The parallel GIN builds perform "freezing" of TID lists when merging
chunks built earlier. This means determining what part of the list can
no longer change, depending on the last received chunk. The frozen part
can be evicted from memory and written out.
The code attempted to freeze items right before merging the old and new
TID list, after already attempting to trim the current buffer. That
means part of the data may get frozen based on the new TID list, but
will be trimmed later (on next loop). This increases memory usage.
This inverts the order, so that we freeze data first (before trimming).
The benefits are likely relatively small, but it's also virtually free
with no other downsides.
Discussion: https://postgr.es/m/CAHLJuCWDwn-PE2BMZE4Kux7x5wWt_6RoWtA0mUQffEDLeZ6sfA@mail.gmail.com
When building intermediate TID lists during parallel GIN builds, split
the sorted lists into smaller chunks, to limit the amount of memory
needed when merging the chunks later.
The leader may need to keep in memory up to one chunk per worker, and
possibly one extra chunk (before evicting some of the data). The code
processing item pointers uses regular palloc/repalloc calls, which means
it's subject to the MaxAllocSize (1GB) limit.
We could fix this by allowing huge allocations, but that'd require
changes in many places without much benefit. Larger chunks do not
actually improve performance, so the memory usage would be wasted.
Fixed by limiting the chunk size to not hit MaxAllocSize. Each worker
gets a fair share.
This requires remembering the number of participating workers, in a
place that can be accessed from the callback. Luckily, the bs_worker_id
field in GinBuildState was unused, so repurpose that.
Report by Greg Smith, investigation and fix by me. Batchpatched to 18,
where parallel GIN builds were introduced.
Reported-by: Gregory Smith <gregsmithpgsql@gmail.com>
Discussion: https://postgr.es/m/CAHLJuCWDwn-PE2BMZE4Kux7x5wWt_6RoWtA0mUQffEDLeZ6sfA@mail.gmail.com
Backpatch-through: 18
When the startup process switches from streaming to archive as WAL
source, we avoid calling ShutdownWalRcv() if the WAL receiver is not
streaming, based on WalRcvStreaming(). WALRCV_STOPPING is a state set
by ShutdownWalRcv(), called only by the startup process, meaning that it
should not be possible to reach this state while in
WaitForWALToBecomeAvailable().
This commit adds an assertion to make sure that a WAL receiver is never
in a WALRCV_STOPPING state should the startup process attempt to reset
InstallXLogFileSegmentActive.
Idea suggested by Noah Misch.
Author: Xuneng Zhou <xunengzhou@gmail.com>
Discussion: https://postgr.es/m/19093-c4fff49a608f82a0@postgresql.org
Commit b4f584f9d2 (affecting v15~, later backpatched down to 13 as of
3635a0a35a) introduced an unconditional WAL receiver shutdown when
switching from streaming to archive WAL sources. This causes problems
during a timeline switch, when a WAL receiver enters WALRCV_WAITING
state but remains alive, waiting for instructions.
The unconditional shutdown can break some monitoring scenarios as the
WAL receiver gets repeatedly terminated and re-spawned, causing
pg_stat_wal_receiver.status to show a "streaming" instead of "waiting"
status, masking the fact that the WAL receiver is waiting for a new TLI
and a new LSN to be able to continue streaming.
This commit changes the WAL receiver behavior so as the shutdown becomes
conditional, with InstallXLogFileSegmentActive being always reset to
prevent the regression fixed by b4f584f9d2: only terminate the WAL
receiver when it is actively streaming (WALRCV_STREAMING,
WALRCV_STARTING, or WALRCV_RESTARTING). When in WALRCV_WAITING state,
just reset InstallXLogFileSegmentActive flag to allow archive
restoration without killing the process. WALRCV_STOPPED and
WALRCV_STOPPING are not reachable states in this code path. For the
latter, the startup process is the one in charge of setting
WALRCV_STOPPING via ShutdownWalRcv(), waiting for the WAL receiver to
reach a WALRCV_STOPPED state after switching walRcvState, so
WaitForWALToBecomeAvailable() cannot be reached while a WAL receiver is
in a WALRCV_STOPPING state.
A regression test is added to check that a WAL receiver is not stopped
on timeline jump, that fails when the fix of this commit is reverted.
Reported-by: Ryan Bird <ryanzxg@gmail.com>
Author: Xuneng Zhou <xunengzhou@gmail.com>
Reviewed-by: Noah Misch <noah@leadboat.com>
Reviewed-by: Michael Paquier <michael@paquier.xyz>
Discussion: https://postgr.es/m/19093-c4fff49a608f82a0@postgresql.org
Backpatch-through: 13
Add comments explaining when and where it is safe for nbtree to treat
row compare keys as if they were simple scalar inequality keys on the
row's most significant column. This is particularly important within
_bt_advance_array_keys, which deals with required inequality keys in a
general and uniform way, without any special handling for row compares.
Also spell out the implications of _bt_check_rowcompare's approach of
_conditionally_ evaluating lower-order row compare subkeys, particularly
when one of its lower-order subkeys might see NULL index tuple values
(these may or may not affect whether the qual as a whole is satisfied).
The behavior in this area isn't particularly intuitive, so these issues
seem worth going into.
In passing, add a few more defensive/documenting row comparison related
assertions to _bt_first and _bt_check_rowcompare.
Follow-up to commits bd3f59fd and ec986020.
Author: Peter Geoghegan <pg@bowt.ie>
Reviewed-By: Victor Yegorov <vyegorov@gmail.com>
Reviewed-By: Chao Li <li.evan.chao@gmail.com>
Discussion: https://postgr.es/m/CAH2-Wznwkak_K7pcAdv9uH8ZfNo8QO7+tHXOaCUddMeTfaCCFw@mail.gmail.com
Backpatch-through: 18
_bt_first reliably uses the same equality key (on each index column) for
initial positioning purposes as the one that _bt_checkkeys can use to
end the scan following commit f09816a0. _bt_first no longer applies its
own independent rules to determine which initial positioning key to use
on each column (for equality and inequality keys alike). Preprocessing
is now fully in control of determining which keys start and end each
scan, ensuring that _bt_first and _bt_checkkeys have symmetric behavior.
Remove obsolete comments that described why _bt_first was expected to
use at least one of the available required equality keys for initial
positioning purposes. The rules in this area are now maximally strict
and uniform, so there's no reason to draw attention to equality keys.
Any column with a required equality key cannot have a redundant required
inequality key (nor can it have a redundant required equality key).
Oversight in commit f09816a0, which removed similar comments from
_bt_first, but missed these comments.
Author: Peter Geoghegan <pg@bowt.ie>
Backpatch-through: 18
Several functions in the codebase accept "Datum *" parameters but do
not modify the pointed-to data. These have been updated to take
"const Datum *" instead, improving type safety and making the
interfaces clearer about their intent. This change helps the compiler
catch accidental modifications and better documents immutability of
arguments.
Most of "Datum *" parameters have a pairing "bool *isnull" parameter,
they are constified as well.
No functional behavior is changed by this patch.
Author: Chao Li <lic@highgo.com>
Discussion: https://www.postgresql.org/message-id/flat/CAEoWx2msfT0knvzUa72ZBwu9LR_RLY4on85w2a9YpE-o2By5HQ@mail.gmail.com
XLogRecordAssemble() may be called multiple times before inserting a
record in XLogInsertRecord(), and the amount of FPIs generated inside
a record whose insertion is attempted multiple times may vary.
The logic added in f9a09aa295 touched directly pgWalUsage in
XLogRecordAssemble(), meaning that it could be possible for pgWalUsage
to be incremented multiple times for a single record. This commit
changes the code to use the same logic as the number of FPIs added to a
record, where XLogRecordAssemble() returns this information and feeds it
to XLogInsertRecord(), updating pgWalUsage only when a record is
inserted.
Reported-by: Shinya Kato <shinya11.kato@gmail.com>
Discussion: https://postgr.es/m/CAOzEurSiSr+rusd0GzVy8Bt30QwLTK=ugVMnF6=5WhsSrukvvw@mail.gmail.com
This new counter, called "wal_fpi_bytes", tracks the total amount in
bytes of full page images (FPIs) generated in WAL. This data becomes
available globally via pg_stat_wal, and for backend statistics via
pg_stat_get_backend_wal().
Previously, this information could only be retrieved with pg_waldump or
pg_walinspect, which may not be available depending on the environment,
and are expensive to execute. It offers hints about how much FPIs
impact the WAL generated, which could be a large percentage for some
workloads, as well as the effects of wal_compression or page holes.
Bump catalog version.
Bump PGSTAT_FILE_FORMAT_ID, due to the addition of wal_fpi_bytes in
PgStat_WalCounters.
Author: Shinya Kato <shinya11.kato@gmail.com>
Reviewed-by: Michael Paquier <michael@paquier.xyz>
Discussion: https://postgr.es/m/CAOzEurQtZEAfg6P0kU3Wa-f9BWQOi0RzJEMPN56wNTOmJLmfaQ@mail.gmail.com
This type is just char * underneath, it provides no real value, no
type safety, and just makes the code one level more mysterious. It is
more idiomatic to refer to blobs of memory by a combination of void *
and size_t, so change it to that.
Also, since this type hides the pointerness, we can't apply qualifiers
to what is pointed to, which requires some unconstify nonsense. This
change allows fixing that.
Extension code that uses the Item type can change its code to use
void * to be backward compatible.
Reviewed-by: Nathan Bossart <nathandbossart@gmail.com>
Reviewed-by: Peter Geoghegan <pg@bowt.ie>
Discussion: https://www.postgresql.org/message-id/flat/c75cccf5-5709-407b-a36a-2ae6570be766@eisentraut.org
Previously, if primary_slot_name was set to an invalid slot name and
the configuration file was reloaded, both the postmaster and all other
backend processes reported a WARNING. With many processes running,
this could produce a flood of duplicate messages. The problem was that
the GUC check hook for primary_slot_name reported errors at WARNING
level via ereport().
This commit changes the check hook to use GUC_check_errdetail() and
GUC_check_errhint() for error reporting. As with other GUC parameters,
this causes non-postmaster processes to log the message at DEBUG3,
so by default, only the postmaster's message appears in the log file.
Backpatch to all supported versions.
Author: Fujii Masao <masao.fujii@gmail.com>
Reviewed-by: Chao Li <lic@highgo.com>
Reviewed-by: Amit Kapila <amit.kapila16@gmail.com>
Reviewed-by: Álvaro Herrera <alvherre@kurilemu.de>
Reviewed-by: Hayato Kuroda <kuroda.hayato@fujitsu.com>
Discussion: https://postgr.es/m/CAHGQGwFud-cvthCTfusBfKHBS6Jj6kdAPTdLWKvP2qjUX6L_wA@mail.gmail.com
Backpatch-through: 13
5983a4cff added CompactAttribute for storing commonly used fields from
FormData_pg_attribute. 5983a4cff didn't go to the trouble of adjusting
every location where we can use CompactAttribute rather than
FormData_pg_attribute, so here we change the remaining ones.
There are some locations where I've left the code using
FormData_pg_attribute. These are mostly in the ALTER TABLE code. Using
CompactAttribute here seems more risky as often the TupleDesc is being
changed and those changes may not have been flushed to the
CompactAttribute yet.
I've also left record_recv(), record_send(), record_cmp(), record_eq()
and record_image_eq() alone as it's not clear to me that accessing the
CompactAttribute is a win here due to the FormData_pg_attribute still
having to be accessed for most cases. Switching the relevant parts to
use CompactAttribute would result in having to access both for common
cases. Careful benchmarking may reveal that something can be done to
make this better, but in absence of that, the safer option is to leave
these alone.
In ReorderBufferToastReplace(), there was a check to skip attnums < 0
while looping over the TupleDesc. Doing this is redundant since
TupleDescs don't store < 0 attnums. Removing that code allows us to
move to using CompactAttribute.
The change in validateDomainCheckConstraint() just moves fetching the
FormData_pg_attribute into the ERROR path, which is cold due to calling
errstart_cold() and results in code being moved out of the common path.
Author: David Rowley <dgrowleyml@gmail.com>
Reviewed-by: Michael Paquier <michael@paquier.xyz>
Discussion: https://postgr.es/m/CAApHDvrMy90o1Lgkt31F82tcSuwRFHq3vyGewSRN=-QuSEEvyQ@mail.gmail.com
A BlockNumber (32-bit) might not be large enough to add bo_pagesPerRange
to when the table contains close to 2^32 pages. At worst, this could
result in a cancellable infinite loop during the BRIN index scan with
power-of-2 pagesPerRange, and slow (inefficient) BRIN index scans and
scanning of unneeded heap blocks for non power-of-2 pagesPerRange.
Backpatch to all supported versions.
Author: sunil s <sunilfeb26@gmail.com>
Reviewed-by: David Rowley <dgrowleyml@gmail.com>
Reviewed-by: Michael Paquier <michael@paquier.xyz>
Discussion: https://postgr.es/m/CAOG6S4-tGksTQhVzJM19NzLYAHusXsK2HmADPZzGQcfZABsvpA@mail.gmail.com
Backpatch-through: 13
The log output functionality of log_autovacuum_min_duration applies to
both VACUUM and ANALYZE, so it is not possible to separate the VACUUM
and ANALYZE log output thresholds. Logs are likely to be output only for
VACUUM and not for ANALYZE.
Therefore, we decided to separate the threshold for log output of VACUUM
by autovacuum (log_autovacuum_min_duration) and the threshold for log
output of ANALYZE by autovacuum (log_autoanalyze_min_duration).
Author: Shinya Kato <shinya11.kato@gmail.com>
Reviewed-by: Kasahara Tatsuhito <kasaharatt@oss.nttdata.com>
Discussion: https://www.postgresql.org/message-id/flat/CAOzEurQtfV4MxJiWT-XDnimEeZAY+rgzVSLe8YsyEKhZcajzSA@mail.gmail.com
After scanning the line pointers on a heap page during the first phase
of vacuum, we use the information collected to decide whether to use
the assembled freeze plans.
Move this decision logic into a helper function to improve readability.
While here, rename a PruneState member and disambiguate some local
variables in heap_page_prune_and_freeze().
Author: Melanie Plageman <melanieplageman@gmail.com>
Reviewed-by: Andres Freund <andres@anarazel.de>
Discussion: https://postgr.es/m/2wk7jo4m4qwh5sn33pfgerdjfujebbccsmmlownybddbh6nawl%40mdyyqpqzxjek
Instead of emitting a separate XLOG_HEAP2_VISIBLE WAL record for each
page that becomes all-visible in vacuum's third phase, specify the VM
changes in the already emitted XLOG_HEAP2_PRUNE_VACUUM_CLEANUP record.
Visibility checks are now performed before marking dead items unused.
This is safe because the heap page is held under exclusive lock for the
entire operation.
This reduces the number of WAL records generated by VACUUM phase III by
up to 50%.
Author: Melanie Plageman <melanieplageman@gmail.com>
Reviewed-by: Andres Freund <andres@anarazel.de>
Reviewed-by: Robert Haas <robertmhaas@gmail.com>
Reviewed-by: Kirill Reshke <reshkekirill@gmail.com>
Discussion: https://postgr.es/m/flat/CAAKRu_ZMw6Npd_qm2KM%2BFwQ3cMOMx1Dh3VMhp8-V7SOLxdK9-g%40mail.gmail.com
Remove a variable that is no longer in use following commit 9a2e2a28.
It's not immediately clear why there were no compiler warnings about
this oversight.
Author: Peter Geoghegan <pg@bowt.ie>
Backpatch-through: 18
An assertion in _bt_killitems expected the scan's currPos state to
contain a valid LSN, saved from when currPos's page was initially read.
The assertion failed to account for the fact that even logged relations
can have leaf pages with an invalid LSN when built with wal_level set to
"minimal". Remove the faulty assertion.
Oversight in commit e6eed40e (though note that the assertion was
backpatched to stable branches before 18 by commit 7c319f54).
Author: Peter Geoghegan <pg@bowt.ie>
Reported-By: Matthijs van der Vleuten <postgresql@zr40.nl>
Bug: #19082
Discussion: https://postgr.es/m/19082-628e62160dbbc1c1@postgresql.org
Backpatch-through: 13
We're planning to merge buffer content locks into BufferDesc.state. To reduce
the size of that patch, centralize calls to BufferDescriptorGetContentLock().
The biggest part of the change is in assertions, by introducing
BufferIsLockedByMe[InMode]() (and removing BufferIsExclusiveLocked()). This
seems like an improvement even without aforementioned plans.
Additionally replace some direct calls to LWLockAcquire() with calls to
LockBuffer().
Reviewed-by: Matthias van de Meent <boekewurm+postgres@gmail.com>
Discussion: https://postgr.es/m/fvfmkr5kk4nyex56ejgxj3uzi63isfxovp2biecb4bspbjrze7@az2pljabhnff
Previously, we attempted to form a posting list tuple even when
ginCompressPostingList() failed to compress the posting list due to
its size. While there was no functional failure, it always wasted one
GinFormTuple() call when item pointers didn't fit in a posting list
tuple.
This commit ensures that a GIN index tuple is formed only when all
item pointers in the posting list are successfully compressed.
Author: Arseniy Mukhin <arseniy.mukhin.dev@gmail.com>
Reviewed-by: Masahiko Sawada <sawada.mshk@gmail.com>
Discussion: https://postgr.es/m/CAE7r3M+C=jcpTD93f_RBHrQp3C+=TAXFs+k4tTuZuuxboK8AvA@mail.gmail.com
The WAL records XLOG_GIN_INSERT and XLOG_GIN_VACUUM_DATA_LEAF_PAGE
included some information about the blocks added to the record.
This information is already provided by XLogRecGetBlockRefInfo() with
much more details about the blocks included in each record, like the
compression information, for example. This commit removes the block
information that existed in the record descriptions specific to GIN.
Author: Kirill Reshke <reshkekirill@gmail.com>
Reviewed-by: Andrey Borodin <x4mmm@yandex-team.ru>
Discussion: https://postgr.es/m/CALdSSPgk=9WRoXhZy5fdk+T1hiau7qbL_vn94w_L1N=gtEdbsg@mail.gmail.com
In similar vein to commit ccc8194e42, a reset instance of a shared
memory TID store happened to occupy the same private memory as the old
one for the entry point, since the chunk freed after the last round
of index vacuuming was put on the context's freelist. The failure
to update the vacrel->dead_items pointer was evident by nudging the
system to allocate memory in a different area. This was not discovered
at the time of the earlier commit since our regression tests didn't
cover multiple index passes with parallel vacuum.
Backpatch to v17, when TidStore came in.
Author: Kevin Oommen Anish <kevin.o@zohocorp.com>
Reviewed-by: Richard Guo <guofenglinux@gmail.com>
Tested-by: Richard Guo <guofenglinux@gmail.com>
Discussion: https://postgr.es/m/199a07cbdfc.7a1c4aac25838.1675074408277594551%40zohocorp.com
Backpatch-through: 17
During recovery, XLogNeedsFlush() checks the minimum recovery LSN point
instead of the flush LSN point. The same condition checks are used when
updating the minimum recovery point in UpdateMinRecoveryPoint(), but are
written in reverse order.
This commit makes the order of the checks consistent between
XLogNeedsFlush() and UpdateMinRecoveryPoint(), improving the code
clarity. Note that the second check (as ordered by this commit) relies
on InRecovery, which is true only in the startup process. So this makes
XLogNeedsFlush() cheaper in the startup process with the first check
acting as a shortcut while doing crash recovery, where
LocalMinRecoveryPoint is an invalid LSN.
Author: Melanie Plageman <melanieplageman@gmail.com>
Reviewed-by: Dilip Kumar <dilipbalaut@gmail.com>
Discussion: https://postgr.es/m/aMIHNRTP6Wj6vw1s%40paquier.xyz
This improves the stability of VACUUM when processing btree indexes,
which was previously able to trigger an assertion failure in
_bt_lock_subtree_parent() when an error was previously thrown outside
the scope of _bt_split() when splitting a btree page. VACUUM would
consider the index as in a corrupted state as the right page would not
be zeroed for the error thrown (allocation failure is one pattern).
In a non-assert build, VACUUM is able to succeed, reporting what it sees
as a corruption while attempting to fix the index. This would manifest
as a LOG message, as of:
LOG: failed to re-find parent key in index "idx" for deletion target
page N
CONTEXT: while vacuuming index "idx" of relation "public.tab"
This commit improves the code to rely on two PGAlignedBlocks that are
used as a temporary space for the left and right pages. The main change
concerns the right page, whose contents are now copied into the
"temporary" PGAlignedBlock page while its original space is zeroed. Its
contents are moved from the PGAlignedBlock page back to the page once we
enter in the critical section used for the split. This simplifies the
split logic, as it is not necessary to zero the right page before
throwing an error anymore. Hence errors can now be thrown outside the
split code. For the left page, this shaves one allocation, with
PageGetTempPage() being previously used.
The previous logic originates from commit 8fa30f906b, at a point where
PGAlignedBlock did not exist yet. This could be argued as something
that should be backpatched, but the lack of complaints indicates that it
may not be necessary.
Author: Konstantin Knizhnik <knizhnik@garret.ru>
Discussion: https://postgr.es/m/566dacaf-5751-47e4-abc6-73de17a5d42a@garret.ru
... which silently propagates a lot of headers into many places
via pgstat.h, as evidenced by the variety of headers that this patch
needs to add to seemingly random places. Add a minimum of typedefs to
conflict.h to be able to remove execnodes.h, and fix the fallout.
Backpatch to 18, where conflict.h first appeared.
Discussion: https://postgr.es/m/202509191927.uj2ijwmho7nv@alvherre.pgsql
This commit corrects several issues in function comments:
* The parameter "rel" was incorrectly referred to as "relation" in the comments
for table_tuple_delete(), table_tuple_update(), and table_tuple_lock().
* In table_tuple_delete(), "changingPart" was listed as an output parameter
in the comments but is actually input.
* In table_tuple_update(), "slot" was listed as an input parameter
in the comments but is actually output.
* The comment for "update_indexes" in table_tuple_update() was mis-indented.
* The comments for heap_lock_tuple() incorrectly referenced a non-existent
"tid" parameter.
Author: Chao Li <lic@highgo.com>
Reviewed-by: Fujii Masao <masao.fujii@gmail.com>
Discussion: https://postgr.es/m/CAEoWx2nB6Ay8g=KEn7L3qbYX_4+sLk9XOMkV0XZqHR4cTY8ZvQ@mail.gmail.com
When deciding which code path to use depending on the state of recovery,
XLogFlush() and XLogNeedsFlush() have been relying on different
criterias:
- XLogFlush() relied on XLogInsertAllowed().
- XLogNeedsFlush() relied on RecoveryInProgress().
Currently, the checkpointer is allowed to insert WAL records while
RecoveryInProgress() returns true for an end-of-recovery checkpoint,
where XLogInsertAllowed() matters. Using RecoveryInProgress() in
XLogNeedsFlush() did not really matter for its existing callers, as the
checkpointer only called XLogFlush(). However, a feature under
discussion, by Melanie Plageman, needs XLogNeedsFlush() to be able to
work in more contexts, the end-of-recovery checkpoint being one.
This commit changes XLogNeedsFlush() to use XLogInsertAllowed() instead
of RecoveryInProgress(), making the checks in both routines more
consistent. While on it, an assertion based on XLogNeedsFlush() is
added at the end of XLogFlush(), triggered when flushing a physical
position (not for the normal recovery patch that checks for updates of
the minimum recovery point). This assertion would fail for example in
the recovery test 015_promotion_pages if XLogNeedsFlush() is changed to
use RecoveryInProgress(). This should be hopefully enough to ensure
that the checks done in both routines remain consistent.
Author: Melanie Plageman <melanieplageman@gmail.com>
Co-authored-by: Dilip Kumar <dilipbalaut@gmail.com>
Reviewed-by: Jeff Davis <pgsql@j-davis.com>
Reviewed-by: Chao Li <li.evan.chao@gmail.com>
Reviewed-by: Michael Paquier <michael@paquier.xyz>
Discussion: https://postgr.es/m/CAAKRu_a1vZRZRWO3_jv_X13RYoqLRVipGO0237g5PKzPa2YX6g@mail.gmail.com
Add logic to _bt_set_startikey that determines whether row compare keys
are guaranteed to be satisfied by every tuple on a page that is about to
be read by _bt_readpage. This works in essentially the same way as the
existing scalar inequality logic. Testing has shown that the new logic
improves performance to about the same degree as the existing scalar
inequality logic (compared to the unoptimized case). In other words,
the new logic makes many row compare scans significantly faster.
Note that the new row compare inequality logic is only effective when
the same individual row member is the deciding subkey for all tuples on
the page (obviously, all tuples have to satisfy the row compare, too).
This is what makes the new row compare logic very similar to the
existing logic for scalar inequalities. Note, in particular, that this
makes it safe to ignore whether all row compare members are against
either ASC or DESC index attributes (i.e. it doesn't matter if
individual subkeys don't all use the same inequality strategy).
Also stop refusing to set pstate.startikey to an offset beyond any
nonrequired key (don't add logic that'll do that for an individual row
compare subkey, either). We can fully rely on our firstchangingattnum
tests instead. This will do the right thing when a page has a group of
tuples with NULLs in a lower-order attribute that makes the tuples fail
to satisfy a row compare key -- we won't incorrectly conclude that all
tuples must satisfy the row compare, just because firsttup and lasttup
happen to. Our firstchangingattnum test prevents that from happening.
(Note that the original "avoid evaluating nbtree scan keys" mechanism
added by commit e0b1ee17 couldn't support row compares due to issues
with tuples that contain NULLs in a lower-order subkey's attribute.
That original mechanism relied on requiredness markings, which the
replacement _bt_set_startikey mechanism never really needed.)
Follow up to commit 8a510275, which added the _bt_set_startikey
optimization. _bt_set_startikey is now feature complete; there's no
remaining kind of nbtree scan key that it still doesn't support.
Author: Peter Geoghegan <pg@bowt.ie>
Reviewed-By: Chao Li <li.evan.chao@gmail.com>
Discussion: https://postgr.es/m/CAH2-WznL6Z3H_GTQze9d8T_Ls=cYbnd-_9f-Jo7aYgTGRUD58g@mail.gmail.com
The TimescaleDB extension expects to be able to change an nbtree scan's
keys across rescans. The issue arises in the extension's implementation
of loose index scan. This is arguably a misuse of the index AM API,
though apparently it worked until recently. It stopped working when the
skipScan flag was added to BTScanOpaqueData by commit 8a510275, though.
The flag wouldn't reliably track whether the scan (actually, the current
rescan) has any skip arrays, leading to confusion in _bt_set_startikey.
nbtree preprocessing will now defensively initialize the scan's skipScan
flag in all cases, including the case where _bt_preprocess_array_keys
returns early due to the (re)scan not using arrays. While nbtree isn't
obligated to support this use case (at least not according to my reading
of the index AM API), it still seems like a good idea to be consistent
here, on general robustness grounds.
Author: Peter Geoghegan <pg@bowt.ie>
Reported-By: Natalya Aksman <natalya@timescale.com>
Discussion: https://postgr.es/m/CAJumhcirfMojbk20+W0YimbNDkwdECvJprQGQ-XqK--ph09nQw@mail.gmail.com
Backpatch-through: 18
