When commit 0f33a719fdbb5d8c43839ea0d2c90cd03e2af2d2 removed the
instructions to start/stop the new cluster before running rsync, it was
now possible for pg_resetwal/pg_resetxlog to leave the final WAL record
at wal_level=minimum, preventing upgraded standby servers from
reconnecting.
This patch fixes that by having pg_upgrade unconditionally start/stop
the new cluster after pg_resetwal/pg_resetxlog has run.
Backpatch through 9.2 since, though the instructions were added in PG
9.5, they worked all the way back to 9.2.
Discussion: https://postgr.es/m/20170620171844.GC24975@momjian.us
Backpatch-through: 9.2
When materialized views were added, psql's \d commands were made to
treat them as a separate object category ... but not everyplace in the
documentation or comments got the memo.
Noted by David Johnston. Back-patch to 9.3 where matviews came in.
Discussion: https://postgr.es/m/CAKFQuwb27M3VXRhHErjCpkWwN9eKThbqWb1=trtoXi9_ejqPXQ@mail.gmail.com
David Rowley found that the "use the smallest per-column selectivity"
heuristic applied in some cases by get_foreign_key_join_selectivity()
was badly off if the FK columns are independent, producing estimates
much worse than we got before that code was added in 9.6.
One case where that heuristic was used was for LEFT and FULL outer joins
with the referenced rel on the outside of the join. But we should not
really need to special-case those here. eqjoinsel() never has had such a
special case; the correction is applied by calc_joinrel_size_estimate()
instead. Let's just estimate such cases like inner joins and rely on that
later adjustment. (I think there was something of a thinko here, in that
the comments seem to be thinking about the selectivity as defined for
semi/anti joins; but that shouldn't apply to left/full joins.) Add a
regression test exercising such a case to show that this is sane in
at least some cases.
The other case where we used that heuristic was for SEMI/ANTI outer joins,
either if the referenced rel was on the outside, or if it was on the inside
but was part of a join within the RHS. In either case, the FK doesn't give
us a lot of traction towards estimating the selectivity. To ensure that
we don't have regressions from what happened before 9.6, let's punt by
ignoring the FK in such cases and applying the traditional selectivity
calculation. (We might be able to improve on that later, but for now
I just want to be sure it's not worse than 9.5.)
Report and patch by David Rowley, simplified a bit by me. Back-patch
to 9.6 where this code was added.
Discussion: https://postgr.es/m/CAKJS1f8NO8oCDcxrteohG6O72uU1saEVT9qX=R8pENr5QWerXw@mail.gmail.com
The combination of -Z -Fp and output to stdout resulted in corrupted
output data, because we left stdout in text mode, resulting in newline
conversion being done on the compressed stream. Switch stdout to binary
mode for this case, at the same place where we do it for non-text output
formats.
Report and patch by Kuntal Ghosh, tested by Ashutosh Sharma and Neha
Sharma. Back-patch to all supported branches.
Discussion: https://postgr.es/m/CAGz5QCJPvbBjXAmJuGx1B_41yVCetAJhp7rtaDf7XQGWuB1GSw@mail.gmail.com
When, during logical decoding, a transaction gets too big, it's
contents get spilled to disk. Not just the top-transaction gets
spilled, but *also* all of its subtransactions, even if they're not
that large themselves. Unfortunately we didn't clean up
such small spilled subtransactions from disk.
Fix that, by keeping better track of whether a transaction has been
spilled to disk.
Author: Andres Freund
Reported-By: Dmitriy Sarafannikov, Fabrízio de Royes Mello
Discussion:
https://postgr.es/m/1457621358.355011041@f382.i.mail.ruhttps://postgr.es/m/CAFcNs+qNMhNYii4nxpO6gqsndiyxNDYV0S=JNq0v_sEE+9PHXg@mail.gmail.com
Backpatch: 9.4-, where logical decoding was introduced
When a new base type is created using the old-style procedure of first
creating the input/output functions with "opaque" in place of the base
type, the "opaque" argument/return type is changed to the final base type,
on CREATE TYPE. However, we did not create a pg_depend record when doing
that, so the functions were left not depending on the type.
Fixes bug #14706, reported by Karen Huddleston.
Discussion: https://www.postgresql.org/message-id/20170614232259.1424.82774@wrigleys.postgresql.org
We had three occurrences of essentially the same coding pattern
wherein we tried to retrieve a query result from a libpq connection
without blocking. In the case where PQconsumeInput failed (typically
indicating a lost connection), all three loops simply gave up and
returned, forgetting to clear any previously-collected PGresult
object. Since those are malloc'd not palloc'd, the oversight results
in a process-lifespan memory leak.
One instance, in libpqwalreceiver, is of little significance because
the walreceiver process would just quit anyway if its connection fails.
But we might as well fix it.
The other two instances, in postgres_fdw, are somewhat more worrisome
because at least in principle the scenario could be repeated, allowing
the amount of memory leaked to build up to something worth worrying
about. Moreover, in these cases the loops contain CHECK_FOR_INTERRUPTS
calls, as well as other calls that could potentially elog(ERROR),
providing another way to exit without having cleared the PGresult.
Here we need to add PG_TRY logic similar to what exists in quite a
few other places in postgres_fdw.
Coverity noted the libpqwalreceiver bug; I found the other two cases
by checking all calls of PQconsumeInput.
Back-patch to all supported versions as appropriate (9.2 lacks
postgres_fdw, so this is really quite unexciting for that branch).
Discussion: https://postgr.es/m/22620.1497486981@sss.pgh.pa.us
It was unsafe to instruct users to start/stop the server after
pg_upgrade was run but before the standby servers were rsync'ed. The
new instructions avoid this.
RELEASE NOTES: This fix should be mentioned in the minor release notes.
Reported-by: Dmitriy Sarafannikov and Sergey Burladyan
Discussion: https://postgr.es/m/87wp8o506b.fsf@seb.koffice.internal
Backpatch-through: 9.5, where standby server upgrade instructions first appeared
During pg_upgrade's restore run, all relfilenode choices should be
overridden by commands in the dump script. If we ever find ourselves
choosing a relfilenode in the ordinary way, someone blew it. Likewise for
pg_type OIDs. Since pg_upgrade might well succeed anyway, if there happens
not to be a conflict during the regression test run, we need assertions
here to keep us on the straight and narrow.
We might someday be able to remove the assertion in GetNewRelFileNode,
if pg_upgrade is rewritten to remove its assumption that old and new
relfilenodes always match. But it's hard to see how to get rid of the
pg_type OID constraint, since those OIDs are embedded in user tables
in some cases.
Back-patch as far as 9.5, because of the risk of back-patches breaking
something here even if it works in HEAD. I'd prefer to go back further,
but 9.4 fails both assertions due to get_rel_infos()'s use of a temporary
table. We can't use the later-branch solution of a CTE for compatibility
reasons (cf commit 5d16332e9), and it doesn't seem worth inventing some
other way to do the query. (I did check, by dint of changing the Asserts
to elog(WARNING), that there are no other cases of unwanted OID assignments
during 9.4's regression test run.)
Discussion: https://postgr.es/m/19785.1497215827@sss.pgh.pa.us
This reverts commit 56b6ef893fee9e9bf47d927a02f4d1ea911f4d9c and instead
makes vcregress.pl parse out PROVE_FLAGS from a command line argument
when doing a TAP test, thus making it consistent with the makefile
treatment.
Discussion: https://postgr.es/m/c26a7416-2fb9-34ab-7991-618c922f896e%402ndquadrant.com
Backpatch to 9.4 like previous patch.
Commit f039eaac7131ef2a4cf63a10cf98486f8bcd09d2, later back-patched
with commit 1b812afb0eafe125b820cc3b95e7ca03821aa675, allowed many of
the queries issued by postgres_fdw to fetch remote data to respond to
cancel interrupts in a timely fashion. However, it didn't do anything
about the transaction control commands, which remained
noninterruptible.
Improve the situation by changing do_sql_command() to retrieve query
results using pgfdw_get_result(), which uses the asynchronous
interface to libpq so that it can check for interrupts every time
libpq returns control. Since this might result in a situation
where we can no longer be sure that the remote transaction state
matches the local transaction state, add a facility to force all
levels of the local transaction to abort if we've lost track of
the remote state; without this, an apparently-successful commit of
the local transaction might fail to commit changes made on the
remote side. Also, add a 60-second timeout for queries issue during
transaction abort; if that expires, give up and mark the state of
the connection as unknown. Drop all such connections when we exit
the local transaction. Together, these changes mean that if we're
aborting the local toplevel transaction anyway, we can just drop the
remote connection in lieu of waiting (possibly for a very long time)
for it to complete an abort.
This still leaves quite a bit of room for improvement. PQcancel()
has no asynchronous interface, so if we get stuck sending the cancel
request we'll still hang. Also, PQsetnonblocking() is not used, which
means we could block uninterruptibly when sending a query. There
might be some other optimizations possible as well. Nonetheless,
this allows us to escape a wait for an unresponsive remote server
quickly in many more cases than previously.
Report by Suraj Kharage. Patch by me and Rafia Sabih. Review
and testing by Amit Kapila and Tushar Ahuja.
Discussion: http://postgr.es/m/CAF1DzPU8Kx+fMXEbFoP289xtm3bz3t+ZfxhmKavr98Bh-C0TqQ@mail.gmail.com
If authentication over an SSL connection fails, with sslmode=prefer,
libpq will reconnect without SSL and retry. However, we did not clear
the variables related to GSS, SSPI, and SASL authentication state, when
reconnecting. Because of that, the second authentication attempt would
always fail with a "duplicate GSS/SASL authentication request" error.
pg_SSPI_startup did not check for duplicate authentication requests like
the corresponding GSS and SASL functions, so with SSPI, you would leak
some memory instead.
Another way this could manifest itself, on version 10, is if you list
multiple hostnames in the "host" parameter. If the first server requests
Kerberos or SCRAM authentication, but it fails, the attempts to connect to
the other servers will also fail with "duplicate authentication request"
errors.
To fix, move the clearing of authentication state from closePGconn to
pgDropConnection, so that it is cleared also when re-connecting.
Patch by Michael Paquier, with some kibitzing by me.
Backpatch down to 9.3. 9.2 has the same bug, but the code around closing
the connection is somewhat different, so that this patch doesn't apply.
To fix this in 9.2, I think we would need to back-port commit 210eb9b743
first, and then apply this patch. However, given that we only bumped into
this in our own testing, we haven't heard any reports from users about
this, and that 9.2 will be end-of-lifed in a couple of months anyway, it
doesn't seem worth the risk and trouble.
Discussion: https://www.postgresql.org/message-id/CAB7nPqRuOUm0MyJaUy9L3eXYJU3AKCZ-0-03=-aDTZJGV4GyWw@mail.gmail.com
Because walsender and normal backends share the same main loop it's
problematic to have two different flag variables, set in signal
handlers, indicating a pending configuration reload. Only certain
walsender commands reach code paths checking for the
variable (START_[LOGICAL_]REPLICATION, CREATE_REPLICATION_SLOT
... LOGICAL, notably not base backups).
This is a bug present since the introduction of walsender, but has
gotten worse in releases since then which allow walsender to do more.
A later patch, not slated for v10, will similarly unify SIGHUP
handling in other types of processes as well.
Author: Petr Jelinek, Andres Freund
Reviewed-By: Michael Paquier
Discussion: https://postgr.es/m/20170423235941.qosiuoyqprq4nu7v@alap3.anarazel.de
Backpatch: 9.2-, bug is present since 9.0
When the checkpointer writes the shutdown checkpoint, it checks
afterwards whether any WAL has been written since it started and
throws a PANIC if so. At that point, only walsenders are still
active, so one might think this could not happen, but walsenders can
also generate WAL, for instance in BASE_BACKUP and logical decoding
related commands (e.g. via hint bits). So they can trigger this panic
if such a command is run while the shutdown checkpoint is being
written.
To fix this, divide the walsender shutdown into two phases. First,
checkpointer, itself triggered by postmaster, sends a
PROCSIG_WALSND_INIT_STOPPING signal to all walsenders. If the backend
is idle or runs an SQL query this causes the backend to shutdown, if
logical replication is in progress all existing WAL records are
processed followed by a shutdown. Otherwise this causes the walsender
to switch to the "stopping" state. In this state, the walsender will
reject any further replication commands. The checkpointer begins the
shutdown checkpoint once all walsenders are confirmed as
stopping. When the shutdown checkpoint finishes, the postmaster sends
us SIGUSR2. This instructs walsender to send any outstanding WAL,
including the shutdown checkpoint record, wait for it to be replicated
to the standby, and then exit.
Author: Andres Freund, based on an earlier patch by Michael Paquier
Reported-By: Fujii Masao, Andres Freund
Reviewed-By: Michael Paquier
Discussion: https://postgr.es/m/20170602002912.tqlwn4gymzlxpvs2@alap3.anarazel.de
Backpatch: 9.4, where logical decoding was introduced
The non-participation in procsignal was a problem for both changes in
master, e.g. parallelism not working for normal statements run in
walsender backends, and older branches, e.g. recovery conflicts and
catchup interrupts not working for logical decoding walsenders.
This commit thus replaces the previous WalSndXLogSendHandler with
procsignal_sigusr1_handler. In branches since db0f6cad48 that can
lead to additional SetLatch calls, but that only rarely seems to make
a difference.
Author: Andres Freund
Reviewed-By: Michael Paquier
Discussion: https://postgr.es/m/20170421014030.fdzvvvbrz4nckrow@alap3.anarazel.de
Backpatch: 9.4, earlier commits don't seem to benefit sufficiently
The current method of computing the record length (excluding the
lenght of full-page images) has been wrong since the WAL format has
been revamped in 2c03216d831160bedd72d45f712601b6f7d03f1c. Only the
main record's length was counted, but that can be significantly too
little if there's data associated with further blocks.
Fix by computing the record length as total_lenght - fpi_length.
Reported-By: Chen Huajun
Bug: #14687
Reviewed-By: Heikki Linnakangas
Discussion: https://postgr.es/m/20170603165939.1436.58887@wrigleys.postgresql.org
Backpatch: 9.5-
Some openssl builds put their lib files in a VC subdirectory, others do
not. Cater for both cases.
Backpatch to all live branches.
From an offline discussion with Leonardo Cecchi.
Previously this was not allowed, as copy.c didn't set the
CURSOR_OPT_PARALLEL_OK flag when planning the query. Set it.
While the lack of parallel query for COPY isn't strictly speaking a
bug, it does prevent parallelism from being used in a facility
commonly used to run long running queries. Thus backpatch to 9.6.
Author: Andres Freund
Discussion: https://postgr.es/m/20170531231958.ihanapplorptykzm@alap3.anarazel.de
Backpatch: 9.6, where parallelism was introduced.
On some platforms, -fpic fails for sufficiently large shared libraries.
We've mostly not hit that boundary yet, but there are some extensions
such as Citus and pglogical where it's becoming a problem. A bit of
research suggests that the penalty for -fPIC is small, in the
single-digit-percentage range --- and there's none at all on popular
platforms such as x86_64. So let's just default to -fPIC everywhere
and provide one less thing for extension developers to worry about.
Per complaint from Christoph Berg. Back-patch to all supported branches.
(I did not bother to touch the recently-removed Makefiles for sco and
unixware in the back branches, though. We'd have no way to test that
it doesn't break anything on those platforms.)
Discussion: https://postgr.es/m/20170529155850.qojdfrwkkqnjb3ap@msg.df7cb.de
Remove some gratuituous message differences by making the AM name
previously embedded in each message be a %s instead. While at it, get
rid of terminology that's unclear and unnecessary in one message.
Discussion: https://postgr.es/m/20170523001557.bq2hbq7hxyvyw62q@alvherre.pgsql
pg_resetwal (formerly pg_resetxlog) doesn't insist on finding a matching
version number in pg_control, and that seems like an important thing to
preserve since recovering from corrupt pg_control is a prime reason to
need to run it. However, that means you can try to run it against a
data directory of a different major version, which is at best useless
and at worst disastrous. So as to provide some protection against that
type of pilot error, inspect PG_VERSION at startup and refuse to do
anything if it doesn't match. PG_VERSION is read-only after initdb,
so it's unlikely to get corrupted, and even if it were corrupted it would
be easy to fix by hand.
This hazard has been there all along, so back-patch to all supported
branches.
Michael Paquier, with some kibitzing by me
Discussion: https://postgr.es/m/f4b8eb91-b934-8a0d-b3cc-68f06e2279d1@enterprisedb.com
The NumericOnly grammar production accepted ICONST, + ICONST, - ICONST,
FCONST, and - FCONST, but for some reason not + FCONST. This led to
strange inconsistencies like
regression=# set random_page_cost = +4;
SET
regression=# set random_page_cost = 4000000000;
SET
regression=# set random_page_cost = +4000000000;
ERROR: syntax error at or near "4000000000"
(because 4000000000 is too large to be an ICONST). While there's
no actual functional reason to need to write a "+", if we allow
it for integers it seems like we should allow it for numerics too.
It's been like that forever, so back-patch to all supported branches.
Discussion: https://postgr.es/m/30908.1496006184@sss.pgh.pa.us
Commit 9aa3c782c added code to allow CREATE TABLE/CREATE TYPE to not fail
when the desired type name conflicts with an autogenerated array type, by
dint of renaming the array type out of the way. But I (tgl) overlooked
that the same case arises in ALTER TABLE/TYPE RENAME. Fix that too.
Back-patch to all supported branches.
Report and patch by Vik Fearing, modified a bit by me
Discussion: https://postgr.es/m/0f4ade49-4f0b-a9a3-c120-7589f01d1eb8@2ndquadrant.com
If an operator class has no operators or functions, and doesn't need
a STORAGE clause, we emitted "CREATE OPERATOR CLASS ... AS ;" which
is syntactically invalid. Fix by forcing a STORAGE clause to be
emitted anyway in this case.
(At some point we might consider changing the grammar to allow CREATE
OPERATOR CLASS without an opclass_item_list. But probably we'd want to
omit the AS in that case, so that wouldn't fix this pg_dump issue anyway.)
It's been like this all along, so back-patch to all supported branches.
Daniel Gustafsson, tweaked by me to avoid a dangling-pointer bug
Discussion: https://postgr.es/m/D9E5FC64-7A37-4F3D-B946-7E4FB468F88A@yesql.se
This patch replaces isspace() calls with scanner_isspace() in functions
that are likely to be presented with non-ASCII input. isspace() has
the small advantage that it will correctly recognize no-break space
in single-byte encodings (such as LATIN1); but it cannot work successfully
for any multibyte character, and depending on platform it might return
false positive results for some fragments of multibyte characters. That's
disastrous for functions that are trying to discard whitespace between
valid strings, as noted in bug #14662 from Justin Muise. Even treating
no-break space as whitespace is pretty questionable for the usages touched
here, because the core scanner would think it is an identifier character.
Affected functions are parse_ident(), parseNameAndArgTypes (underlying
regprocedurein() and siblings), SplitIdentifierString (used for parsing
GUCs and options that are qualified names or lists of names), and
SplitDirectoriesString (used for parsing GUCs that are lists of
directories).
All the functions adjusted here are parsing SQL identifiers and similar
constructs, so it's reasonable to insist that their definition of
whitespace match the core scanner. So we can hope that this won't cause
many backwards-compatibility problems. I've left alone isspace() calls
in places that aren't really expecting any non-ASCII input characters,
such as float8in().
Back-patch to all supported branches.
Discussion: https://postgr.es/m/10129.1495302480@sss.pgh.pa.us
The cash_div_intX functions applied rint() to the result of the division.
That's not merely useless (because the result is already an integer) but
it causes precision loss for values larger than 2^52 or so, because of
the forced conversion to float8.
On the other hand, the cash_mul_fltX functions neglected to apply rint() to
their multiplication results, thus possibly causing off-by-one outputs.
Per C standard, arithmetic between any integral value and a float value is
performed in float format. Thus, cash_mul_flt4 and cash_div_flt4 produced
answers good to only about six digits, even when the float value is exact.
We can improve matters noticeably by widening the float inputs to double.
(It's tempting to consider using "long double" arithmetic if available,
but that's probably too much of a stretch for a back-patched fix.)
Also, document that cash_div_intX operators truncate rather than round.
Per bug #14663 from Richard Pistole. Back-patch to all supported branches.
Discussion: https://postgr.es/m/22403.1495223615@sss.pgh.pa.us
This is more secure, and saves a redirect since we no longer accept
plain HTTP connections on the website.
References in code comments should probably be updated too, but
that doesn't seem to need back-patching, whereas this does.
Also, in the 9.2 branch, remove suggestion that you can get the
source code via FTP, since that service will be shut down soon.
Daniel Gustafsson, with a few additional changes by me
Discussion: https://postgr.es/m/9A2C89A7-0BB8-41A8-B288-8B7BD09D7D44@yesql.se
When stdin is a terminal, it's possible to end a COPY FROM STDIN with
a keyboard EOF signal (typically control-D), and then keep on issuing
SQL commands. One would expect another COPY FROM STDIN to work as well,
but on some platforms it did not. This turns out to be because we were
not resetting the stream's feof() flag, and BSD-ish versions of fread()
and fgets() won't attempt to read more data if that's set.
The misbehavior is observed on BSDen (including macOS), but not Linux,
Windows, or SysV-ish Unixen, which makes this a portability bug not
just a missing feature.
Add a clearerr() call to fix the behavior, and improve the prompt that's
issued when copying from a TTY to mention that EOF signals work.
It's been like this forever, so back-patch to all supported branches.
Thomas Munro
Discussion: https://postgr.es/m/CAEepm=0MCGfYf=JAMiYhO6JPtv9-3ZfBo8fcGeCZ8oMzaw+Z+Q@mail.gmail.com
Modifying the permissions of a persistent file isn't really much nicer
than modifying its contents, even if git doesn't currently notice it.
Adjust the test script to make a copy and set the permissions of that
instead.
Michael Paquier, per a gripe from me. Back-patch to 9.5 where these
tests were introduced.
Discussion: https://postgr.es/m/14836.1494885946@sss.pgh.pa.us
The CommentStmt made by RebuildConstraintComment() has to pstrdup the
relation name, else it will contain a dangling pointer after that
relcache entry is flushed. (I'm less sure that pstrdup'ing conname
is necessary, but let's be safe.) Failure to do this leads to weird
errors or crashes, as reported by Marko Elezovic.
Bug introduced by commit e42375fc8, so back-patch to 9.5 as that was.
Fix by David Rowley, regression test by Michael Paquier
Discussion: https://postgr.es/m/DB6PR03MB30775D58E732D4EB0C13725B9AE00@DB6PR03MB3077.eurprd03.prod.outlook.com
Commit 60690a6fe attempted to fix the wait_for_stats() function in this
test so that it would wait properly if the tenk2 scans were done in
parallel workers instead of the main session (typically as a consequence of
force_parallel_mode being turned on). However, we made it test for whether
the main session's actions had been reported by looking for inserts on
'trunc_stats_test'. This is the Wrong Thing, because those aren't the last
updates we expect the main session to do. As shown by recent failures on
buildfarm member frogmouth, it's entirely likely that the trunc_stats_test
updates will be reported in a separate message from later updates, which
means there can be a window in which wait_for_stats() will exit but not all
the updates we are expecting to see will have arrived. We should test for
the last updates we're expecting, namely those on 'trunc_stats_test4'.
Unfortunately, I doubt that this explains frogmouth's failures, because
there's no reason to believe that it's running the tenk2 queries in
parallel. Still, the test is wrong on its own terms, so fix and back-patch
to 9.6 where parallel query came in.
Ultra-modern versions of the perl Data::Dumper module have apparently
changed how they indent output. Instead of trying to keep up we choose
to tell it to supporess all indentation in the hstore_plperl regression
tests.
Backpatch to 9.5 where this feature was introduced.
Before 955a684e0401 logical decoding snapshot maintenance needed to
cope with transactions it might not have seen in their entirety. For
such transactions we'd to assume they modified the catalog (could have
happened before we were watching), and thus a new snapshot had to be
built, and distributed to concurrently running transactions.
That's problematic because building a new snapshot isn't that cheap ,
especially as the the array of committed transactions needs to be
sorted. When creating a slot on a server with a lot of transactions,
this could make logical slot creation infeasibly expensive.
After 955a684e0401 there's no need to deal with transaction that
aren't guaranteed to be fully observable. That allows to avoid
building snapshots for transactions that haven't modified catalog,
even before reaching consistency.
While this isn't necessarily a bugfix, slot creation being impossible
in some production workloads, is severe enough to warrant
backpatching.
Author: Andres Freund, based on a quite different patch from Petr Jelinek
Analyzed-By: Petr Jelinek
Reviewed-By: Petr Jelinek
Discussion: https://postgr.es/m/f37e975c-908f-858e-707f-058d3b1eb214@2ndquadrant.com
Backpatch: 9.4-, where logical decoding has been introduced
The snapshot assembly during the creation of logical slots relied
waiting for transactions in xl_running_xacts to end, by checking for
their commit/abort records. Unfortunately, despite locking, it is
possible to see an xl_running_xact record listing transactions as
ready, that have already WAL-logged an commit/abort record, as the
locking just prevents the ProcArray to be adjusted, and the commit
record has to be logged first.
That lead to either delayed or hanging snapshot creation, because
snapbuild.c would wait "forever" to see commit/abort records for some
transactions. That hang resolved only if a xl_running_xacts record
without any running transactions happened to be logged, far from
certain on a busy server.
It's impractical to prevent that via more heavyweight locking, the
likelihood of deadlocks and significantly increased contention would
be too big.
Instead change the initial snapshot creation to be solely based on
tracking the oldest running transaction via
xl_running_xacts->oldestRunningXid - that actually ends up
significantly simplifying the code. That has two disadvantages:
1) Because we cannot fully "trust" the contents of xl_running_xacts,
we cannot use it to build the initial snapshot. Instead we have to
wait twice for all running transactions to finish.
2) Previously a slot, unless the race occurred, could be created when
the all transaction perceived as running based on commit/abort
records, now we have to wait for the next xl_running_xacts record.
To address that, trigger logging new xl_running_xacts record from
within snapbuild.c exactly when necessary.
Unfortunately snabuild.c's SnapBuild is stored on disk, one of the
stupider ideas of a certain Mr Freund, so we can't change it in a
minor release. As this is going to be backpatched, we have to hack
around a bit to keep on-disk compatibility. A later commit will
rejigger that on master.
Author: Andres Freund, based on a quite different patch from Petr Jelinek
Analyzed-By: Petr Jelinek
Reviewed-By: Petr Jelinek
Discussion: https://postgr.es/m/f37e975c-908f-858e-707f-058d3b1eb214@2ndquadrant.com
Backpatch: 9.4-, where logical decoding has been introduced
We have now grown enough registerable syscache-invalidation callback
functions that the original assumption that there would be few of them
is causing performance problems. In particular, let's fix things so that
CallSyscacheCallbacks doesn't have to search the whole array to find
which callback(s) to invoke for a given cache ID. Preserve the original
behavior that callbacks are called in order of registration, just in
case there's someplace that depends on that (which I doubt).
In support of this, export the number of syscaches from syscache.h.
People could have found that out anyway from the enum, but adding a
#define makes that much safer.
This provides a useful additional speedup in Mathieu Fenniak's
logical-decoding test case, although we're reaching the point of
diminishing returns there. I think any further improvement will have
to come from reducing the number of cache invalidations that are
triggered in the first place. Still, we can hope that this change
gives some incremental benefit for all invalidation scenarios.
Back-patch to 9.4 where logical decoding was introduced.
Discussion: https://postgr.es/m/CAHoiPjzea6N0zuCi=+f9v_j94nfsy6y8SU7-=bp4=7qw6_i=Rg@mail.gmail.com
This has to be backpatched to all supported releases so release markup
added to HEAD and copied to back branches matches the existing markup.
Reported-by: Peter Eisentraut
Discussion: 2b8a2552-fffa-f7c8-97c5-14db47a87731@2ndquadrant.com
Author: initial patch and sample markup by Peter Eisentraut
Backpatch-through: 9.2
A test case provided by Mathieu Fenniak shows that hash_seq_search'ing
this hashtable can consume a very significant amount of overhead during
logical decoding, which triggers frequent cache invalidation. Testing
suggests that the actual population of the hashtable is often no more
than a few dozen entries, so we can cut the overhead just by dropping
the initial number of buckets down from 1024 --- I chose to cut it to 64.
(In situations where we do have a significant number of entries, we
shouldn't get any real penalty from doing this, as the dynahash.c code
will resize the hashtable automatically.)
This gives a further factor-of-two savings in Mathieu's test case.
That may be overly optimistic for real-world benefit, as real cases
may have larger average table populations, but it's hard to see it
turning into a net negative for any workload.
Back-patch to 9.4 where relfilenodemap.c was introduced.
Discussion: https://postgr.es/m/CAHoiPjzea6N0zuCi=+f9v_j94nfsy6y8SU7-=bp4=7qw6_i=Rg@mail.gmail.com
A test case provided by Mathieu Fenniak shows that the initial search for
the target catcache in CatalogCacheIdInvalidate consumes a very significant
amount of overhead in cases where cache invalidation is triggered but has
little useful work to do. There is no good reason for that search to exist
at all, as the index array maintained by syscache.c allows direct lookup of
the catcache from its ID. We just need a frontend function in syscache.c,
matching the division of labor for most other cache-accessing operations.
While there's more that can be done in this area, this patch alone reduces
the runtime of Mathieu's example by 2X. We can hope that it offers some
useful benefit in other cases too, although usually cache invalidation
overhead is not such a striking fraction of the total runtime.
Back-patch to 9.4 where logical decoding was introduced. It might be
worth going further back, but presently the only case we know of where
cache invalidation is really a significant burden is in logical decoding.
Also, older branches have fewer catcaches, reducing the possible benefit.
(Note: although this nominally changes catcache's API, we have always
documented CatalogCacheIdInvalidate as a private function, so I would
have little sympathy for an external module calling it directly. So
backpatching should be fine.)
Discussion: https://postgr.es/m/CAHoiPjzea6N0zuCi=+f9v_j94nfsy6y8SU7-=bp4=7qw6_i=Rg@mail.gmail.com
On MSVC builds and on back branches that means removing the hardcoded
--verbose setting. On master for Unix that means removing the empty
setting in the global Makefile so that the value can be acquired from
the environment as well as from the make arguments.
Backpatch to 9.4 where we introduced TAP tests
On Unix this path is detected via the use of xml2-config, but that's not
available on Windows. This means that users building with libxml2 will
no longer need to move things around from the standard libxml2
installation for MSVC builds.
Backpatch to all live branches.
