Currently, B-tree code matches every scan key to every item on the page.
Imagine the ordered B-tree scan for the query like this.
SELECT * FROM tbl WHERE col > 'a' AND col < 'b' ORDER BY col;
The (col > 'a') scan key will be always matched once we find the location to
start the scan. The (col < 'b') scan key will match every item on the page
as long as it matches the last item on the page.
This patch implements prechecking of the scan keys required for directional
scan on beginning of page scan. If precheck is successful we can skip this
scan keys check for the items on the page. That could lead to significant
acceleration especially if the comparison operator is expensive.
Idea from patch by Konstantin Knizhnik.
Discussion: https://postgr.es/m/079c3f8e-3371-abe2-e93c-fc8a0ae3f571%40garret.ru
Reviewed-by: Peter Geoghegan, Pavel Borisov
BuildDescForRelation() main job is to convert ColumnDef lists to
pg_attribute/tuple descriptor arrays, which is really mostly an
internal subroutine of DefineRelation() and some related functions,
which is more the remit of tablecmds.c and doesn't have much to do
with the basic tuple descriptor interfaces in tupdesc.c. This is also
supported by observing the header includes we can remove in tupdesc.c.
By moving it over, we can also (in the future) make
BuildDescForRelation() use more internals of tablecmds.c that are not
sensible to be exposed in tupdesc.c.
Discussion: https://www.postgresql.org/message-id/flat/52a125e4-ff9a-95f5-9f61-b87cf447e4da@eisentraut.org
Make a few newish calls to appendStringInfo() which have no special
formatting use appendStringInfoString() instead. Also, adjust usages of
appendStringInfoString() which only append a string containing a single
character to make use of appendStringInfoChar() instead.
This makes the code marginally faster, but primarily this change is so
we use the StringInfo type as it was intended to be used.
Discussion: https://postgr.es/m/CAApHDvpXKQmL+r=VDNS98upqhr9yGBhv2Jw3GBFFk_wKHcB39A@mail.gmail.com
This commit changes the WAL reader routines so as a FATAL for the
backend or exit(FAILURE) for the frontend is triggered if an allocation
for a WAL record decode fails in walreader.c, rather than treating this
case as bogus data, which would be equivalent to the end of WAL. The
key is to avoid palloc_extended(MCXT_ALLOC_NO_OOM) in walreader.c,
relying on plain palloc() calls.
The previous behavior could make WAL replay finish too early than it
should. For example, crash recovery finishing earlier may corrupt
clusters because not all the WAL available locally was replayed to
ensure a consistent state. Out-of-memory failures would show up
randomly depending on the memory pressure on the host, but one simple
case would be to generate a large record, then replay this record after
downsizing a host, as Ethan Mertz originally reported.
This relies on bae868caf2, as the WAL reader routines now do the
memory allocation required for a record only once its header has been
fully read and validated, making xl_tot_len trustable. Making the WAL
reader react differently on out-of-memory or bogus record data would
require ABI changes, so this is the safest choice for stable branches.
Also, it is worth noting that 3f1ce97346 has been using a plain
palloc() in this code for some time now.
Thanks to Noah Misch and Thomas Munro for the discussion.
Like the other commit, backpatch down to 12, leaving out v11 that will
be EOL'd soon. The behavior of considering a failed allocation as bogus
data comes originally from 0ffe11abd3, where the record length
retrieved from its header was not entirely trustable.
Reported-by: Ethan Mertz
Discussion: https://postgr.es/m/ZRKKdI5-RRlta3aF@paquier.xyz
Backpatch-through: 12
The retry loop is needed because heap_page_prune() calls
HeapTupleSatisfiesVacuum() and then lazy_scan_prune() does the same
thing again, and they might get different answers due to concurrent
clog updates. But this patch makes heap_page_prune() return the
HeapTupleSatisfiesVacuum() results that it computed back to the
caller, which allows lazy_scan_prune() to avoid needing to recompute
those values in the first place. That's nice both because it eliminates
the need for a retry loop and also because it's cheaper.
Melanie Plageman, reviewed by David Geier, Andres Freund, and me.
Discussion: https://postgr.es/m/CAAKRu_br124qsGJieuYA0nGjywEukhK1dKBfRdby_4yY3E9SXA%40mail.gmail.com
nbtree's mark/restore processing failed to correctly handle an edge case
involving array key advancement and related search-type scan key state.
Scans with ScalarArrayScalarArrayOpExpr quals requiring mark/restore
processing (for a merge join) could incorrectly conclude that an
affected array/scan key must not have advanced during the time between
marking and restoring the scan's position.
As a result of all this, array key handling within btrestrpos could skip
a required call to _bt_preprocess_keys(). This confusion allowed later
primitive index scans to overlook tuples matching the true current array
keys. The scan's search-type scan keys would still have spurious values
corresponding to the final array element(s) -- not values matching the
first/now-current array element(s).
To fix, remember that "array key wraparound" has taken place during the
ongoing btrescan in a flag variable stored in the scan's state, and use
that information at the point where btrestrpos decides if another call
to _bt_preprocess_keys is required.
Oversight in commit 70bc5833, which taught nbtree to handle array keys
during mark/restore processing, but missed this subtlety. That commit
was itself a bug fix for an issue in commit 9e8da0f7, which taught
nbtree to handle ScalarArrayOpExpr quals natively.
Author: Peter Geoghegan <pg@bowt.ie>
Discussion: https://postgr.es/m/CAH2-WzkgP3DDRJxw6DgjCxo-cu-DKrvjEv_ArkP2ctBJatDCYg@mail.gmail.com
Backpatch: 11- (all supported branches).
Previously, one of the values in the struct was returned as the return
value, and another was returned via an output parameter. In
preparation for returning more stuff, consolidate both values into a
struct returned via an output parameter.
Melanie Plageman, reviewed by Andres Freund and by me.
Discussion: https://postgr.es/m/CAAKRu_br124qsGJieuYA0nGjywEukhK1dKBfRdby_4yY3E9SXA%40mail.gmail.com
This unifies some repetitive code.
Note: I didn't push the "not found" error message into the new
function, even though all existing callers would be able to make use
of it. Using the existing error handling as-is would probably require
exposing the Relation type via tupdesc.h, which doesn't seem
desirable. (Or even if we changed it to just report the OID, it would
inject the concept of a relation containing the tuple descriptor into
tupdesc.h, which might be a layering violation. Perhaps some further
improvements could be considered here separately.)
Discussion: https://www.postgresql.org/message-id/flat/52a125e4-ff9a-95f5-9f61-b87cf447e4da%40eisentraut.org
Yet another bug in the ilk of commits a7ee7c851 and 741b88435. In
741b88435, we took care to clear the memorized location of the
downlink when we split the parent page, because splitting the parent
page can move the downlink. But we missed that even *updating* a tuple
on the parent can move it, because updating a tuple on a gist page is
implemented as a delete+insert, so the updated tuple gets moved to the
end of the page.
This commit fixes the bug in two different ways (belt and suspenders):
1. Clear the downlink when we update a tuple on the parent page, even
if it's not split. This the same approach as in commits a7ee7c851
and 741b88435.
I also noticed that gistFindCorrectParent did not clear the
'downlinkoffnum' when it stepped to the right sibling. Fix that
too, as it seems like a clear bug even though I haven't been able
to find a test case to hit that.
2. Change gistFindCorrectParent so that it treats 'downlinkoffnum'
merely as a hint. It now always first checks if the downlink is
still at that location, and if not, it scans the page like before.
That's more robust if there are still more cases where we fail to
clear 'downlinkoffnum' that we haven't yet uncovered. With this,
it's no longer necessary to meticulously clear 'downlinkoffnum',
so this makes the previous fixes unnecessary, but I didn't revert
them because it still seems nice to clear it when we know that the
downlink has moved.
Also add the test case using the same test data that Alexander
posted. I tried to reduce it to a smaller test, and I also tried to
reproduce this with different test data, but I was not able to, so
let's just include what we have.
Backpatch to v12, like the previous fixes.
Reported-by: Alexander Lakhin
Discussion: https://www.postgresql.org/message-id/18129-caca016eaf0c3702@postgresql.org
bae868ca removed a check that was still needed. If you had an
xl_tot_len at the end of a page that was too small for a record header,
but not big enough to span onto the next page, we'd immediately perform
the CRC check using a bogus large length. Because of arbitrary coding
differences between the CRC implementations on different platforms,
nothing very bad happened on common modern systems. On systems using
the _sb8.c fallback we could segfault.
Restore that check, add a new assertion and supply a test for that case.
Back-patch to 12, like bae868ca.
Tested-by: Tom Lane <tgl@sss.pgh.pa.us>
Tested-by: Alexander Lakhin <exclusion@gmail.com>
Discussion: https://postgr.es/m/CA%2BhUKGLCkTT7zYjzOxuLGahBdQ%3DMcF%3Dz5ZvrjSOnW4EDhVjT-g%40mail.gmail.com
xl_tot_len comes first in a WAL record. Usually we don't trust it to be
the true length until we've validated the record header. If the record
header was split across two pages, previously we wouldn't do the
validation until after we'd already tried to allocate enough memory to
hold the record, which was bad because it might actually be garbage
bytes from a recycled WAL file, so we could try to allocate a lot of
memory. Release 15 made it worse.
Since 70b4f82a4b, we'd at least generate an end-of-WAL condition if the
garbage 4 byte value happened to be > 1GB, but we'd still try to
allocate up to 1GB of memory bogusly otherwise. That was an
improvement, but unfortunately release 15 tries to allocate another
object before that, so you could get a FATAL error and recovery could
fail.
We can fix both variants of the problem more fundamentally using
pre-existing page-level validation, if we just re-order some logic.
The new order of operations in the split-header case defers all memory
allocation based on xl_tot_len until we've read the following page. At
that point we know that its first few bytes are not recycled data, by
checking its xlp_pageaddr, and that its xlp_rem_len agrees with
xl_tot_len on the preceding page. That is strong evidence that
xl_tot_len was truly the start of a record that was logged.
This problem was most likely to occur on a standby, because
walreceiver.c recycles WAL files without zeroing out trailing regions of
each page. We could fix that too, but it wouldn't protect us from rare
crash scenarios where the trailing zeroes don't make it to disk.
With reliable xl_tot_len validation in place, the ancient policy of
considering malloc failure to indicate corruption at end-of-WAL seems
quite surprising, but changing that is left for later work.
Also included is a new TAP test to exercise various cases of end-of-WAL
detection by writing contrived data into the WAL from Perl.
Back-patch to 12. We decided not to put this change into the final
release of 11.
Author: Thomas Munro <thomas.munro@gmail.com>
Author: Michael Paquier <michael@paquier.xyz>
Reported-by: Alexander Lakhin <exclusion@gmail.com>
Reviewed-by: Noah Misch <noah@leadboat.com> (the idea, not the code)
Reviewed-by: Michael Paquier <michael@paquier.xyz>
Reviewed-by: Sergei Kornilov <sk@zsrv.org>
Reviewed-by: Alexander Lakhin <exclusion@gmail.com>
Discussion: https://postgr.es/m/17928-aa92416a70ff44a2%40postgresql.org
In older branches, COMMIT/ROLLBACK AND CHAIN failed to propagate
the current transaction's properties to the new transaction if
there was any open subtransaction (unreleased savepoint).
Instead, some previous transaction's properties would be restored.
This is because the "if (s->chain)" check in CommitTransactionCommand
examined the wrong instance of the "chain" flag and falsely
concluded that it didn't need to save transaction properties.
Our regression tests would have noticed this, except they used
identical transaction properties for multiple tests in a row,
so that the faulty behavior was not distinguishable from correct
behavior.
Commit 12d768e70 fixed the problem in v15 and later, but only rather
accidentally, because I removed the "if (s->chain)" test to avoid a
compiler warning, while not realizing that the warning was flagging a
real bug.
In v14 and before, remove the if-test and save transaction properties
unconditionally; just as in the newer branches, that's not expensive
enough to justify thinking harder.
Add the comment and extra regression test to v15 and later to
forestall any future recurrence, but there's no live bug in those
branches.
Patch by me, per bug #18118 from Liu Xiang. Back-patch to v12 where
the AND CHAIN feature was added.
Discussion: https://postgr.es/m/18118-4b72fcbb903aace6@postgresql.org
When tracking IO timing for WAL, the duration is what we calculate
based on the start and end timestamps, it's not what the variable
contains. Rename the timestamp variable to end to better communicate
what it contains. Original patch by Krishnakumar with additional
hacking to fix another occurrence by me.
Author: Krishnakumar R <kksrcv001@gmail.com>
Discussion: https://postgr.es/m/CAPMWgZ9f9o8awrQpjo8oxnNQ=bMDVPx00NE0QcDzvHD_ZrdLPw@mail.gmail.com
This became safe after commit 4b4798e138. The smgrcreate() call will
now register the segment for syncing at the next checkpoint, so we
don't need to sync it here. If a checkpoint happens before the
creation is WAL-logged, the records will be replayed when starting
recovery from the checkpoint. If a checkpoint happens after the WAL
logging, the checkpoint will fsync() it.
In the passing, clarify a comment in smgrDoPendingSyncs().
Discussion: https://www.postgresql.org/message-id/6e5bbc08-cdfc-b2b3-9e23-1a914b9850a9%40iki.fi
Reviewed-by: Robert Haas
The majority of all filenames are quoted in user facing error and
log messages, but a few were still printed without quotes. While
these filenames do not risk causing any ambiguity as their format
is strict, quote them anyways to be consistent across all logs.
Also concatenate a message to keep it one line to make it easier
to grep for in the code.
Reviewed-by: Peter Eisentraut <peter@eisentraut.org>
Reviewed-by: Michael Paquier <michael@paquier.xyz>
Discussion: https://postgr.es/m/080EEABE-6645-4A46-AB20-6285ADAC44FE@yesql.se
It's entirely possible for a logical slot to have a confirmed_flush LSN
higher than the last value saved on disk while not being marked as dirty.
Currently, it is not a major problem but a later patch adding support for
the upgrade of slots relies on that value being properly flushed to disk.
It can also help avoid processing the same transactions again in some
boundary cases after the clean shutdown and restart. Say, we process
some transactions for which we didn't send anything downstream (the
changes got filtered) but the confirm_flush LSN is updated due to
keepalives. As we don't flush the latest value of confirm_flush LSN, it
may lead to processing the same changes again without this patch.
The approach taken by this patch has been suggested by Ashutosh Bapat.
Author: Vignesh C, Julien Rouhaud, Kuroda Hayato
Reviewed-by: Amit Kapila, Dilip Kumar, Michael Paquier, Ashutosh Bapat, Peter Smith, Hou Zhijie
Discussion: http://postgr.es/m/CAA4eK1JzJagMmb_E8D4au=GYQkxox0AfNBm1FbP7sy7t4YWXPQ@mail.gmail.com
Discussion: http://postgr.es/m/TYAPR01MB58664C81887B3AF2EB6B16E3F5939@TYAPR01MB5866.jpnprd01.prod.outlook.com
The comment in heapgettup_advance_block() says that it reports the
scan position before checking for end of scan, but that didn't match
the code. The code was refactored in commit 7ae0ab0ad9, which
inadvertently changed the order of the check and reporting. Change it
back.
This caused a few regression test failures with a small shared_buffers
setting like 10 MB. The 'portals' and 'cluster' tests perform seqscans
that are large enough that sync seqscans kick in. When the sync scan
position is not updated at end of scan, the next seq scan doesn't
start at the beginning of the table, and the test queries are
sensitive to that.
Reviewed-by: Melanie Plageman, David Rowley
Discussion: https://www.postgresql.org/message-id/6f991389-ae22-d844-a9d8-9aceb7c01a9a@iki.fi
Backpatch-through: 16
Since its introduction in 10074651e3, pg_promote() has been returning
a false status in three cases:
- SIGUSR1 not sent to the postmaster process.
- Postmaster death during standby promotion.
- Standby not promoted within the specified wait time.
An application calling this function will have a hard time understanding
what a false state returned actually means.
Per discussion, this switches the two first states to fail rather than
return a "false" status, making the second case more consistent with the
existing CHECK_FOR_INTERRUPTS in the wait loop. False is only returned
when the promotion is not completed within the specified time (60s by
default).
Author: Ashutosh Sharma
Reviewed-by: Fujii Masao, Laurenz Albe, Michael Paquier
Discussion: https://postgr.es/m/CAE9k0P=QTrwptL0t4J0fuBRDDjgsT-0PVKd-ikd96i1hyL7Bcg@mail.gmail.com
Make the primary messages more compact and make the detail messages
uniform. In initdb.c and pg_resetwal.c, use the newish
option_parse_int() to simplify some of the option parsing. For the
backend GUC wal_segment_size, add a GUC check hook to do the
verification instead of coding it in bootstrap.c. This might be
overkill, but that way the check is in the right place and it becomes
more self-documenting.
In passing, make pg_controldata use the logging API for warning
messages.
Reviewed-by: Aleksander Alekseev <aleksander@timescale.com>
Discussion: https://www.postgresql.org/message-id/flat/9939aa8a-d7be-da2c-7715-0a0b5535a1f7@eisentraut.org
Some of the ambuildempty functions used smgrwrite() directly, followed
by smgrimmedsync(). A few small problems with that:
Firstly, one is supposed to use smgrextend() when extending a
relation, not smgrwrite(). It doesn't make much difference in
production builds. smgrextend() updates the relation size cache, so
you miss that, but that's harmless because we never use the cached
relation size of an init fork. But if you compile with
CHECK_WRITE_VS_EXTEND, you get an assertion failure.
Secondly, the smgrwrite() calls were performed before WAL-logging, so
the page image written to disk had 0/0 as the LSN, not the LSN of the
WAL record. That's also harmless in practice, but seems sloppy.
Thirdly, it's better to use the buffer cache, because then you don't
need to smgrimmedsync() the relation to disk, which adds latency.
Bypassing the cache makes sense for bulk operations like index
creation, but not when you're just initializing an empty index.
Creation of unlogged tables is hardly performance bottleneck in any
real world applications, but nevertheless.
Backpatch to v16, but no further. These issues should be harmless in
practice, so better to not rock the boat in older branches.
Reviewed-by: Robert Haas
Discussion: https://www.postgresql.org/message-id/6e5bbc08-cdfc-b2b3-9e23-1a914b9850a9@iki.fi
This commit introduces descriptively-named macros for the
identifiers used in wire protocol messages. These new macros are
placed in a new header file so that they can be easily used by
third-party code.
Author: Dave Cramer
Reviewed-by: Alvaro Herrera, Tatsuo Ishii, Peter Smith, Robert Haas, Tom Lane, Peter Eisentraut, Michael Paquier
Discussion: https://postgr.es/m/CADK3HHKbBmK-PKf1bPNFoMC%2BoBt%2BpD9PH8h5nvmBQskEHm-Ehw%40mail.gmail.com
Commit 31966b15 invented a way for functions dealing with relation
extension to accept a Relation in online code and an SMgrRelation in
recovery code. It seems highly likely that future bufmgr.c interfaces
will face the same problem, and need to do something similar.
Generalize the names so that each interface doesn't have to re-invent
the wheel.
Back-patch to 16. Since extension AM authors might start using the
constructor macros once 16 ships, we agreed to do the rename in 16
rather than waiting for 17.
Reviewed-by: Peter Geoghegan <pg@bowt.ie>
Reviewed-by: Andres Freund <andres@anarazel.de>
Discussion: https://postgr.es/m/CA%2BhUKG%2B6tLD2BhpRWycEoti6LVLyQq457UL4ticP5xd8LqHySA%40mail.gmail.com
Attribute missing values might be needed past the lifetime of the tuple
descriptors from which they are extracted. To avoid possibly using
pointers for by-reference values which might thus be left dangling, we
cache a datumCopy'd version of the datum in the TopMemoryContext. Since
we first search for the value this only needs to be done once per
session for any such value.
Original complaint from Tom Lane, idea for mitigation by Andrew Dunstan,
tweaked by Tom Lane.
Backpatch to version 11 where missing values were introduced.
Discussion: https://postgr.es/m/1306569.1687978174@sss.pgh.pa.us
The new relation extension logic, introduced in 00d1e02be2, could lead to
slowdowns in some scenarios. E.g., when loading narrow rows into a table using
COPY, the caller of RelationGetBufferForTuple() will only request a small
number of pages. Without concurrency, we just extended using pwritev() in that
case. However, if there is *some* concurrency, we switched between extending
by a small number of pages and a larger number of pages, depending on the
number of waiters for the relation extension logic. However, some
filesystems, XFS in particular, do not perform well when switching between
extending files using fallocate() and pwritev().
To avoid that issue, remember the number of prior relation extensions in
BulkInsertState and extend more aggressively if there were prior relation
extensions. That not just avoids the aforementioned slowdown, but also leads
to noticeable performance gains in other situations, primarily due to
extending more aggressively when there is no concurrency. I should have done
it this way from the get go.
Reported-by: Masahiko Sawada <sawada.mshk@gmail.com>
Author: Andres Freund <andres@anarazel.de>
Discussion: https://postgr.es/m/CAD21AoDvDmUQeJtZrau1ovnT_smN940=Kp6mszNGK3bq9yRN6g@mail.gmail.com
Backpatch: 16-, where the new relation extension code was added
pg_logical_emit_message(false, '_', repeat('x', 1069547465)) failed with
self-contradictory message "WAL record would be 1069547520 bytes (of
maximum 1069547520 bytes)". There's no particular benefit from allowing
or denying one byte in either direction; XLogRecordMaxSize could rise a
few megabytes without trouble. Hence, this is just for cleanliness.
Back-patch to v16, where this check first appeared.
This commit adds a few comments about what LWLockWaitForVar() relies on
when a backend waits for a variable update on its LWLocks for WAL
insertions up to an expected LSN.
First, LWLockWaitForVar() does not include a memory barrier, relying on
a spinlock taken at the beginning of WaitXLogInsertionsToFinish(). This
was hidden behind two layers of routines in lwlock.c. This assumption
is now documented at the top of LWLockWaitForVar(), and detailed at bit
more within LWLockConflictsWithVar().
Second, document why WaitXLogInsertionsToFinish() does not include
memory barriers, relying on a spinlock at its top, which is, per Andres'
input, fine for two different reasons, both depending on the fact that
the caller of WaitXLogInsertionsToFinish() is waiting for a LSN up to a
certain value.
This area's documentation and assumptions could be improved more in the
future, but at least that's a beginning.
Author: Bharath Rupireddy, Andres Freund
Reviewed-by: Michael Paquier
Discussion: https://postgr.es/m/CALj2ACVF+6jLvqKe6xhDzCCkr=rfd6upaGc3477Pji1Ke9G7Bg@mail.gmail.com
The WAL insertion lock variable insertingAt is currently being read
and written with the help of the LWLock wait list lock to avoid any read
of torn values. This wait list lock can become a point of contention on
a highly concurrent write workloads.
This commit switches insertingAt to a 64b atomic variable that provides
torn-free reads/writes. On platforms without 64b atomic support, the
fallback implementation uses spinlocks to provide the same guarantees
for the values read. LWLockWaitForVar(), through
LWLockConflictsWithVar(), reads the new value to check if it still needs
to wait with a u64 atomic operation. LWLockUpdateVar() updates the
variable before waking up the waiters with an exchange_u64 (full memory
barrier). LWLockReleaseClearVar() now uses also an exchange_u64 to
reset the variable. Before this commit, all these steps relied on
LWLockWaitListLock() and LWLockWaitListUnlock().
This reduces contention on LWLock wait list lock and improves
performance of highly-concurrent write workloads. Here are some
numbers using pg_logical_emit_message() (HEAD at d6677b93) with various
arbitrary record lengths and clients up to 1k on a rather-large machine
(64 vCPUs, 512GB of RAM, 16 cores per sockets, 2 sockets), in terms of
TPS numbers coming from pgbench:
message_size_b | 16 | 64 | 256 | 1024
--------------------+--------+--------+--------+-------
patch_4_clients | 83830 | 82929 | 80478 | 73131
patch_16_clients | 267655 | 264973 | 250566 | 213985
patch_64_clients | 380423 | 378318 | 356907 | 294248
patch_256_clients | 360915 | 354436 | 326209 | 263664
patch_512_clients | 332654 | 321199 | 287521 | 240128
patch_1024_clients | 288263 | 276614 | 258220 | 217063
patch_2048_clients | 252280 | 243558 | 230062 | 192429
patch_4096_clients | 212566 | 213654 | 205951 | 166955
head_4_clients | 83686 | 83766 | 81233 | 73749
head_16_clients | 266503 | 265546 | 249261 | 213645
head_64_clients | 366122 | 363462 | 341078 | 261707
head_256_clients | 132600 | 132573 | 134392 | 165799
head_512_clients | 118937 | 114332 | 116860 | 150672
head_1024_clients | 133546 | 115256 | 125236 | 151390
head_2048_clients | 137877 | 117802 | 120909 | 138165
head_4096_clients | 113440 | 115611 | 120635 | 114361
Bharath has been measuring similar improvements, where the limit of the
WAL insertion lock begins to be felt when more than 256 concurrent
clients are involved in this specific workload.
An extra patch has been discussed to introduce a fast-exit path in
LWLockUpdateVar() when there are no waiters, still this does not
influence the write-heavy workload cases discussed as there are always
waiters. This will be considered separately.
Author: Bharath Rupireddy
Reviewed-by: Nathan Bossart, Andres Freund, Michael Paquier
Discussion: https://postgr.es/m/CALj2ACVF+6jLvqKe6xhDzCCkr=rfd6upaGc3477Pji1Ke9G7Bg@mail.gmail.com
This has been missed in cb0cca1, noticed before buildfarm member koel
has been able to complain while poking at a different patch. Like the
other commit, backpatch all the way down to limit the odds of merge
conflicts.
Backpatch-through: 11
A crash in the middle of a checkpoint with some two-phase state data
already flushed to disk by this checkpoint could cause a follow-up crash
recovery to recover twice the same transaction, once from what has been
found in pg_twophase/ at the beginning of recovery and a second time
when replaying its corresponding record.
This would lead to FATAL failures in the startup process during
recovery, where the same transaction would have a state recovered twice
instead of once:
LOG: recovering prepared transaction 731 from shared memory
LOG: recovering prepared transaction 731 from shared memory
FATAL: lock ExclusiveLock on object 731/0/0 is already held
This issue is fixed by skipping the addition of any 2PC state coming
from a record whose equivalent 2PC state file has already been loaded in
TwoPhaseState at the beginning of recovery by restoreTwoPhaseData(),
which is OK as long as the system has not reached a consistent state.
The timing to get a messed up recovery processing is very racy, and
would very unlikely happen. The thread that has reported the issue has
demonstrated the bug using injection points to force a PANIC in the
middle of a checkpoint.
Issue introduced in 728bd99, so backpatch all the way down.
Reported-by: "suyu.cmj" <mengjuan.cmj@alibaba-inc.com>
Author: "suyu.cmj" <mengjuan.cmj@alibaba-inc.com>
Author: Michael Paquier
Discussion: https://postgr.es/m/109e6994-b971-48cb-84f6-829646f18b4c.mengjuan.cmj@alibaba-inc.com
Backpatch-through: 11
This commit adds two columns: indexes_total and indexes_processed, to
pg_stat_progress_vacuum system view to show the index vacuum
progress. These numbers are reported in the "vacuuming indexes" and
"cleaning up indexes" phases.
This uses the new parallel message type for progress reporting added
by be06506e7.
Bump catversion because this changes the definition of
pg_stat_progress_vacuum.
Author: Sami Imseih
Reviewed by: Masahiko Sawada, Michael Paquier, Nathan Bossart, Andres Freund
Discussion: https://www.postgresql.org/message-id/flat/5478DFCD-2333-401A-B2F0-0D186AB09228@amazon.com
This commit adds a new type of parallel message 'P' to allow a
parallel worker to poke at a leader to update the progress.
Currently it supports only incremental progress reporting but it's
possible to allow for supporting of other backend progress APIs in the
future.
There are no users of this new message type as of this commit. That
will follow in future commits.
Idea from Andres Freund.
Author: Sami Imseih
Reviewed by: Michael Paquier, Masahiko Sawada
Discussion: https://www.postgresql.org/message-id/flat/5478DFCD-2333-401A-B2F0-0D186AB09228@amazon.com
Commit 19d8e2308b allowed a weaker check for HOT with summarizing
indexes, but it did not update README.HOT. So do that now.
Patch by Matthias van de Meent, minor changes by me. Backpatch to 16,
where the optimization was introduced.
Author: Matthias van de Meent
Reviewed-by: Tomas Vondra
Backpatch-through: 16
Discussion: https://postgr.es/m/CAEze2WiEOm8V+c9kUeYp2BPhbEc5s473fUf51xNeqvSFGv44Ew@mail.gmail.com
The ginfast.c code previously checked for conflicts in before locking
the relevant buffer, leaving a window where a RW conflict could be
missed. Re-order.
There was also a place where buffer ID and block number were confused
while trying to predicate-lock a page, noted by visual inspection.
Back-patch to all supported releases. Fixes one more problem discovered
with the reproducer from bug #17949, in this case when Dmitry tried
other index types.
Reported-by: Artem Anisimov <artem.anisimov.255@gmail.com>
Reported-by: Dmitry Dolgov <9erthalion6@gmail.com>
Reviewed-by: Heikki Linnakangas <hlinnaka@iki.fi>
Discussion: https://postgr.es/m/17949-a0f17035294a55e2%40postgresql.org
When performing a bitmap heap scan, we don't want to miss concurrent
writes that occurred after we observed the heap's rs_nblocks, but before
we took predicate locks on index pages. Therefore, we can't skip
fetching any heap tuples that are referenced by the index, because we
need to test them all with CheckForSerializableConflictOut(). The
old optimization that would ignore any references to blocks >=
rs_nblocks gets in the way of that requirement, because it means that
concurrent writes in that window are ignored.
Removing that optimization shouldn't affect correctness at any isolation
level, because any new tuples shouldn't be visible to an MVCC snapshot.
There also shouldn't be any error-causing references to heap blocks past
the end, because we should have held at least an AccessShareLock on the
table before the index scan. It can't get smaller while our transaction
is running. For now, though, we'll keep the optimization at lower
levels to avoid making unnecessary changes in a bug fix.
Back-patch to all supported releases. In release 11, the code is in a
different place but not fundamentally different. Fixes one aspect of
bug #17949.
Reported-by: Artem Anisimov <artem.anisimov.255@gmail.com>
Reviewed-by: Dmitry Dolgov <9erthalion6@gmail.com>
Reviewed-by: Heikki Linnakangas <hlinnaka@iki.fi>
Discussion: https://postgr.es/m/17949-a0f17035294a55e2%40postgresql.org
When predicate-locking btrees, we have a special case for completely
empty btrees, since there is no page to lock. This was racy, because,
without buffer lock held, a matching key could be inserted between the
_bt_search() and the PredicateLockRelation() calls.
Fix, by rechecking _bt_search() after taking the relation-level SIREAD
lock, if using SERIALIZABLE isolation and an empty btree is discovered.
Back-patch to all supported releases. Fixes one aspect of bug #17949.
Reported-by: Artem Anisimov <artem.anisimov.255@gmail.com>
Reviewed-by: Dmitry Dolgov <9erthalion6@gmail.com>
Reviewed-by: Heikki Linnakangas <hlinnaka@iki.fi>
Discussion: https://postgr.es/m/17949-a0f17035294a55e2%40postgresql.org
When table_block_relation_estimate_size() estimated the number of tuples
in a relation without statistics (e.g. right after load), it did not
consider fillfactor when calculating density. With non-default
fillfactor values, this may result in significant overestimate of the
number of tuples - up to 10x with the minimum 10% fillfactor. This may
have unexpected consequences, e.g. when creating hash indexes.
This considers the current fillfactor value in the "no statistics" code
path. If the fillfactor changes after loading data into the table, the
estimate may be off. But that seems much less likely than changing the
fillfactor before the data load.
Reviewed-by: Corey Huinker, Peter Eisentraut
Discussion: https://postgr.es/m/cf154ef9-6bac-d268-b735-67a3443debba@enterprisedb.com
