andres@anarazel.de>
Reviewed-by: Michael Paquier <michael@paquier.xyz>
Discussion: https://postgr.es/m/CA%2BhUKGKpRWQ9SxdxxDmTBCJoR0YnFpMBe7kyzY8SUQk%2BHeskxg%40mail.gmail.com
Get rid of the three-valued logic for the Boolean variables to track
whether the value was been specified and what the new value should be.
Instead, we can use the "dfoo" variables to determine whether the
value was specified and should be applied. This was already done in
some cases, so this makes this more uniform and removes one layer of
indirection.
Reviewed-by: Pavel Stehule <pavel.stehule@gmail.com>
Discussion: https://www.postgresql.org/message-id/flat/8c1a2e37-c68d-703c-5a83-7a6077f4f997@enterprisedb.com
Corruption of redirect item pointers often only becomes visible well after
being corrupted, as e.g. bug #17255 shows: In the original reproducer,
gigabyte of WAL were between the source of the corruption and the corruption
becoming visible.
To make it easier to find / prevent such bugs, verify whether redirect
pointers are sensible at the end of heap_page_prune_execute(). 5cd7eb1f1c
introduced related assertions while modifying the page, but they can't easily
detect marking the target of an existing redirect as unused. Sometimes the
corruption will be detected later, but that's harder to diagnose.
Author: Andres Freund <andres@andres@anarazel.de>
Reviewed-By: Peter Geoghegan <pg@bowt.ie>
Discussion: https://postgr.es/m/20211122175914.ayk6gg6nvdwuhrzb@alap3.anarazel.de
Since dc7420c2c9 the horizon used for pruning is determined "lazily". A more
accurate horizon is built on-demand, rather than in GetSnapshotData(). If a
horizon computation is triggered between two HeapTupleSatisfiesVacuum() calls
for the same tuple, the result can change from RECENTLY_DEAD to DEAD.
heap_page_prune() can process the same tid multiple times (once following an
update chain, once "directly"). When the result of HeapTupleSatisfiesVacuum()
of a tuple changes from RECENTLY_DEAD during the first access, to DEAD in the
second, the "tuple is DEAD and doesn't chain to anything else" path in
heap_prune_chain() can end up marking the target of a LP_REDIRECT ItemId
unused.
Initially not easily visible,
Once the target of a LP_REDIRECT ItemId is marked unused, a new tuple version
can reuse it. At that point the corruption may become visible, as index
entries pointing to the "original" redirect item, now point to a unrelated
tuple.
To fix, compute HTSV for all tuples on a page only once. This fixes the entire
class of problems of HTSV changing inside heap_page_prune(). However,
visibility changes can obviously still occur between HTSV checks inside
heap_page_prune() and outside (e.g. in lazy_scan_prune()).
The computation of HTSV is now done in bulk, in heap_page_prune(), rather than
on-demand in heap_prune_chain(). Besides being a bit simpler, it also is
faster: Memory accesses can happen sequentially, rather than in the order of
HOT chains.
There are other causes of HeapTupleSatisfiesVacuum() results changing between
two visibility checks for the same tuple, even before dc7420c2c9. E.g.
HEAPTUPLE_INSERT_IN_PROGRESS can change to HEAPTUPLE_DEAD when a transaction
aborts between the two checks. None of the these other visibility status
changes are known to cause corruption, but heap_page_prune()'s approach makes
it hard to be confident.
A patch implementing a more fundamental redesign of heap_page_prune(), which
fixes this bug and simplifies pruning substantially, has been proposed by
Peter Geoghegan in
https://postgr.es/m/CAH2-WzmNk6V6tqzuuabxoxM8HJRaWU6h12toaS-bqYcLiht16A@mail.gmail.com
However, that redesign is larger change than desirable for backpatching. As
the new design still benefits from the batched visibility determination
introduced in this commit, it makes sense to commit this narrower fix to 14
and master, and then commit Peter's improvement in master.
The precise sequence required to trigger the bug is complicated and hard to do
exercise in an isolation test (until we have wait points). Due to that the
isolation test initially posted at
https://postgr.es/m/20211119003623.d3jusiytzjqwb62p%40alap3.anarazel.de
and updated in
https://postgr.es/m/20211122175914.ayk6gg6nvdwuhrzb%40alap3.anarazel.de
isn't committable.
A followup commit will introduce additional assertions, to detect problems
like this more easily.
Bug: #17255
Reported-By: Alexander Lakhin <exclusion@gmail.com>
Debugged-By: Andres Freund <andres@anarazel.de>
Debugged-By: Peter Geoghegan <pg@bowt.ie>
Author: Andres Freund <andres@andres@anarazel.de>
Reviewed-By: Peter Geoghegan <pg@bowt.ie>
Discussion: https://postgr.es/m/20211122175914.ayk6gg6nvdwuhrzb@alap3.anarazel.de
Backpatch: 14-, the oldest branch containing dc7420c2c9
Commit 7745bc352 intended to ensure that whole-row Vars would be
printed with "::type" decoration in all contexts where plain
"var.*" notation would result in star-expansion, notably in
ROW() and VALUES() constructs. However, it missed the case of
INSERT with a single-row VALUES, as reported by Timur Khanjanov.
Nosing around ruleutils.c, I found a second oversight: the
code for RowCompareExpr generates ROW() notation without benefit
of an actual RowExpr, and naturally it wasn't in sync :-(.
(The code for FieldStore also does this, but we don't expect that
to generate strictly parsable SQL anyway, so I left it alone.)
Back-patch to all supported branches.
Discussion: https://postgr.es/m/efaba6f9-4190-56be-8ff2-7a1674f9194f@intrans.baku.az
Tom Lane
Daniel Gustafsson
Robert Haas
Michael Paquier
Thomas Munro
Peter Eisentraut
John Naylor
Andres Freund
Amit Kapila
Tomas Vondra
Fujii Masao
Peter Geoghegan