This is similar to 2065ddf5e34c, but this time for pg_logical/ itself
and its contents, like the paths for snapshots, mappings or origin
checkpoints.
Author: Bertrand Drouvot
Reviewed-by: Ashutosh Bapat, Yugo Nagata, Michael Paquier
Discussion: https://postgr.es/m/ZryVvjqS9SnV1GPP@ip-10-97-1-34.eu-west-3.compute.internal
This commit replaces most of the hardcoded values of "pg_replslot" by a
new PG_REPLSLOT_DIR #define. This makes the style more consistent with
the existing PG_STAT_TMP_DIR, for example. More places will follow a
similar change.
Author: Bertrand Drouvot
Reviewed-by: Ashutosh Bapat, Yugo Nagata, Michael Paquier
Discussion: https://postgr.es/m/ZryVvjqS9SnV1GPP@ip-10-97-1-34.eu-west-3.compute.internal
The conflict types 'update_differ' and 'delete_differ' indicate that a row
to be modified was previously altered by another origin. Rename those to
'update_origin_differs' and 'delete_origin_differs' to clarify their
meaning.
Author: Hou Zhijie
Reviewed-by: Shveta Malik, Peter Smith
Discussion: https://postgr.es/m/CAA4eK1+HEKwG_UYt4Zvwh5o_HoCKCjEGesRjJX38xAH3OxuuYA@mail.gmail.com
Commit 5bec1d6bc5e changed the memory usage updates of the
ReorderBufferTXN to zero all at once by subtracting txn->size, rather
than updating it for each change. However, if TOAST reconstruction
data remained in the transaction when freeing it, there were cases
where it further subtracted the memory counter from zero, resulting in
an assertion failure.
This change calculates the memory size for each change and updates the
memory usage to precisely the amount that has been freed.
Backpatch to v17, where this was introducd.
Reviewed-by: Amit Kapila, Shlok Kyal
Discussion: https://postgr.es/m/CAD21AoAqkNUvicgKPT_dXzNoOwpPkVTg0QPPxEcWmzT0moCJ1g%40mail.gmail.com
Backpatch-through: 17
We advance origin progress during abort on successful streaming and
application of ROLLBACK in parallel streaming mode. But the origin
shouldn't be advanced during an error or unsuccessful apply due to
shutdown. Otherwise, it will result in a transaction loss as such a
transaction won't be sent again by the server.
Reported-by: Hou Zhijie
Author: Hayato Kuroda and Shveta Malik
Reviewed-by: Amit Kapila
Backpatch-through: 16
Discussion: https://postgr.es/m/TYAPR01MB5692FAC23BE40C69DA8ED4AFF5B92@TYAPR01MB5692.jpnprd01.prod.outlook.com
This patch provides the additional logging information in the following
conflict scenarios while applying changes:
insert_exists: Inserting a row that violates a NOT DEFERRABLE unique constraint.
update_differ: Updating a row that was previously modified by another origin.
update_exists: The updated row value violates a NOT DEFERRABLE unique constraint.
update_missing: The tuple to be updated is missing.
delete_differ: Deleting a row that was previously modified by another origin.
delete_missing: The tuple to be deleted is missing.
For insert_exists and update_exists conflicts, the log can include the origin
and commit timestamp details of the conflicting key with track_commit_timestamp
enabled.
update_differ and delete_differ conflicts can only be detected when
track_commit_timestamp is enabled on the subscriber.
We do not offer additional logging for exclusion constraint violations because
these constraints can specify rules that are more complex than simple equality
checks. Resolving such conflicts won't be straightforward. This area can be
further enhanced if required.
Author: Hou Zhijie
Reviewed-by: Shveta Malik, Amit Kapila, Nisha Moond, Hayato Kuroda, Dilip Kumar
Discussion: https://postgr.es/m/OS0PR01MB5716352552DFADB8E9AD1D8994C92@OS0PR01MB5716.jpnprd01.prod.outlook.com
The apply worker was using XactLastCommitEnd as local end_lsn for applying
prepare and rollback_prepare. The XactLastCommitEnd value is the end lsn
of the last commit applied before the prepare transaction which makes no
sense. This LSN is used to decide whether we can send the acknowledgment
of the corresponding remote LSN to the server.
It is okay not to set the local_end LSN with the actual WAL position for
the prepare because we always flush the prepare record. So, we can send
the acknowledgment of the remote_end LSN as soon as prepare is finished.
The current code is misleading but as such doesn't create any problem, so
decided not to backpatch.
Author: Hayato Kuroda
Reviewed-by: Shveta Malik, Amit Kapila
Discussion: https://postgr.es/m/TYAPR01MB5692FA4926754B91E9D7B5F0F5AA2@TYAPR01MB5692.jpnprd01.prod.outlook.com
Before Bison 3.4, the generated parser implementation files run afoul
of -Wmissing-variable-declarations (in spite of commit ab61c40bfa2)
because declarations for yylval and possibly yylloc are missing. The
generated header files contain an extern declaration, but the
implementation files don't include the header files. Since Bison 3.4,
the generated implementation files automatically include the generated
header files, so then it works.
To make this work with older Bison versions as well, include the
generated header file from the .y file.
(With older Bison versions, the generated implementation file contains
effectively a copy of the header file pasted in, so including the
header file is redundant. But we know this works anyway because the
core grammar uses this arrangement already.)
Discussion: https://www.postgresql.org/message-id/flat/e0a62134-83da-4ba4-8cdb-ceb0111c95ce@eisentraut.org
This adds extern declarations for some global variables produced by
Bison that are not already declared in its generated header file.
This is a workaround to be able to add -Wmissing-variable-declarations
to the global set of warning options in the near future.
Another longer-term solution would be to convert these grammars to
"pure" parsers in Bison, to avoid global variables altogether. Note
that the core grammar is already pure, so this patch did not need to
touch it.
Reviewed-by: Andres Freund <andres@anarazel.de>
Discussion: https://www.postgresql.org/message-id/flat/e0a62134-83da-4ba4-8cdb-ceb0111c95ce@eisentraut.org
The two_phase option is controlled by both the publisher (as a slot
option) and the subscriber (as a subscription option), so the slot option
must also be modified.
Changing the 'two_phase' option for a subscription from 'true' to 'false'
is permitted only when there are no pending prepared transactions
corresponding to that subscription. Otherwise, the changes of already
prepared transactions can be replicated again along with their corresponding
commit leading to duplicate data or errors.
To avoid data loss, the 'two_phase' option for a subscription can only be
changed from 'false' to 'true' once the initial data synchronization is
completed. Therefore this is performed later by the logical replication worker.
Author: Hayato Kuroda, Ajin Cherian, Amit Kapila
Reviewed-by: Peter Smith, Hou Zhijie, Amit Kapila, Vitaly Davydov, Vignesh C
Discussion: https://postgr.es/m/8fab8-65d74c80-1-2f28e880@39088166
Commit f4b54e1ed9, which introduced macros for protocol characters,
missed updating a few places. It also did not introduce macros for
messages sent from parallel workers to their leader processes.
This commit adds a new section in protocol.h for those.
Author: Aleksander Alekseev
Discussion: https://postgr.es/m/CAJ7c6TNTd09AZq8tGaHS3LDyH_CCnpv0oOz2wN1dGe8zekxrdQ%40mail.gmail.com
Backpatch-through: 17
When creating and initializing a logical slot, the restart_lsn is set
to the latest WAL insertion point (or the latest replay point on
standbys). Subsequently, WAL records are decoded from that point to
find the start point for extracting changes in the
DecodingContextFindStartpoint() function. Since the initial
restart_lsn could be in the middle of a transaction, the start point
must be a consistent point where we won't see the data for partial
transactions.
Previously, when not building a full snapshot, serialized snapshots
were restored, and the SnapBuild jumps to the consistent state even
while finding the start point. Consequently, the slot's restart_lsn
and confirmed_flush could be set to the middle of a transaction. This
could lead to various unexpected consequences. Specifically, there
were reports of logical decoding decoding partial transactions, and
assertion failures occurred because only subtransactions were decoded
without decoding their top-level transaction until decoding the commit
record.
To resolve this issue, the changes prevent restoring the serialized
snapshot and jumping to the consistent state while finding the start
point.
On v17 and HEAD, a flag indicating whether snapshot restores should be
skipped has been added to the SnapBuild struct, and SNAPBUILD_VERSION
has been bumpded.
On backbranches, the flag is stored in the LogicalDecodingContext
instead, preserving on-disk compatibility.
Backpatch to all supported versions.
Reported-by: Drew Callahan
Reviewed-by: Amit Kapila, Hayato Kuroda
Discussion: https://postgr.es/m/2444AA15-D21B-4CCE-8052-52C7C2DAFE5C%40amazon.com
Backpatch-through: 12
Previously, the comment incorrectly stated that libpqrcv_check_conninfo()
returns true or false based on the connection string check.
However, this function actually has a void return type and
raises an error if the check fails.
Author: Rintaro Ikeda
Reviewed-by: Jelte Fennema-Nio, Fujii Masao
Discussion: https://postgr.es/m/6a1ca81b27fec4da0ccdfaaaec787982@oss.nttdata.com
Commit 0fdab27ad6 changed the code to wait for WAL to be available before
determining the timeline but forgot to move the failure check.
This change is to make the related code easier to understand and enhance
otherwise there is no bug in the current code.
In the passing, improve the nearby comments to explain why we determine
am_cascading_walsender after waiting for the required WAL.
Author: Peter Smith
Reviewed-by: Bertrand Drouvot, Amit Kapila
Discussion: https://postgr.es/m/CAHut+PvqX49fusLyXspV1Mmd_EekPtXG0oT146vZjcb9XDvNgw@mail.gmail.com
All the errors triggered in the code paths patched here would cause the
backend to issue an internal_error errcode, which is a state that should
be used only for "can't happen" situations. However, these code paths
are reachable by the regression tests, and could be seen by users in
valid cases. Some regression tests expect internal errcodes as they
manipulate the backend state to cause corruption (like checksums), or
use elog() because it is more convenient (like injection points), these
have no need to change.
This reduces the number of internal failures triggered in a check-world
by more than half, while providing correct errcodes for these valid
cases.
Reviewed-by: Robert Haas
Discussion: https://postgr.es/m/Zic_GNgos5sMxKoa@paquier.xyz
After several refactoring iterations, auxiliary processes are no
longer initialized from the bootstrapper. Using the InitProcessing
mode for initializing auxiliary processes is more appropriate. Since
the global variable Mode is initialized to InitProcessing, we can just
remove the redundant calls of SetProcessingMode(InitProcessing).
Author: Xing Guo <higuoxing@gmail.com>
Discussion: https://www.postgresql.org/message-id/CACpMh%2BDBHVT4xPGimzvex%3DwMdMLQEu9PYhT%2BkwwD2x2nu9dU_Q%40mail.gmail.com
Up to now, committing a transaction has caused CurrentMemoryContext to
get set to TopMemoryContext. Most callers did not pay any particular
heed to this, which is problematic because TopMemoryContext is a
long-lived context that never gets reset. If the caller assumes it
can leak memory because it's running in a limited-lifespan context,
that behavior translates into a session-lifespan memory leak.
The first-reported instance of this involved ProcessIncomingNotify,
which is called from the main processing loop that normally runs in
MessageContext. That outer-loop code assumes that whatever it
allocates will be cleaned up when we're done processing the current
client message --- but if we service a notify interrupt, then whatever
gets allocated before the next switch to MessageContext will be
permanently leaked in TopMemoryContext. sinval catchup interrupts
have a similar problem, and I strongly suspect that some places in
logical replication do too.
To fix this in a generic way, let's redefine the behavior as
"CommitTransactionCommand restores the memory context that was current
at entry to StartTransactionCommand". This clearly fixes the issue
for the notify and sinval cases, and it seems to match the mental
model that's in use in the logical replication code, to the extent
that anybody thought about it there at all.
For consistency, likewise make subtransaction exit restore the context
that was current at subtransaction start (rather than always selecting
the CurTransactionContext of the parent transaction level). This case
has less risk of resulting in a permanent leak than the outer-level
behavior has, but it would not meet the principle of least surprise
for some CommitTransactionCommand calls to restore the previous
context while others don't.
While we're here, also change xact.c so that we reset
TopTransactionContext at transaction exit and then re-use it in later
transactions, rather than dropping and recreating it in each cycle.
This probably doesn't save a lot given the context recycling mechanism
in aset.c, but it should save a little bit. Per suggestion from David
Rowley. (Parenthetically, the text in src/backend/utils/mmgr/README
implies that this is how I'd planned to implement it as far back as
commit 1aebc3618 --- but the code actually added in that commit just
drops and recreates it each time.)
This commit also cleans up a few places outside xact.c that were
needlessly making TopMemoryContext current, and thus risking more
leaks of the same kind. I don't think any of them represent
significant leak risks today, but let's deal with them while the
issue is top-of-mind.
Per bug #18512 from wizardbrony. Commit to HEAD only, as this change
seems to have some risk of breaking things for some callers. We'll
apply a narrower fix for the known-broken cases in the back branches.
Discussion: https://postgr.es/m/3478884.1718656625@sss.pgh.pa.us
The standby_slot_names GUC allows the specification of physical standby
slots that must be synchronized before the logical walsenders associated
with logical failover slots. However, for this purpose, the GUC name is
too generic.
Author: Hou Zhijie
Reviewed-by: Bertrand Drouvot, Masahiko Sawada
Backpatch-through: 17
Discussion: https://postgr.es/m/ZnWeUgdHong93fQN@momjian.us
In pgoutput, when converting the child table's tuple format to match the
parent table's, we temporarily create a new slot to store the converted
tuple. However, we missed to drop such temporary slots, leading to
resource leakage.
Reported-by: Bowen Shi
Author: Hou Zhijie
Reviewed-by: Amit Kapila
Backpatch-through: 15
Discussion: https://postgr.es/m/CAM_vCudv8dc3sjWiPkXx5F2b27UV7_YRKRbtSCcE-pv=cVACGA@mail.gmail.com
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 during Postgres 17 development.
pg_bsd_indent still has a couple of similar inconsistencies, which I
(pgeoghegan) have left untouched for now.
This commit was written with help from clang-tidy, by mechanically
applying the same rules as similar clean-up commits (the earliest such
commit was commit 035ce1fe).
Commit e0b2eed047 assumed that the confirmed_flush LSN can't go backward.
However, it is possible that confirmed_flush LSN can go backward
temporarily when the client acknowledges a prior value of flush location.
This can happen when the client (subscriber in this case) acknowledges an
LSN it doesn't have to do anything for (say for DDLs) and thus didn't
store persistently. After restart, the client sends the prior value of
flush LSN which it had stored persistently and the server updates the
confirmed_flush LSN with that value.
The fix is to remove the assumption and not allow the prior value of
confirmed_flush LSN to be flushed to the disk.
Author: Vignesh C
Reviewed-by: Amit Kapila, Shlok Kyal
Discussion: https://postgr.es/m/CALDaNm3hgow2+oEov5jBk4iYP5eQrUCF1yZtW7+dV3J__p4KLQ@mail.gmail.com
After further review, we want to move in the direction of always
quoting GUC names in error messages, rather than the previous (PG16)
wildly mixed practice or the intermittent (mid-PG17) idea of doing
this depending on how possibly confusing the GUC name is.
This commit applies appropriate quotes to (almost?) all mentions of
GUC names in error messages. It partially supersedes a243569bf65 and
8d9978a7176, which had moved things a bit in the opposite direction
but which then were abandoned in a partial state.
Author: Peter Smith <smithpb2250@gmail.com>
Discussion: https://www.postgresql.org/message-id/flat/CAHut%2BPv-kSN8SkxSdoHano_wPubqcg5789ejhCDZAcLFceBR-w%40mail.gmail.com
Allow pg_sync_replication_slots() to error out during promotion of standby.
This makes the behavior of the SQL function consistent with the slot sync
worker. We also ensured that pg_sync_replication_slots() cannot be
executed if sync_replication_slots is enabled and the slotsync worker is
already running to perform the synchronization of slots. Previously, it
would have succeeded in cases when the worker is idle and failed when it
is performing sync which could confuse users.
This patch fixes another issue in the slot sync worker where
SignalHandlerForShutdownRequest() needs to be registered *before* setting
SlotSyncCtx->pid, otherwise, the slotsync worker could miss handling
SIGINT sent by the startup process(ShutDownSlotSync) if it is sent before
worker could register SignalHandlerForShutdownRequest(). To be consistent,
all signal handlers' registration is moved to a prior location before we
set the worker's pid.
Ensure that we clean up synced temp slots at the end of
pg_sync_replication_slots() to avoid such slots being left over after
promotion.
Ensure that ShutDownSlotSync() captures SlotSyncCtx->pid under spinlock to
avoid accessing invalid value as it can be reset by concurrent slot sync
exit due to an error.
Author: Shveta Malik
Reviewed-by: Hou Zhijie, Bertrand Drouvot, Amit Kapila, Masahiko Sawada
Discussion: https://postgr.es/m/CAJpy0uBefXUS_TSz=oxmYKHdg-fhxUT0qfjASW3nmqnzVC3p6A@mail.gmail.com
We missed performing table sync if the invalidation happened while the
non-ready tables list was being prepared. This occurs because the sync
state was set to valid at the end of non-ready table list preparation
irrespective of the invalidations processed while the list is being
prepared.
Fix it by changing the boolean variable to a tri-state enum and by setting
table state to valid only if no invalidations have occurred while the list
is being prepared.
Reprted-by: Alexander Lakhin
Diagnosed-by: Alexander Lakhin
Author: Vignesh C
Reviewed-by: Hou Zhijie, Alexander Lakhin, Ajin Cherian, Amit Kapila
Backpatch-through: 15
Discussion: https://postgr.es/m/711a6afe-edb7-1211-cc27-1bef8239eec7@gmail.com
This fixes various typos, duplicated words, and tiny bits of whitespace
mainly in code comments but also in docs.
Author: Daniel Gustafsson <daniel@yesql.se>
Author: Heikki Linnakangas <hlinnaka@iki.fi>
Author: Alexander Lakhin <exclusion@gmail.com>
Author: David Rowley <dgrowleyml@gmail.com>
Author: Nazir Bilal Yavuz <byavuz81@gmail.com>
Discussion: https://postgr.es/m/3F577953-A29E-4722-98AD-2DA9EFF2CBB8@yesql.se
Ensure that when updating the catalog_xmin of the synced slots, it is
first written to disk before changing the in-memory value
(effective_catalog_xmin). This is to prevent a scenario where the
in-memory value change triggers a vacuum to remove catalog tuples before
the catalog_xmin is written to disk. In the event of a crash before the
catalog_xmin is persisted, we would not know that some required catalog
tuples have been removed and the synced slot would be invalidated.
Change the sanity check to ensure that remote_slot's confirmed_flush LSN
can't precede the local/synced slot during slot sync. Note that the
restart_lsn of the synced/local slot can be ahead of remote_slot. This can
happen when slot advancing machinery finds a running xacts record after
reaching the consistent state at a later point than the primary where it
serializes the snapshot and updates the restart_lsn.
Make the check to sync slots robust by allowing to sync only when the
confirmed_lsn, restart_lsn, or catalog_xmin of the remote slot is ahead of
the synced/local slot.
Reported-by: Amit Kapila and Shveta Malik
Author: Hou Zhijie, Shveta Malik
Reviewed-by: Amit Kapila, Bertrand Drouvot
Discussion: https://postgr.es/m/OS0PR01MB57162B67D3CB01B2756FBA6D94062@OS0PR01MB5716.jpnprd01.prod.outlook.com
Discussion: https://postgr.es/m/CAJpy0uCSS5zmdyUXhvw41HSdTbRqX1hbYqkOfHNj7qQ+2zn0AQ@mail.gmail.com
This reverts commit b840508644 and bcb14f4abc. These commits were made
for commit 5bec1d6bc5 (Improve eviction algorithm in ReorderBuffer
using max-heap for many subtransactions). However, per discussion,
commit efb8acc0d0 replaced binary heap + index with pairing heap, and
made these commits unnecessary.
Reported-by: Jeff Davis
Discussion: https://postgr.es/m/12747c15811d94efcc5cda72d6b35c80d7bf3443.camel%40j-davis.com
A pairing heap can perform the same operations as the binary heap +
index, with as good or better algorithmic complexity, and that's an
existing data structure so that we don't need to invent anything new
compared to v16. This commit makes the new binaryheap functionality
that was added in commits b840508644 and bcb14f4abc unnecessary, but
they will be reverted separately.
Remove the optimization to only build and maintain the heap when the
amount of memory used is close to the limit, becuase the bookkeeping
overhead with the pairing heap seems to be small enough that it
doesn't matter in practice.
Reported-by: Jeff Davis
Author: Heikki Linnakangas
Reviewed-by: Michael Paquier, Hayato Kuroda, Masahiko Sawada
Discussion: https://postgr.es/m/12747c15811d94efcc5cda72d6b35c80d7bf3443.camel%40j-davis.com
This adjusts various appendStringInfo* function calls to use a more
appropriate and efficient function with the same behavior. For example,
use appendStringInfoChar() when appending a single character rather than
appendStringInfo() and appendStringInfoString() when no formatting is
required rather than using appendStringInfo().
All adjustments made here are in code that's new to v17, so it makes
sense to fix these now rather than wait a few years and make
backpatching harder.
Discussion: https://postgr.es/m/CAApHDvojY2UvMiO+9_55ArTj10P1LBNJyyoGB+C65BLDNT0GsQ@mail.gmail.com
Reviewed-by: Nathan Bossart, Tom Lane
This patch generalizes libpq's existing single-row mode to allow
individual partial-result PGresults to contain up to N rows, rather
than always one row. This reduces malloc overhead compared to plain
single-row mode, and it is very useful for psql's FETCH_COUNT feature,
since otherwise we'd have to add code (and cycles) to either merge
single-row PGresults into a bigger one or teach psql's
results-printing logic to accept arrays of PGresults.
To avoid API breakage, PQsetSingleRowMode() remains the same, and we
add a new function PQsetChunkedRowsMode() to invoke the more general
case. Also, PGresults obtained the old way continue to carry the
PGRES_SINGLE_TUPLE status code, while if PQsetChunkedRowsMode() is
used then their status code is PGRES_TUPLES_CHUNK. The underlying
logic is the same either way, though.
Daniel Vérité, reviewed by Laurenz Albe and myself (and whacked
around a bit by me, so any remaining bugs are my fault)
Discussion: https://postgr.es/m/CAKZiRmxsVTkO928CM+-ADvsMyePmU3L9DQCa9NwqjvLPcEe5QA@mail.gmail.com
This commit does two things:
1) Maintains inactive_since for sync slots whenever the slot is released
just like any other regular slot.
2) Ensures the value is set to the current timestamp during the promotion
of standby to help correctly interpret the time after promotion. We don't
want the slots to appear inactive for a long time after promotion if they
haven't been synchronized recently. This would also avoid the invalidation
of such slots immediately after promotion if tomorrow we have a feature
that invalidates slots based on their inactivity time. Whoever acquires
the slot i.e. makes the slot active will reset it to NULL.
Author: Bharath Rupireddy
Reviewed-by: Bertrand Drouvot, Amit Kapila, Shveta Malik, Masahiko Sawada
Discussion: https://postgr.es/m/CAA4eK1KrPGwfZV9LYGidjxHeW+rxJ=E2ThjXvwRGLO=iLNuo=Q@mail.gmail.com
Discussion: https://postgr.es/m/CALj2ACW4aUe-_uFQOjdWCEN-xXoLGhmvRFnL8SNw_TZ5nJe+aw@mail.gmail.com
Discussion: https://postgr.es/m/CA+Tgmob_Ta-t2ty8QrKHBGnNLrf4ZYcwhGHGFsuUoFrAEDw4sA@mail.gmail.com
We were directly copying the LSN locations while syncing the slots on the
standby. Now, it is possible that at some particular restart_lsn there are
some running xacts, which means if we start reading the WAL from that
location after promotion, we won't reach a consistent snapshot state at
that point. However, on the primary, we would have already been in a
consistent snapshot state at that restart_lsn so we would have just
serialized the existing snapshot.
To avoid this problem we will use the advance_slot functionality unless
the snapshot already exists at the synced restart_lsn location. This will
help us to ensure that snapbuilder/slot statuses are updated properly
without generating any changes. Note that the synced slot will remain as
RS_TEMPORARY till the decoding from corresponding restart_lsn can reach a
consistent snapshot state after which they will be marked as
RS_PERSISTENT.
Per buildfarm
Author: Hou Zhijie
Reviewed-by: Bertrand Drouvot, Shveta Malik, Bharath Rupireddy, Amit Kapila
Discussion: https://postgr.es/m/OS0PR01MB5716B3942AE49F3F725ACA92943B2@OS0PR01MB5716.jpnprd01.prod.outlook.com
Previously, when selecting the transaction to evict during logical
decoding, we check all transactions to find the largest
transaction. This could lead to a significant replication lag
especially in the case where there are many subtransactions.
This commit improves the eviction algorithm in ReorderBuffer using the
max-heap with transaction size as the key to efficiently find the
largest transaction.
The max-heap starts with empty. While the max-heap is empty, we don't
do anything for the max-heap when updating the memory
counter. Therefore, we get the largest transaction in O(N) time, where
N is the number of transactions including top-level transactions and
subtransactions.
We build the max-heap just before selecting the largest transactions
if the number of transactions being decoded is higher than the
threshold, MAX_HEAP_TXN_COUNT_THRESHOLD. After building the max-heap,
we also update the max-heap when updating the memory counter. The
intention is to efficiently find the largest transaction in O(1) time
instead of incurring the cost of memory counter updates (O(log
N)). Once the number of transactions got lower than the threshold, we
reset the max-heap.
The performance benchmark results showed significant speed up (more
than x30 speed up on my machine) in decoding a transaction with 100k
subtransactions, whereas there is no visible overhead in other cases.
Reviewed-by: Amit Kapila, Hayato Kuroda, Vignesh C, Ajin Cherian,
Tomas Vondra, Shubham Khanna, Peter Smith, Álvaro Herrera,
Euler Taveira
Discussion: https://postgr.es/m/CAD21AoAfKTgrBrLq96GcTv9d6k97zaQcDM-rxfKEt4GSe0qnaQ%40mail.gmail.com
Previously, binaryheap didn't support updating a key and removing a
node in an efficient way. For example, in order to remove a node from
the binaryheap, the caller had to pass the node's position within the
array that the binaryheap internally has. Removing a node from the
binaryheap is done in O(log n) but searching for the key's position is
done in O(n).
This commit adds a hash table to binaryheap in order to track the
position of each nodes in the binaryheap. That way, by using newly
added functions such as binaryheap_update_up() etc., both updating a
key and removing a node can be done in O(1) on an average and O(log n)
in worst case. This is known as the indexed binary heap. The caller
can specify to use the indexed binaryheap by passing indexed = true.
The current code does not use the new indexing logic, but it will be
used by an upcoming patch.
Reviewed-by: Vignesh C, Peter Smith, Hayato Kuroda, Ajin Cherian,
Tomas Vondra, Shubham Khanna
Discussion: https://postgr.es/m/CAD21AoDffo37RC-eUuyHJKVEr017V2YYDLyn1xF_00ofptWbkg%40mail.gmail.com
