An inplace update's invalidation messages are part of its transaction's
commit record. However, the update survives even if its transaction
aborts or we stop recovery before replaying its transaction commit.
After recovery, a backend that started in recovery could update the row
without incorporating the inplace update. That could result in a table
with an index, yet relhasindex=f. That is a source of index corruption.
This bulk invalidation avoids the functional consequences. A future
change can fix the !RecoveryInProgress() scenario without changing the
WAL format. Back-patch to v17 - v12 (all supported versions). v18 will
instead add invalidations to WAL.
Discussion: https://postgr.es/m/20240618152349.7f.nmisch@google.com
The previous commit fixed some ways of losing an inplace update. It
remained possible to lose one when a backend working toward a
heap_update() copied a tuple into memory just before inplace update of
that tuple. In catalogs eligible for inplace update, use LOCKTAG_TUPLE
to govern admission to the steps of copying an old tuple, modifying it,
and issuing heap_update(). This includes MERGE commands. To avoid
changing most of the pg_class DDL, don't require LOCKTAG_TUPLE when
holding a relation lock sufficient to exclude inplace updaters.
Back-patch to v12 (all supported versions). In v13 and v12, "UPDATE
pg_class" or "UPDATE pg_database" can still lose an inplace update. The
v14+ UPDATE fix needs commit 86dc90056d,
and it wasn't worth reimplementing that fix without such infrastructure.
Reviewed by Nitin Motiani and (in earlier versions) Heikki Linnakangas.
Discussion: https://postgr.es/m/20231027214946.79.nmisch@google.com
Commit 5b562644fe added a comment that
SetRelationHasSubclass() callers must hold this lock. When commit
17f206fbc8 extended use of this column to
partitioned indexes, it didn't take the lock. As the latter commit
message mentioned, we currently never reset a partitioned index to
relhassubclass=f. That largely avoids harm from the lock omission. The
cause for fixing this now is to unblock introducing a rule about locks
required to heap_update() a pg_class row. This might cause more
deadlocks. It gives minor user-visible benefits:
- If an ALTER INDEX SET TABLESPACE runs concurrently with ALTER TABLE
ATTACH PARTITION or CREATE PARTITION OF, one transaction blocks
instead of failing with "tuple concurrently updated". (Many cases of
DDL concurrency still fail that way.)
- Match ALTER INDEX ATTACH PARTITION in choosing to lock the index.
While not user-visible today, we'll need this if we ever make something
set the flag to false for a partitioned index, like ANALYZE does today
for tables. Back-patch to v12 (all supported versions), the plan for
the commit relying on the new rule. In back branches, add
LockOrStrongerHeldByMe() instead of adding a LockHeldByMe() parameter.
Reviewed (in an earlier version) by Robert Haas.
Discussion: https://postgr.es/m/20240611024525.9f.nmisch@google.com
We did not recover the subtransaction IDs of prepared transactions
when starting a hot standby from a shutdown checkpoint. As a result,
such subtransactions were considered as aborted, rather than
in-progress. That would lead to hint bits being set incorrectly, and
the subtransactions suddenly becoming visible to old snapshots when
the prepared transaction was committed.
To fix, update pg_subtrans with prepared transactions's subxids when
starting hot standby from a shutdown checkpoint. The snapshots taken
from that state need to be marked as "suboverflowed", so that we also
check the pg_subtrans.
Backport to all supported versions.
Discussion: https://www.postgresql.org/message-id/6b852e98-2d49-4ca1-9e95-db419a2696e0@iki.fi
If temp tables have dependencies (such as sequences) then it's
possible for autovacuum's cleanup of orphan temp tables to deadlock
against an incoming backend that's trying to clean out the temp
namespace for its own use. That can happen because RemoveTempRelations'
performDeletion call can visit objects within the namespace in
an order different from the order in which a per-table deletion
will visit them.
To fix, observe that performDeletion will begin by taking an exclusive
lock on the temp namespace (even though it won't actually delete it).
So, if we can get a shared lock on the namespace, we can be sure we're
not running concurrently with RemoveTempRelations, while also not
conflicting with ordinary use of the namespace. This requires
introducing a conditional version of LockDatabaseObject, but that's no
big deal. (It's surprising we've got along without that this long.)
Report and patch by Mikhail Zhilin. Back-patch to all supported
branches.
Discussion: https://postgr.es/m/c43ce028-2bc2-4865-9b89-3f706246eed5@postgrespro.ru
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
Commit b91dd9de was concerned with a theoretical problem with our
non-atomic condition variable operations. If you stop sleeping, and
then cancel the sleep in a separate step, you might be signaled in
between, and that could be lost. That doesn't matter for callers of
ConditionVariableBroadcast(), but callers of ConditionVariableSignal()
might be upset if a signal went missing like this.
Commit bc971f4025 interacted badly with that logic, because it doesn't
use ConditionVariableSleep(), which would normally put us back in the
wait list. ConditionVariableCancelSleep() would be confused and think
we'd received an extra signal, and try to forward it to another backend,
resulting in wakeup storms.
New idea: ConditionVariableCancelSleep() can just return true if we've
been signaled. Hypothetical users of ConditionVariableSignal() would
then still have a way to deal with rare lost signals if they are
concerned about that problem.
Back-patch to 16, where bc971f4025 arrived.
Reported-by: Tomas Vondra <tomas.vondra@enterprisedb.com>
Reviewed-by: Andres Freund <andres@anarazel.de>
Discussion: https://postgr.es/m/2840876b-4cfe-240f-0a7e-29ffd66711e7%40enterprisedb.com
Run pgindent, pgperltidy, and reformat-dat-files.
This set of diffs is a bit larger than typical. We've updated to
pg_bsd_indent 2.1.2, which properly indents variable declarations that
have multi-line initialization expressions (the continuation lines are
now indented one tab stop). We've also updated to perltidy version
20230309 and changed some of its settings, which reduces its desire to
add whitespace to lines to make assignments etc. line up. Going
forward, that should make for fewer random-seeming changes to existing
code.
Discussion: https://postgr.es/m/20230428092545.qfb3y5wcu4cm75ur@alvherre.pgsql
28e626bde0 added the concept of IOOps but neglected to include writeback
operations. ac8d53dae5 added time spent doing these I/O operations. Without
counting writeback, checkpointer write time in the log often differed
substantially from that in pg_stat_io. To fix this, add IOOp IOOP_WRITEBACK
and track writeback in pg_stat_io.
Bumps catversion.
Author: Melanie Plageman <melanieplageman@gmail.com>
Reviewed-by: Kyotaro Horiguchi <horikyota.ntt@gmail.com>
Reported-by: Andres Freund <andres@anarazel.de>
Discussion: https://postgr.es/m/20230419172326.dhgyo4wrrhulovt6%40awork3.anarazel.de
For clarity of review, renaming the function parameter "context" in
ScheduleBufferTagForWriteback() and IssuePendingWritebacks() to
"wb_context" is a separate commit. The next commit adds an "io_context"
parameter and "wb_context" makes it more clear which is which.
Author: Melanie Plageman <melanieplageman@gmail.com>
Discussion: https://postgr.es/m/CAAKRu_acc6iL4M3hvOTeztf_ZPpsB3Pqio5aVHgZ5q=Pi3BZKg@mail.gmail.com
The recently added inclusion of guc.h in smgr.h is not necessary and
introduces more server-related stuff. Removing the directive helps
avoid potential issues with including sgmr.h in frontends.
Author: Kyotaro Horiguchi <horikyota.ntt@gmail.com>
Discussion: https://postgr.es/m/20230425.115748.2130383825066921512.horikyota.ntt%40gmail.com
vacuum_defer_cleanup_age was introduced before hot_standby_feedback and
replication slots existed. It is hard to use reasonably - commonly it will
either be set too low (not preventing recovery conflicts, while still causing
some bloat), or too high (causing a lot of bloat). The alternatives do not
have that issue.
That on its own might not be sufficient reason to remove
vacuum_defer_cleanup_age, but it also complicates computation of xid
horizons. See e.g. the bug fixed in be504a3e97. It also is untested.
This commit removes TransactionIdRetreatSafely(), as there are no users
anymore. There might be potential future users, hence noting that here.
Reviewed-by: Daniel Gustafsson <daniel@yesql.se>
Reviewed-by: Justin Pryzby <pryzby@telsasoft.com>
Reviewed-by: Alvaro Herrera <alvherre@alvh.no-ip.org>
Discussion: https://postgr.es/m/20230317230930.nhsgk3qfk7f4axls@awork3.anarazel.de
Make sure that function declarations use names that exactly match the
corresponding names from function definitions in a few places. These
inconsistencies were all introduced relatively recently, after the code
base had parameter name mismatches fixed in bulk (see commits starting
with commits 4274dc22 and 035ce1fe).
pg_bsd_indent still has a couple of similar inconsistencies, which I
(pgeoghegan) have left untouched for now.
Like all earlier commits that cleaned up function parameter names, this
commit was written with help from clang-tidy.
During WAL replay on the standby, when a conflict with a logical slot is
identified, invalidate such slots. There are two sources of conflicts:
1) Using the information added in 6af1793954, logical slots are invalidated if
required rows are removed
2) wal_level on the primary server is reduced to below logical
Uses the infrastructure introduced in the prior commit. FIXME: add commit
reference.
Change InvalidatePossiblyObsoleteSlot() to use a recovery conflict to
interrupt use of a slot, if called in the startup process. The new recovery
conflict is added to pg_stat_database_conflicts, as confl_active_logicalslot.
See 6af1793954 for an overall design of logical decoding on a standby.
Bumps catversion for the addition of the pg_stat_database_conflicts column.
Bumps PGSTAT_FILE_FORMAT_ID for the same reason.
Author: "Drouvot, Bertrand" <bertranddrouvot.pg@gmail.com>
Author: Andres Freund <andres@anarazel.de>
Author: Amit Khandekar <amitdkhan.pg@gmail.com> (in an older version)
Reviewed-by: "Drouvot, Bertrand" <bertranddrouvot.pg@gmail.com>
Reviewed-by: Andres Freund <andres@anarazel.de>
Reviewed-by: Robert Haas <robertmhaas@gmail.com>
Reviewed-by: Fabrízio de Royes Mello <fabriziomello@gmail.com>
Reviewed-by: Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>
Reviewed-by: Amit Kapila <amit.kapila16@gmail.com>
Reviewed-by: Alvaro Herrera <alvherre@alvh.no-ip.org>
Discussion: https://postgr.es/m/20230407075009.igg7be27ha2htkbt@awork3.anarazel.de
Provide a way to ask the kernel to use O_DIRECT (or local equivalent)
where available for data and WAL files, to avoid or minimize kernel
caching. This hurts performance currently and is not intended for end
users yet. Later proposed work would introduce our own I/O clustering,
read-ahead, etc to replace the facilities the kernel disables with this
option.
The only user-visible change, if the developer-only GUC is not used, is
that this commit also removes the obscure logic that would activate
O_DIRECT for the WAL when wal_sync_method=open_[data]sync and
wal_level=minimal (which also requires max_wal_senders=0). Those are
non-default and unlikely settings, and this behavior wasn't (correctly)
documented. The same effect can be achieved with io_direct=wal.
Author: Thomas Munro <thomas.munro@gmail.com>
Author: Andres Freund <andres@anarazel.de>
Author: Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>
Reviewed-by: Justin Pryzby <pryzby@telsasoft.com>
Reviewed-by: Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>
Discussion: https://postgr.es/m/CA%2BhUKGK1X532hYqJ_MzFWt0n1zt8trz980D79WbjwnT-yYLZpg%40mail.gmail.com
In order to have the option to use O_DIRECT/FILE_FLAG_NO_BUFFERING in a
later commit, we need the addresses of user space buffers to be well
aligned. The exact requirements vary by OS and file system (typically
sectors and/or memory pages). The address alignment size is set to
4096, which is enough for currently known systems: it matches modern
sectors and common memory page size. There is no standard governing
O_DIRECT's requirements so we might eventually have to reconsider this
with more information from the field or future systems.
Aligning I/O buffers on memory pages is also known to improve regular
buffered I/O performance.
Three classes of I/O buffers for regular data pages are adjusted:
(1) Heap buffers are now allocated with the new palloc_aligned() or
MemoryContextAllocAligned() functions introduced by commit 439f6175.
(2) Stack buffers now use a new struct PGIOAlignedBlock to respect
PG_IO_ALIGN_SIZE, if possible with this compiler. (3) The buffer
pool is also aligned in shared memory.
WAL buffers were already aligned on XLOG_BLCKSZ. It's possible for
XLOG_BLCKSZ to be configured smaller than PG_IO_ALIGNED_SIZE and thus
for O_DIRECT WAL writes to fail to be well aligned, but that's a
pre-existing condition and will be addressed by a later commit.
BufFiles are not yet addressed (there's no current plan to use O_DIRECT
for those, but they could potentially get some incidental speedup even
in plain buffered I/O operations through better alignment).
If we can't align stack objects suitably using the compiler extensions
we know about, we disable the use of O_DIRECT by setting PG_O_DIRECT to
0. This avoids the need to consider systems that have O_DIRECT but
can't align stack objects the way we want; such systems could in theory
be supported with more work but we don't currently know of any such
machines, so it's easier to pretend there is no O_DIRECT support
instead. That's an existing and tested class of system.
Add assertions that all buffers passed into smgrread(), smgrwrite() and
smgrextend() are correctly aligned, unless PG_O_DIRECT is 0 (= stack
alignment tricks may be unavailable) or the block size has been set too
small to allow arrays of buffers to be all aligned.
Author: Thomas Munro <thomas.munro@gmail.com>
Author: Andres Freund <andres@anarazel.de>
Reviewed-by: Justin Pryzby <pryzby@telsasoft.com>
Discussion: https://postgr.es/m/CA+hUKGK1X532hYqJ_MzFWt0n1zt8trz980D79WbjwnT-yYLZpg@mail.gmail.com
Add new options to the VACUUM and ANALYZE commands called
BUFFER_USAGE_LIMIT to allow users more control over how large to make the
buffer access strategy that is used to limit the usage of buffers in
shared buffers. Larger rings can allow VACUUM to run more quickly but
have the drawback of VACUUM possibly evicting more buffers from shared
buffers that might be useful for other queries running on the database.
Here we also add a new GUC named vacuum_buffer_usage_limit which controls
how large to make the access strategy when it's not specified in the
VACUUM/ANALYZE command. This defaults to 256KB, which is the same size as
the access strategy was prior to this change. This setting also
controls how large to make the buffer access strategy for autovacuum.
Per idea by Andres Freund.
Author: Melanie Plageman
Reviewed-by: David Rowley
Reviewed-by: Andres Freund
Reviewed-by: Justin Pryzby
Reviewed-by: Bharath Rupireddy
Discussion: https://postgr.es/m/20230111182720.ejifsclfwymw2reb@awork3.anarazel.de
The primary bottlenecks for relation extension are:
1) The extension lock is held while acquiring a victim buffer for the new
page. Acquiring a victim buffer can require writing out the old page
contents including possibly needing to flush WAL.
2) When extending via ReadBuffer() et al, we write a zero page during the
extension, and then later write out the actual page contents. This can
nearly double the write rate.
3) The existing bulk relation extension infrastructure in hio.c just amortized
the cost of acquiring the relation extension lock, but none of the other
costs.
Unfortunately 1) cannot currently be addressed in a central manner as the
callers to ReadBuffer() need to acquire the extension lock. To address that,
this this commit moves the responsibility for acquiring the extension lock
into bufmgr.c functions. That allows to acquire the relation extension lock
for just the required time. This will also allow us to improve relation
extension further, without changing callers.
The reason we write all-zeroes pages during relation extension is that we hope
to get ENOSPC errors earlier that way (largely works, except for CoW
filesystems). It is easier to handle out-of-space errors gracefully if the
page doesn't yet contain actual tuples. This commit addresses 2), by using the
recently introduced smgrzeroextend(), which extends the relation, without
dirtying the kernel page cache for all the extended pages.
To address 3), this commit introduces a function to extend a relation by
multiple blocks at a time.
There are three new exposed functions: ExtendBufferedRel() for extending the
relation by a single block, ExtendBufferedRelBy() to extend a relation by
multiple blocks at once, and ExtendBufferedRelTo() for extending a relation up
to a certain size.
To avoid duplicating code between ReadBuffer(P_NEW) and the new functions,
ReadBuffer(P_NEW) now implements relation extension with
ExtendBufferedRel(), using a flag to tell ExtendBufferedRel() that the
relation lock is already held.
Note that this commit does not yet lead to a meaningful performance or
scalability improvement - for that uses of ReadBuffer(P_NEW) will need to be
converted to ExtendBuffered*(), which will be done in subsequent commits.
Reviewed-by: Heikki Linnakangas <hlinnaka@iki.fi>
Reviewed-by: Melanie Plageman <melanieplageman@gmail.com>
Discussion: https://postgr.es/m/20221029025420.eplyow6k7tgu6he3@awork3.anarazel.de
A future patch will add support for extending relations by multiple blocks at
once. To be concurrency safe, the buffers for those blocks need to be marked
as BM_IO_IN_PROGRESS. Until now we only had infrastructure for recovering from
an IO error for a single buffer. This commit extends that infrastructure to
multiple buffers by using the resource owner infrastructure.
This commit increases the size of the ResourceOwnerData struct, which appears
to have a just about measurable overhead in very extreme workloads. Medium
term we are planning to substantially shrink the size of
ResourceOwnerData. Short term the increase is small enough to not worry about
it for now.
Reviewed-by: Melanie Plageman <melanieplageman@gmail.com>
Discussion: https://postgr.es/m/20221029025420.eplyow6k7tgu6he3@awork3.anarazel.de
Discussion: https://postgr.es/m/20221029200025.w7bvlgvamjfo6z44@awork3.anarazel.de
This adds a few more assertions against a buffer being local in places we
don't expect, and extracts the check for a buffer being pinned exactly once
from LockBufferForCleanup() into its own function. Later commits will use this
function.
Reviewed-by: Heikki Linnakangas <hlinnaka@iki.fi>
Reviewed-by: Melanie Plageman <melanieplageman@gmail.com>
Discussion: http://postgr.es/m/419312fd-9255-078c-c3e3-f0525f911d7f@iki.fi
smgrzeroextend() uses FileFallocate() to efficiently extend files by multiple
blocks. When extending by a small number of blocks, use FileZero() instead, as
using posix_fallocate() for small numbers of blocks is inefficient for some
file systems / operating systems. FileZero() is also used as the fallback for
FileFallocate() on platforms / filesystems that don't support fallocate.
A big advantage of using posix_fallocate() is that it typically won't cause
dirty buffers in the kernel pagecache. So far the most common pattern in our
code is that we smgrextend() a page full of zeroes and put the corresponding
page into shared buffers, from where we later write out the actual contents of
the page. If the kernel, e.g. due to memory pressure or elapsed time, already
wrote back the all-zeroes page, this can lead to doubling the amount of writes
reaching storage.
There are no users of smgrzeroextend() as of this commit. That will follow in
future commits.
Reviewed-by: Melanie Plageman <melanieplageman@gmail.com>
Reviewed-by: Heikki Linnakangas <hlinnaka@iki.fi>
Reviewed-by: Kyotaro Horiguchi <horikyota.ntt@gmail.com>
Reviewed-by: David Rowley <dgrowleyml@gmail.com>
Reviewed-by: John Naylor <john.naylor@enterprisedb.com>
Discussion: https://postgr.es/m/20221029025420.eplyow6k7tgu6he3@awork3.anarazel.de
Among other things, this should make it easier to calculate a useful cache hit
ratio by excluding buffer reads via buffer access strategies. As buffer access
strategies reuse buffers (and thus evict the prior buffer contents), it is
normal to see reads on repeated scans of the same data.
Author: Melanie Plageman <melanieplageman@gmail.com>
Reviewed-by: Bertrand Drouvot <bertranddrouvot.pg@gmail.com>
Reviewed-by: Andres Freund <andres@anarazel.de>
Discussion: https://postgr.es/m/CAAKRu_beMa9Hzih40%3DXPYqhDVz6tsgUGTrhZXRo%3Dunp%2Bszb%3DUA%40mail.gmail.com
This function has been a no-op for over a decade. Even if bufmgr
regains a need to be called during commit, it seems unlikely that
the most appropriate call points would be precisely here, so it's not
doing us much good as a placeholder either. Now, removing it probably
doesn't save any noticeable number of cycles --- but the main call is
inside the commit critical section, and the less work done there the
better.
Matthias van de Meent
Discussion: https://postgr.es/m/CAEze2Wi1=tLKbxZnXzcD+8fYKyKqBtivVakLQC_mYBsP4Y8qVA@mail.gmail.com
Commit 28e626bde0 introduced the infrastructure for tracking more detailed IO
statistics. This commit adds the actual collection of the new IO statistics
for relations and temporary relations. See aforementioned commit for goals and
high-level design.
The changes in this commit are fairly straight-forward. The bulk of the change
is to passing sufficient information to the callsites of pgstat_count_io_op().
A somewhat unsightly detail is that it currently is hard to find a better
place to count fsyncs than in md.c, whereas the other pgstat_count_io_op()
calls are in bufmgr.c/localbuf.c. As the number of fsyncs is tied to md.c
implementation details, it's not obvious there is a better answer.
Author: Melanie Plageman <melanieplageman@gmail.com>
Reviewed-by: Andres Freund <andres@anarazel.de>
Discussion: https://postgr.es/m/20200124195226.lth52iydq2n2uilq@alap3.anarazel.de
Invent separate macros for "invalid" values of these types, so that
we needn't embed knowledge of their representations into calling code.
These are all zeroes anyway ATM, so this is not fixing any live bug,
but it makes the code cleaner and more future-proof.
I (tgl) also chose to move DSM_HANDLE_INVALID into dsm_impl.h,
since it seems like it should live beside the typedef for dsm_handle.
Hou Zhijie, Nathan Bossart, Kyotaro Horiguchi, Tom Lane
Discussion: https://postgr.es/m/OS0PR01MB5716860B1454C34E5B179B6694C99@OS0PR01MB5716.jpnprd01.prod.outlook.com
Enforce wal_retrieve_retry_interval on a per-subscription basis,
rather than globally, and arrange to skip that delay in case of
an intentional worker exit. This probably makes little difference
in the field, where apply workers wouldn't be restarted often;
but it has a significant impact on the runtime of our logical
replication regression tests (even though those tests use
artificially-small wal_retrieve_retry_interval settings already).
Nathan Bossart, with mostly-cosmetic editorialization by me
Discussion: https://postgr.es/m/20221122004119.GA132961@nathanxps13
This provides a way to reserve connection slots for non-superusers.
The slots reserved via the new GUC are available only to users who
have the new predefined role pg_use_reserved_connections.
superuser_reserved_connections remains as a final reserve in case
reserved_connections has been exhausted.
Patch by Nathan Bossart. Reviewed by Tushar Ahuja and by me.
Discussion: http://postgr.es/m/20230119194601.GA4105788@nathanxps13
This makes sure that the internal logic of these functions does not
attempt to change the value of the arguments constified, and it removes
one unconstify() in basic_archive.c.
Author: Nathan Bossart
Reviewed-by: Andrew Dunstan, Peter Eisentraut
Discussion: https://postgr.es/m/20230114231126.GA2580330@nathanxps13
Most callers of BufFileRead() want to check whether they read the full
specified length. Checking this at every call site is very tedious.
This patch provides additional variants BufFileReadExact() and
BufFileReadMaybeEOF() that include the length checks.
I considered changing BufFileRead() itself, but this function is also
used in extensions, and so changing the behavior like this would
create a lot of problems there. The new names are analogous to the
existing LogicalTapeReadExact().
Reviewed-by: Amit Kapila <amit.kapila16@gmail.com>
Discussion: https://www.postgresql.org/message-id/flat/f3501945-c591-8cc3-5ef0-b72a2e0eaa9c@enterprisedb.com
Currently, for large transactions, the publisher sends the data in
multiple streams (changes divided into chunks depending upon
logical_decoding_work_mem), and then on the subscriber-side, the apply
worker writes the changes into temporary files and once it receives the
commit, it reads from those files and applies the entire transaction. To
improve the performance of such transactions, we can instead allow them to
be applied via parallel workers.
In this approach, we assign a new parallel apply worker (if available) as
soon as the xact's first stream is received and the leader apply worker
will send changes to this new worker via shared memory. The parallel apply
worker will directly apply the change instead of writing it to temporary
files. However, if the leader apply worker times out while attempting to
send a message to the parallel apply worker, it will switch to
"partial serialize" mode - in this mode, the leader serializes all
remaining changes to a file and notifies the parallel apply workers to
read and apply them at the end of the transaction. We use a non-blocking
way to send the messages from the leader apply worker to the parallel
apply to avoid deadlocks. We keep this parallel apply assigned till the
transaction commit is received and also wait for the worker to finish at
commit. This preserves commit ordering and avoid writing to and reading
from files in most cases. We still need to spill if there is no worker
available.
This patch also extends the SUBSCRIPTION 'streaming' parameter so that the
user can control whether to apply the streaming transaction in a parallel
apply worker or spill the change to disk. The user can set the streaming
parameter to 'on/off', or 'parallel'. The parameter value 'parallel' means
the streaming will be applied via a parallel apply worker, if available.
The parameter value 'on' means the streaming transaction will be spilled
to disk. The default value is 'off' (same as current behaviour).
In addition, the patch extends the logical replication STREAM_ABORT
message so that abort_lsn and abort_time can also be sent which can be
used to update the replication origin in parallel apply worker when the
streaming transaction is aborted. Because this message extension is needed
to support parallel streaming, parallel streaming is not supported for
publications on servers < PG16.
Author: Hou Zhijie, Wang wei, Amit Kapila with design inputs from Sawada Masahiko
Reviewed-by: Sawada Masahiko, Peter Smith, Dilip Kumar, Shi yu, Kuroda Hayato, Shveta Mallik
Discussion: https://postgr.es/m/CAA4eK1+wyN6zpaHUkCLorEWNx75MG0xhMwcFhvjqm2KURZEAGw@mail.gmail.com
An epoll fd belonging to the parent should be closed in the child. A
kqueue fd is automatically closed by fork(), but we should still adjust
our counter. For poll and Windows systems, nothing special is required.
On all systems we free the memory.
No caller yet, but we'll need this if we start using WaitEventSet in the
postmaster as planned.
Reviewed-by: Andres Freund <andres@anarazel.de>
Discussion: https://postgr.es/m/CA%2BhUKG%2BZ-HpOj1JsO9eWUP%2Bar7npSVinsC_npxSy%2BjdOMsx%3DGg%40mail.gmail.com
To be able to handle incoming connections on a server socket with
the WaitEventSet API, we'll need a new kind of event to indicate that
the the socket is ready to accept a connection.
On Unix, it's just the same as WL_SOCKET_READABLE, but on Windows there
is a different underlying kernel event that we need to map our
abstraction to.
No user yet, but a proposed patch would use this.
Reviewed-by: Andres Freund <andres@anarazel.de>
Discussion: https://postgr.es/m/CA%2BhUKG%2BZ-HpOj1JsO9eWUP%2Bar7npSVinsC_npxSy%2BjdOMsx%3DGg%40mail.gmail.com
A new function pg_stat_get_backend_subxact() can be used to get
information about the number of subtransactions in the cache of
a particular backend and whether that cache has overflowed. This
can be useful for tracking down performance problems that can
result from overflowed snapshots.
Dilip Kumar, reviewed by Zhihong Yu, Nikolay Samokhvalov,
Justin Pryzby, Nathan Bossart, Ashutosh Sharma, Julien
Rouhaud. Additional design comments from Andres Freund,
Tom Lane, Bruce Momjian, and David G. Johnston.
Discussion: http://postgr.es/m/CAFiTN-ut0uwkRJDQJeDPXpVyTWD46m3gt3JDToE02hTfONEN=Q@mail.gmail.com
Because we added StaticAssertStmt() first before StaticAssertDecl(),
some uses as well as the instructions in c.h are now a bit backwards
from the "native" way static assertions are meant to be used in C.
This updates the guidance and moves some static assertions to better
places.
Specifically, since the addition of StaticAssertDecl(), we can put
static assertions at the file level. This moves a number of static
assertions out of function bodies, where they might have been stuck
out of necessity, to perhaps better places at the file level or in
header files.
Also, when the static assertion appears in a position where a
declaration is allowed, then using StaticAssertDecl() is more native
than StaticAssertStmt().
Reviewed-by: John Naylor <john.naylor@enterprisedb.com>
Discussion: https://www.postgresql.org/message-id/flat/941a04e7-dd6f-c0e4-8cdf-a33b3338cbda%40enterprisedb.com
Make the argument types of the File API match stdio better:
- Change the data buffer to void *, from char *.
- Change FileWrite() data buffer to const on top of that.
- Change amounts to size_t, from int.
In passing, change the FilePrefetch() amount argument from int to
off_t, to match the underlying posix_fadvise().
Discussion: https://www.postgresql.org/message-id/flat/11dda853-bb5b-59ba-a746-e168b1ce4bdb%40enterprisedb.com
Previously, we'd compress only when the active range of array entries
reached Max(4 * PROCARRAY_MAXPROCS, 2 * pArray->numKnownAssignedXids).
If max_connections is large, the first term could result in not
compressing for a long time, resulting in much wastage of cycles in
hot-standby backends scanning the array to take snapshots. Get rid
of that term, and just bound it to 2 * pArray->numKnownAssignedXids.
That however creates the opposite risk, that we might spend too much
effort compressing. Hence, consider compressing only once every 128
commit records. (This frequency was chosen by benchmarking. While
we only tried one benchmark scenario, the results seem stable over
a fairly wide range of frequencies.)
Also, force compression when processing RecoveryInfo WAL records
(which should be infrequent); the old code could perform compression
then, but would do so only after the same array-range check as for
the transaction-commit path.
Also, opportunistically run compression if the startup process is about
to wait for WAL, though not oftener than once a second. This should
prevent cases where we waste lots of time by leaving the array
not-compressed for long intervals due to low WAL traffic.
Lastly, add a simple check to keep us from uselessly compressing
when the array storage is already compact.
Back-patch, as the performance problem is worse in pre-v14 branches
than in HEAD.
Simon Riggs and Michail Nikolaev, with help from Tom Lane and
Andres Freund.
Discussion: https://postgr.es/m/CALdSSPgahNUD_=pB_j=1zSnDBaiOtqVfzo8Ejt5J_k7qZiU1Tw@mail.gmail.com
Until now LWLockDequeueSelf() sequentially searched the list of waiters to see
if the current proc is still is on the list of waiters, or has already been
removed. In extreme workloads, where the wait lists are very long, this leads
to a quadratic behavior. #backends iterating over a list #backends
long. Additionally, the likelihood of needing to call LWLockDequeueSelf() in
the first place also increases with the increased length of the wait queue, as
it becomes more likely that a lock is released while waiting for the wait list
lock, which is held for longer during lock release.
Due to the exponential back-off in perform_spin_delay() this is surprisingly
hard to detect. We should make that easier, e.g. by adding a wait event around
the pg_usleep() - but that's a separate patch.
The fix is simple - track whether a proc is currently waiting in the wait list
or already removed but waiting to be woken up in PGPROC->lwWaiting.
In some workloads with a lot of clients contending for a small number of
lwlocks (e.g. WALWriteLock), the fix can substantially increase throughput.
As the quadratic behavior arguably is a bug, we might want to decide to
backpatch this fix in the future.
Author: Andres Freund <andres@anarazel.de>
Reviewed-by: Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>
Discussion: https://postgr.es/m/20221027165914.2hofzp4cvutj6gin@awork3.anarazel.de
Discussion: https://postgr.es/m/CALj2ACXktNbG=K8Xi7PSqbofTZozavhaxjatVc14iYaLu4Maag@mail.gmail.com
