When row triggers exist in partitioned partitions that are not either
part of FKs or deferred unique constraints, they are not correctly
cloned to their partitions. That's because they are marked "internal",
and those are purposefully skipped when doing the clone triggers dance.
Fix by relaxing the condition on which internal triggers are skipped.
Amit Langote initially diagnosed the problem and proposed a fix, but I
used a different approach.
Reported-by: Petr Fedorov
Discussion: https://postgr.es/m/6b3f0646-ba8c-b3a9-c62d-1c6651a1920f@phystech.edu
Currently while decoding changes, if the number of changes exceeds a
certain threshold, we spill those to disk. And this happens for each
(sub)transaction. Now, while reading all these files, we don't close them
until we read all the files. While reading these files, if the number of
such files exceeds the maximum number of file descriptors, the operation
errors out.
Use PathNameOpenFile interface to open these files as that internally has
the mechanism to release kernel FDs as needed to get us under the
max_safe_fds limit.
Reported-by: Amit Khandekar
Author: Amit Khandekar
Reviewed-by: Amit Kapila
Backpatch-through: 9.4
Discussion: https://postgr.es/m/CAJ3gD9c-sECEn79zXw4yBnBdOttacoE-6gAyP0oy60nfs_sabQ@mail.gmail.com
This operation was possible for the owner of the schema or a superuser.
Down to 9.4, doing this operation would cause inconsistencies in a
session whose temporary schema was dropped, particularly if trying to
create new temporary objects after the drop. A more annoying
consequence is a crash of autovacuum on an assertion failure when
logging information about an orphaned temp table dropped. Note that
because of 246a6c8 (present in v11~), which has made the removal of
orphaned temporary tables more aggressive, the failure could be
triggered more easily, but it is possible to reproduce down to 9.4.
Reported-by: Mahendra Singh, Prabhat Sahu
Author: Michael Paquier
Reviewed-by: Kyotaro Horiguchi, Mahendra Singh
Discussion: https://postgr.es/m/CAKYtNAr9Zq=1-ww4etHo-VCC-k120YxZy5OS01VkaLPaDbv2tg@mail.gmail.com
Backpatch-through: 9.4
When revalidate_rectypeid() acts to update a stale record type OID
in plpgsql's data structures, it fixes the active PLpgSQL_rec struct
as well as the PLpgSQL_type struct it references. However, the latter
is shared across function executions while the former is not. In a
later function execution, the PLpgSQL_rec struct would be reinitialized
by copy_plpgsql_datums and would then contain a stale type OID,
typically leading to "could not open relation with OID NNNN" errors.
revalidate_rectypeid() can easily fix this, fortunately, just by
treating typ->typoid as authoritative.
Per report and diagnosis from Ashutosh Sharma, though this is not his
suggested fix. Back-patch to v11 where this code came in.
Discussion: https://postgr.es/m/CAE9k0Pkd4dZwt9J5pS9xhJFWpUtqs05C9xk_GEwPzYdV=GxwWg@mail.gmail.com
Our algorithm for choosing batch numbers turned out not to work
effectively for multi-billion key inner relations. We would use
more hash bits than we have, and effectively concentrate all tuples
into a smaller number of batches than we intended. While ideally
we should switch to wider hashes, for now, change the algorithm to
one that effectively gives up bits from the bucket number when we
don't have enough bits. That means we'll finish up with longer
bucket chains than would be ideal, but that's better than having
batches that don't fit in work_mem and can't be divided.
Batch-patch to all supported releases.
Author: Thomas Munro
Reviewed-by: Tom Lane, thanks also to Tomas Vondra, Alvaro Herrera, Andres Freund for testing and discussion
Reported-by: James Coleman
Discussion: https://postgr.es/m/16104-dc11ed911f1ab9df%40postgresql.org
While building a hash map of categories in load_categories_hash,
resulting category names have not thus far been checked to ensure
they are not null. Prior to pg12 null category names worked to the
extent that they did not crash on some platforms. This is because
those system libraries have an snprintf which can deal with being
passed a null pointer argument for a string. But even in those cases
null categories did nothing useful. And on some platforms it crashed.
As of pg12, our own version of snprintf gets called, and it does
not deal with null pointer arguments at all, and crashes consistently.
Fix that by disallowing null categories. They never worked usefully,
and no one has ever asked for them to work previously. Back-patch to
all supported branches.
Reported-By: Ireneusz Pluta
Discussion: https://postgr.es/m/16176-7489719b05e4303c@postgresql.org
This wasn't checked originally, but it should have been, because
in general pseudo-types can't be stored to and retrieved from disk.
Notably, partition bound values of type "record" would not be
interpretable by another session.
In v12 and HEAD, add another flag to CheckAttributeType's repertoire
so that it can produce a specific error message for this case. That's
infeasible in older branches without an ABI break, so fall back to
a slightly-less-nicely-worded error message in v10 and v11.
Problem noted by Amit Langote, though this patch is not his initial
solution. Back-patch to v10 where partitioning was introduced.
Discussion: https://postgr.es/m/CA+HiwqFUzjfj9HEsJtYWcr1SgQ_=iCAvQ=O2Sx6aQxoDu4OiHw@mail.gmail.com
We probably should have thought of this case when ranges were added,
but we didn't. (It's not the fault of commit eb51af71f, because
ranges didn't exist then.)
It's an old bug, so back-patch to all supported branches.
Discussion: https://postgr.es/m/7782.1577051475@sss.pgh.pa.us
If the first transaction block in these tests were entered exactly
at midnight (California time), they'd report a bogus failure due
to 'now' and 'midnight' having the same values. Commit 8c2ac75c5
had dismissed this as being of negligible probability, but we've
now seen it happen in the buildfarm, so let's prevent it. We can
get pretty much the same test coverage without an it's-not-midnight
assumption by moving the does-'now'-work cases into their own test step.
While here, apply commit 47169c255's s/DELETE/TRUNCATE/ change to
timestamptz as well as timestamp (not sure why that didn't
occur to me at the time; the risk of failure is the same).
Back-patch to all supported branches, since the main point is
to get rid of potential buildfarm failures.
Discussion: https://postgr.es/m/14821.1577031117@sss.pgh.pa.us
This fixes a performance problem introduced by commit 6d7547c21.
ERROR_ACCESS_DENIED is returned in some other cases besides the
delete-pending case considered by that commit; notably, if the
given path names a directory instead of a plain file. In that
case we'll uselessly loop for 1 second before returning the
failure condition. That slows down some usage scenarios enough
to cause test timeout failures on our Windows buildfarm critters.
To fix, try to stat() the file, and sleep/loop only if that fails.
It will fail in the delete-pending case, and also in the case where
the deletion completed before we could stat(), so we have the cases
where we want to loop covered. In the directory case, the stat()
should succeed, letting us exit without a wait.
One case where we'll still wait uselessly is if the access-denied
problem pertains to a directory in the given pathname. But we don't
expect that to happen in any performance-critical code path.
There might be room to refine this further, but I'll push it now
in hopes of making the buildfarm green again.
Back-patch, like the preceding commit.
Alexander Lakhin and Tom Lane
Discussion: https://postgr.es/m/23073.1576626626@sss.pgh.pa.us
We realized years ago that it's better for libpq to accept all
connection parameters syntactically, even if some are ignored or
restricted due to lack of the feature in a particular build.
However, that lesson from the SSL support was for some reason never
applied to the GSSAPI support. This is causing various buildfarm
members to have problems with a test case added by commit 6136e94dc,
and it's just a bad idea from a user-experience standpoint anyway,
so fix it.
While at it, fix some places where parameter-related infrastructure
was added with the aid of a dartboard, or perhaps with the aid of
the anti-pattern "add new stuff at the end". It should be safe
to rearrange the contents of struct pg_conn even in released
branches, since that's private to libpq (and we'd have to move
some fields in some builds to fix this, anyway).
Back-patch to all supported branches.
Discussion: https://postgr.es/m/11297.1576868677@sss.pgh.pa.us
The "auth-methods" <sect1> used to include descriptions of all our
authentication methods. Commit 56811e573 promoted its child <sect2>'s
to <sect1>'s, which has advantages but also created some issues:
* The auth-methods page itself is essentially empty/useless.
* Links that pointed to "auth-methods" as a placeholder for all
auth methods were rendered a bit nonsensical.
* DocBook no longer provides a subsection table-of-contents here,
which formerly was a useful if terse summary of available auth methods.
To improve matters, add a handwritten list of all the auth methods.
Per gripe from Dave Cramer. Back-patch to v11 where the previous
commit came in.
Discussion: https://postgr.es/m/CADK3HH+xQLhcPgg=kWqfogtXGGZr-JdSo=x=WQC0PkAVyxUWyQ@mail.gmail.com
This patch allows building the local relmap cache for a subscribed
relation after processing pending invalidation messages and potential
relcache updates. Without this, the attributes in the local cache don't
tally with the updated relcache entry leading to invalid memory access.
Reported-by Jehan-Guillaume de Rorthais
Author: Jehan-Guillaume de Rorthais and Vignesh C
Reviewed-by: Amit Kapila
Backpatch-through: 10
Discussion: https://postgr.es/m/20191025175929.7e90dbf5@firost
This changes the routines in charge of recycling WAL segments past the
last redo LSN to not use anymore "RedoRecPtr" as a local variable, which
is also available in the context of the session as a static declaration,
replacing it with "lastredoptr". This confusion has been introduced by
d9fadbf, so backpatch down to v11 like the other commit.
Thanks to Tom Lane, Robert Haas, Alvaro Herrera, Mark Dilger and Kyotaro
Horiguchi for the input provided.
Author: Ranier Vilela
Discussion: https://postgr.es/m/MN2PR18MB2927F7B5F690065E1194B258E35D0@MN2PR18MB2927.namprd18.prod.outlook.com
Backpatch-through: 11
If CheckAttributeType() threw an error about the datatype of an
index expression column, it would report an empty column name,
which is pretty unhelpful and certainly not the intended behavior.
I (tgl) evidently broke this in commit cfc5008a5, by not noticing
that the column's attname was used above where I'd placed the
assignment of it.
In HEAD and v12, this is trivially fixable by moving up the
assignment of attname. Before v12 the code is a bit more messy;
to avoid doing substantial refactoring, I took the lazy way out
and just put in two copies of the assignment code.
Report and patch by Amit Langote. Back-patch to all supported
branches.
Discussion: https://postgr.es/m/CA+HiwqFA+BGyBFimjiYXXMa2Hc3fcL0+OJOyzUNjhU4NCa_XXw@mail.gmail.com
This Assert thought that an overflowed transaction can never get registered
for the group update. But that is not true, because even when the number
of children for a transaction got reduced, the overflow flag is not
changed. And, for group update, we only care about the current number of
children for a transaction that is being committed.
Based on comments by Andres Freund, remove a redundant Assert in
TransactionIdSetPageStatus as we already had a static Assert for the same
condition a few lines earlier.
Reported-by: Vignesh C
Author: Dilip Kumar
Reviewed-by: Amit Kapila
Backpatch-through: 11
Discussion: https://postgr.es/m/CAFiTN-s5=uJw-Z6JC9gcqtBSjXsrHnU63PXBrA=pnBjqnkm5UA@mail.gmail.com
Attempting to open a file fails with ERROR_ACCESS_DENIED if the file
is flagged for deletion but not yet actually gone (another in a long
list of reasons why Windows is broken, if you ask me). This seems
likely to explain a lot of irreproducible failures we see in the
buildfarm. This state generally persists for only a millisecond or so,
so just wait a bit and retry. If it's a real permissions problem,
we'll eventually give up and report it as such. If it's the pending
deletion case, we'll see file-not-found and report that after the
deletion completes, and the caller will treat that in an appropriate
way.
In passing, rejigger the existing retry logic for some other error
cases so that we don't uselessly wait an extra time when we're
not going to retry anymore.
Alexander Lakhin (with cosmetic tweaks by me). Back-patch to all
supported branches, since this seems like a pretty safe change and
the problem is definitely real.
Discussion: https://postgr.es/m/16161-7a985d2f1bbe8f71@postgresql.org
Commit a7ee7c8513 fixed a bug in GiST page split during index creation,
where we failed to re-find the position of a downlink after the page
containing it was split. However, that fix was incomplete; the other call
to gistinserttuples() in the same function needs to also clear
'downlinkoffnum'.
Fixes bug #16134 reported by Alexander Lakhin, for real this time. The
previous fix was enough to fix the crash with the reproducer script for
bug #16162, but the original script for #16134 was still crashing.
Backpatch to v12, like the previous incomplete fix.
Discussion: https://www.postgresql.org/message-id/d869f537-abe4-d2ea-0510-38cd053f5152%40gmail.com
The RTE_RESULT simplification logic added by commit 4be058fe9 had a
flaw: it would collapse out a RTE_RESULT that is due to compute a
PlaceHolderVar, and reassign the PHV to the parent join level, even if
another input relation of the join contained a lateral reference to
the PHV. That can't work because the PHV would be computed too late.
In practice it led to failures of internal sanity checks later in
planning (either assertion failures or errors such as "failed to
construct the join relation").
To fix, add code to check for the presence of such PHVs in relevant
portions of the query tree. Notably, this required refactoring
range_table_walker so that a caller could ask to walk individual RTEs
not the whole list. (It might be a good idea to refactor
range_table_mutator in the same way, if only to keep those functions
looking similar; but I didn't do so here as it wasn't necessary for
the bug fix.)
This exercise also taught me that find_dependent_phvs(), as it stood,
could only safely be used on the entire Query, not on subtrees.
Adjust its API to reflect that; which in passing allows it to have
a fast path for the common case of no PHVs anywhere.
Per report from Will Leinweber. Back-patch to v12 where the bug
was introduced.
Discussion: https://postgr.es/m/CALLb-4xJMd4GZt2YCecMC95H-PafuWNKcmps4HLRx2NHNBfB4g@mail.gmail.com
The previous commit failed to consider that FileGetRawDesc() might
not return a valid fd, as discovered on the build farm. Switch to
using the File interface only.
Back-patch to 12, like the previous commit.
_mdfd_getseg() opens all segments up to the requested one. That
causes problems for mdsyncfiletag(), if mdunlinkfork() has
already unlinked other segment files. Open the file we want
directly by name instead, if it's not already open.
The consequence of this bug was a rare panic in the checkpointer,
made more likely if you saturated the sync request queue so that
the SYNC_FORGET_REQUEST messages for a given relation were more
likely to be absorbed in separate cycles by the checkpointer.
Back-patch to 12. Defect in commit 3eb77eba.
Author: Thomas Munro
Reported-by: Justin Pryzby
Discussion: https://postgr.es/m/20191119115759.GI30362%40telsasoft.com
The bug was similar to the one that was fixed in commit 22251686f0. When
we split page X and insert the downlink for the new page, the parent page
might also need to be split. When that happens, the downlink offset number
we remembered for X is no longer valid. We correctly called
gistFindCorrectParent() to re-find it, but gistFindCorrectParent() doesn't
do anything if the LSN of the page hasn't changed, and we stopped updating
LSNs during index build in commit 9155580fd5. The buggy codepath was taken
if the page was split into three or more pages, and inserting the downlink
caused the parent page to split. To fix, explicitly mark the downlink
offset number as invalid, to force gistFindCorrectParent() to re-find it.
Fixes bug #16134 reported by Alexander Lakhin, reported again as #16162 by
Andreas Kunert. Thanks to Jeff Janes, Tom Lane and Tomas Vondra for
debugging. Backpatch to v12, where we stopped WAL-logging during index
build.
Discussion: https://www.postgresql.org/message-id/16134-0423f729671dec64%40postgresql.org
Discussion: https://www.postgresql.org/message-id/16162-45d21b7b6c1a3105%40postgresql.org
The DTK_DOW/DTK_ISODOW and DTK_DOY switch cases in timestamp_part() and
timestamptz_part() contained calls of timestamp2tm() that were fully
redundant with the ones done just above the switch. This evidently crept
in during commit 258ee1b63, which relocated that code from another place
where the calls were indeed needed. Just delete the redundant calls.
I (tgl) noted that our test coverage of these functions left quite a
bit to be desired, so extend timestamp.sql and timestamptz.sql to
cover all the branches.
Back-patch to all supported branches, as the previous commit was.
There's no real issue here other than some wasted cycles in some
not-too-heavily-used code paths, but the test coverage seems valuable.
Report and patch by Li Japin; test case adjustments by me.
Discussion: https://postgr.es/m/SG2PR06MB37762CAE45DB0F6CA7001EA9B6550@SG2PR06MB3776.apcprd06.prod.outlook.com
On Windows, we use CMD.EXE to redirect the postmaster's stdout/stderr
into a log file. CMD.EXE will open that file with non-sharing-friendly
parameters, and the file will remain open for a short time after the
postmaster has removed postmaster.pid. This can result in an
ERROR_SHARING_VIOLATION failure if we attempt to start a new postmaster
immediately with the same log file (e.g. during "pg_ctl restart").
This seems to explain intermittent buildfarm failures we've been seeing
on Windows machines.
To fix, just open and close the log file using our own pgwin32_open(),
which will wait if necessary to avoid the failure. (Perhaps someday
we should stop using CMD.EXE, but that would be a far more complex
patch, and it doesn't seem worth the trouble ... yet.)
Back-patch to v12. This only solves the problem when frontend fopen()
is redirected to pgwin32_fopen(), which has only been true since commit
0ba06e0bf. Hence, no point in back-patching further, unless we care
to back-patch that change too.
Diagnosis and patch by Alexander Lakhin (bug #16154).
Discussion: https://postgr.es/m/16154-1ccf0b537b24d5e0@postgresql.org
AfterTriggerExecute() retrieves a fresh tuple or pair of tuples from a
tuplestore and then stores the tuple(s) in the passed-in slot(s) if
AFTER_TRIGGER_FDW_FETCH, while it uses the most-recently-retrieved
tuple(s) stored in the slot(s) if AFTER_TRIGGER_FDW_REUSE. This was
done correctly before 12, but commit ff11e7f4b broke it by mistakenly
clearing the tuple(s) stored in the slot(s) in that function, leading to
an assertion failure as reported in bug #16139 from Alexander Lakhin.
Also, fix some other issues with the aforementioned commit in passing:
* For tg_newslot, which is a slot added to the TriggerData struct by the
commit to store new updated tuples, it didn't ensure the slot was NULL
if there was no such tuple.
* The commit failed to update the documentation about the trigger
interface.
Author: Etsuro Fujita
Backpatch-through: 12
Discussion: https://postgr.es/m/16139-94f9ccf0db6119ec%40postgresql.org
pg_signal_dispatch_thread() responded to the client (signal sender)
and disconnected the pipe before actually setting the shared variables
that make the signal visible to the backend process's main thread.
In the worst case, it seems, effective delivery of the signal could be
postponed for as long as the machine has any other work to do.
To fix, just move the pg_queue_signal() call so that we do it before
responding to the client. This essentially makes pgkill() synchronous,
which is a stronger guarantee than we have on Unix. That may be
overkill, but on the other hand we have not seen comparable timing bugs
on any Unix platform.
While at it, add some comments to this sadly underdocumented code.
Problem diagnosis and fix by Amit Kapila; I just added the comments.
Back-patch to all supported versions, as it appears that this can cause
visible NOTIFY timing oddities on all of them, and there might be
other misbehavior due to slow delivery of other signals.
Discussion: https://postgr.es/m/32745.1575303812@sss.pgh.pa.us
isolationtester.c had a hard-wired limit of 3 minutes per test step.
It now emerges that this isn't quite enough for some of the slowest
buildfarm animals. This isn't the first time we've had to raise
this limit (cf. 1db439ad4), so let's make it configurable. This
patch raises the default to 5 minutes, and introduces an environment
variable PGISOLATIONTIMEOUT that can be set if more time is needed,
following the precedent of PGCTLTIMEOUT.
Also, modify isolationtester so that when the timeout is hit,
it explicitly reports having sent a cancel. This makes the regression
failure log considerably more intelligible. (In the worst case, a
timed-out test might actually be reported as "passing" without this
extra output, so arguably this is a bug fix in itself.)
In passing, update the README file, which had apparently not gotten
touched when we added "make check" support here.
Back-patch to 9.6; older versions don't have comparable timeout logic.
Discussion: https://postgr.es/m/22964.1575842935@sss.pgh.pa.us
Commit 5770172cb0 wrote, incorrectly, that
certain schema usage patterns are secure against CREATEROLE users and
database owners. When an untrusted user is the database owner or holds
CREATEROLE privilege, a query is secure only if its session started with
SELECT pg_catalog.set_config('search_path', '', false) or equivalent.
Back-patch to 9.4 (all supported versions).
Discussion: https://postgr.es/m/20191013013512.GC4131753@rfd.leadboat.com
This function is supported down to OpenSSL 0.9.8, which is the oldest
version supported since 593d4e4 (from Postgres 10 onwards), and is used
since e3bdb2d (from 11 onwards). It is defined as a macro from OpenSSL
0.9.8 to 1.0.2, and as a function in 1.1.0 and newer versions. However,
the configure check present is only adapted for functions. So, even if
the code would be able to compile, configure fails to detect the macro,
causing it to be ignored when compiling the code with OpenSSL from 0.9.8
to 1.0.2.
The code needs a configure check as per a364dfa, which has fixed a
compilation issue with a past version of LibreSSL in NetBSD 5.1. On
HEAD, just remove the configure check as the last release of NetBSD 5 is
from 2014 (and we have no more buildfarm members for it). In 11 and 12,
improve the configure logic so as both macros and functions are
correctly detected. This makes NetBSD 5 still work on already-released
branches, but not for 13 onwards.
The patch for HEAD is from me, and Daniel has written the version to use
for the back-branches.
Author: Michael Paquier, Daniel Gustaffson
Reviewed-by: Tom Lane
Discussion: https://postgr.es/m/20191205083252.GE5064@paquier.xyz
Discussion: https://postgr.es/m/98F7F99E-1129-41D8-B86B-FE3B1E286881@yesql.se
Backpatch-through: 11
The dict_int text search dictionary template accepts maxlen parameter,
which is then used to cap the length of input strings. The value was
not properly checked, and the code simply does
txt[d->maxlen] = '\0';
to insert a terminator, leading to segfaults with negative values.
This commit simply rejects values less than 1. The issue was there since
dct_int was introduced in 9.3, so backpatch all the way back to 9.4
which is the oldest supported version.
Reported-by: cili
Discussion: https://postgr.es/m/16144-a36a5bef7657047d@postgresql.org
Backpatch-through: 9.4
On Windows, all the hosts spawned by the TAP tests bind to 127.0.0.1.
Hence, if there is a port conflict, starting a cluster would immediately
fail. One of the test scripts of pg_ctl initializes a node without
PostgresNode.pm, using the default port 5432. This could cause
unexpected startup failures in the tests if an independent server was up
and running on the same host (the reverse is also possible, though more
unlikely). Fix this issue by assigning properly a free port to the node
configured, in the same range used as for the other nodes part of the
tests.
Author: Michael Paquier
Reviewed-by: Andrew Dunstan
Discussion: https://postgr.es/m/20191202031444.GC1696@paquier.xyz
Backpatch-through: 11
We implement ON COMMIT DELETE ROWS by truncating tables marked that
way, which requires also truncating/rebuilding their indexes. But
RelationTruncateIndexes asks the relcache for up-to-date copies of any
index expressions, which may cause execution of eval_const_expressions
on them, which can result in actual execution of subexpressions.
This is a bad thing to have happening during ON COMMIT. Manuel Rigger
reported that use of a SQL function resulted in crashes due to
expectations that ActiveSnapshot would be set, which it isn't.
The most obvious fix perhaps would be to push a snapshot during
PreCommit_on_commit_actions, but I think that would just open the door
to more problems: CommitTransaction explicitly expects that no
user-defined code can be running at this point.
Fortunately, since we know that no tuples exist to be indexed, there
seems no need to use the real index expressions or predicates during
RelationTruncateIndexes. We can set up dummy index expressions
instead (we do need something that will expose the right data type,
as there are places that build index tupdescs based on this), and
just ignore predicates and exclusion constraints.
In a green field it'd likely be better to reimplement ON COMMIT DELETE
ROWS using the same "init fork" infrastructure used for unlogged
relations. That seems impractical without catalog changes though,
and even without that it'd be too big a change to back-patch.
So for now do it like this.
Per private report from Manuel Rigger. This has been broken forever,
so back-patch to all supported branches.
When using %b or %B patterns to format a date, the code was simply using
tm_mon as an index into array of month names. But that is wrong, because
tm_mon is 1-based, while array indexes are 0-based. The result is we
either use name of the next month, or a segfault (for December).
Fix by subtracting 1 from tm_mon for both patterns, and add a regression
test triggering the issue. Backpatch to all supported versions (the bug
is there far longer, since at least 2003).
Reported-by: Paul Spencer
Backpatch-through: 9.4
Discussion: https://postgr.es/m/16143-0d861eb8688d3fef%40postgresql.org
When picking the best extended statistics object for a list of clauses,
it's not enough to look at attnums extracted from the clause list as a
whole. Consider for example this query with OR clauses:
SELECT * FROM t WHERE (t.a = 1) OR (t.b = 1) OR (t.c = 1)
with a statistics defined on columns (a,b). Relying on attnums extracted
from the whole OR clause, we'd consider the statistics usable. That does
not work, as we see the conditions as a single OR-clause, referencing an
attribute not covered by the statistic, leading to empty list of clauses
to be estimated using the statistics and an assert failure.
This changes choose_best_statistics to check which clauses are actually
covered, and only using attributes from the fully covered ones. For the
previous example this means the statistics object will not be considered
as compatible with the OR-clause.
Backpatch to 12, where MCVs were introduced. The issue does not affect
older versions because functional dependencies don't handle OR clauses.
Author: Tomas Vondra
Reviewed-by: Dean Rasheed
Reported-By: Manuel Rigger
Discussion: https://postgr.es/m/CA+u7OA7H5rcE2=8f263w4NZD6ipO_XOrYB816nuLXbmSTH9pQQ@mail.gmail.com
Backpatch-through: 12