This reverts much of commit f03a9ca436,
but leaves the relpages/reltuples probe in select_parallel.sql.
The pg_stat_all_tables probes are unstable enough to be annoying,
and it no longer seems likely that they will teach us anything more
about the underlying problem. I'd still like some more confirmation
though that the observed plan instability is caused by VACUUM leaving
relpages/reltuples as zero for one of these tables.
Discussion: https://postgr.es/m/CA+hUKG+0CxrKRWRMf5ymN3gm+BECHna2B-q1w8onKBep4HasUw@mail.gmail.com
We're seeing occasional instability in the plans generated for
parallel queries on the "a_star" table hierarchy. This suggests
that something is changing the planner's stats for those tables,
but that should not be happening within a regression test run.
To try to gather some information about what's happening, insert
additional queries to check the basic page/tuple counts for these
tables, as well as whether any vacuums or analyzes have happened
on them. (We expect that only the database-wide VACUUM in
sanity_check.sql will have touched them.)
I added the probes not only in select_parallel.sql itself, but
also in stats.sql, bearing in mind that the stats collector's
lag may prevent the initial query from reporting current truth.
If any extra vacuum/analyze has happened, the recheck in stats.sql
definitely ought to see it.
This commit can be reverted once we figure out what's going on.
Per suggestion from David Rowley, though I changed the queries around.
Discussion: https://postgr.es/m/CA+hUKG+0CxrKRWRMf5ymN3gm+BECHna2B-q1w8onKBep4HasUw@mail.gmail.com
The stanza of ExecuteTruncate[Guts] that truncates a target table's toast
relation re-used the loop local variable "rel" to reference the toast rel.
This was safe enough when written, but commit d42358efb added code below
that that supposed "rel" still pointed to the parent table. Therefore,
the stats counter update was applied to the wrong relcache entry (the
toast rel not the user rel); and if we were unlucky and that relcache
entry had been flushed during reindex_relation, very bad things could
ensue.
(I'm surprised that CLOBBER_CACHE_ALWAYS testing hasn't found this.
I'm even more surprised that the problem wasn't detected during the
development of d42358efb; it must not have been tested in any case
with a toast table, as the incorrect stats counts are very obvious.)
To fix, replace use of "rel" in that code branch with a more local
variable. Adjust test cases added by d42358efb so that some of them
use tables with toast tables.
Per bug #15540 from Pan Bian. Back-patch to 9.5 where d42358efb came in.
Discussion: https://postgr.es/m/15540-01078812338195c0@postgresql.org
If we don't have to return any columns from heap tuples, and there's
no need to recheck qual conditions, and the heap page is all-visible,
then we can skip fetching the heap page altogether.
Skip prefetching pages too, when possible, on the assumption that the
recheck flag will remain the same from one page to the next. While that
assumption is hardly bulletproof, it seems like a good bet most of the
time, and better than prefetching pages we don't need.
This commit installs the executor infrastructure, but doesn't change
any planner cost estimates, thus possibly causing bitmap scans to
not be chosen in cases where this change renders them the best choice.
I (tgl) am not entirely convinced that we need to account for this
behavior in the planner, because I think typically the bitmap scan would
get chosen anyway if it's the best bet. In any case the submitted patch
took way too many shortcuts, resulting in too many clearly-bad choices,
to be committable.
Alexander Kuzmenkov, reviewed by Alexey Chernyshov, and whacked around
rather heavily by me.
Discussion: https://postgr.es/m/239a8955-c0fc-f506-026d-c837e86c827b@postgrespro.ru
pg_stat_get_snapshot_timestamp() returns the timestamp seen in the "global"
stats file. Because pgstat_write_statsfiles() writes per-DB stats files
before the global file (or at least before renaming it into place), there
is a window where the test backend can see all the stats updates that
wait_for_stats() was checking for (all of which come from the per-DB file)
but also see the same global stats file it had seen at the start of the
test script. This results in a failure in only the "snapshot_newer" query,
as reported by a couple of buildfarm members recently.
I suspect that this ought to be back-patched. Commit 4e37b3e15 has
evidently increased the probability of this window getting hit, but
it's not apparent why it could not have been hit before. I'll refrain
for the moment though.
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.
On faster machines, the overall runtime for running the core regression
tests is under twenty seconds these days, of which the hard-wired delays
in the stats test are a significant fraction. But on closer inspection,
it seems like we shouldn't need those.
The initial 2-second delay is there only to reduce the risk of the test's
stats messages not getting sent due to contention. But analysis of the
last ten years' worth of buildfarm runs shows no evidence that such
failures actually occur. (We do see failures that look like stats
messages not getting sent, particularly on Windows; but there is little
reason to believe that the initial delay reduces their frequency.)
The later 1-second delay is there to ensure that our session's stats
will have gotten sent. But we could also do that by starting a fresh
session, which takes well under 1 second even on very slow machines.
Hence, let's remove both delays and see what happens. The first delay
was the only test of pg_sleep_for() in the regression tests, but we can
move that responsibility into wait_for_stats().
Discussion: https://postgr.es/m/17795.1493869423@sss.pgh.pa.us
Historically, the wait_for_stats() function in this test has simply checked
for a report of an indexscan on tenk2, corresponding to the last command
issued before we expect stats updates to appear. However, with parallel
query that indexscan could be done by a parallel worker that will emit
its stats counters to the collector before the session's main backend does
(a full second before, in fact, thanks to the "pg_sleep(1.0)" added by
commit 957d08c81f). That leaves a sizable window in which an
autovacuum-triggered write of the stats files would present a state in
which the indexscan on tenk2 appears to have been done, but none of the
write updates performed by the test have been. This is evidently the
explanation for intermittent failures seen by me and on buildfarm member
mandrill.
To fix, we should check separately for both the tenk2 seqscan and indexscan
counts, since those might be reported by different processes that could be
delayed arbitrarily on an overloaded test machine. And we need to check
for at least one update-related count. If we ever allow parallel workers
to do writes, this will get even more complicated ... but in view of all
the other hard problems that will entail, I don't feel a need to solve this
one today.
Per research by Rahila Syed and myself; part of this patch is Rahila's.
This works by keeping a per-subtransaction record of the ins/upd/del
counters before the truncate, and then resetting them; this record is
useful to return to the previous state in case the truncate is rolled
back, either in a subtransaction or whole transaction. The state is
propagated upwards as subtransactions commit.
When the per-table data is sent to the stats collector, a flag indicates
to reset the live/dead counters to zero as well.
Catalog version bumped due to the change in pgstat format.
Author: Alexander Shulgin
Discussion: 1007.1207238291@sss.pgh.pa.us
Discussion: 548F7D38.2000401@BlueTreble.com
Reviewed-by: Álvaro Herrera, Jim Nasby
Per discussion, this could be useful for purposes such as programmatically
detecting a nonresponding stats collector. We already have the timestamp
anyway, it's just a matter of providing a SQL-accessible function to fetch
it.
Matt Kelly, reviewed by Jim Nasby
This bollixes the test because it's expecting to see the idx_tup_fetch
counter increase, which won't happen if heap fetches were avoided by use
of an index-only scan. Per buildfarm results.
While at it, let's just make sure that enable_seqscan and enable_indexscan
are ON for this test ...
* stats_start_collector goes away; we always start the collector process,
unless prevented by a problem with setting up the stats UDP socket.
* stats_reset_on_server_start goes away; it seems useless in view of the
availability of pg_stat_reset().
* stats_block_level and stats_row_level are merged into a single variable
"track_counts", which controls all reports sent to the collector process.
* stats_command_string is renamed to track_activities.
* log_autovacuum is renamed to log_autovacuum_min_duration to better reflect
its meaning.
The log_autovacuum change is not a compatibility issue since it didn't exist
before 8.3 anyway. The other changes need to be release-noted.
messages to the stats collector. This avoids the problem that enabling
stats_row_level for autovacuum has a significant overhead for short
read-only transactions, as noted by Arjen van der Meijden. We can avoid
an extra gettimeofday call by piggybacking on the one done for WAL-logging
xact commit or abort (although that doesn't help read-only transactions,
since they don't WAL-log anything).
In my proposal for this, I noted that we could change the WAL log entries
for commit/abort to record full TimestampTz precision, instead of only
time_t as at present. That's not done in this patch, but will be committed
separately.
already collected in the current transaction; this allows plpgsql functions to
watch for stats updates even though they are confined to a single transaction.
Use this instead of the previous kluge involving pg_stat_file() to wait for
the stats collector to update in the stats regression test. Internally,
decouple storage of stats snapshots from transaction boundaries; they'll
now stick around until someone calls pgstat_clear_snapshot --- which xact.c
still does at transaction end, to maintain the previous behavior. This makes
the logic a lot cleaner, at the price of a couple dozen cycles per transaction
exit.
changes (with an upper limit of 30 seconds), and record the delay time in
the postmaster log. This should give us some info about what's happening
with the intermittent stats failures in buildfarm. After an idea of
Andrew Dunstan's.
activity quiesce. Possibly this will fix the large increase in
non-reproducible stats test failures we've noted since turning on
stats_row_level by default.
generated by bitmap index scans. Along the way, simplify and speed up
the code for counting sequential and index scans; it was both confusing
and inefficient to be taking care of that in the per-tuple loops, IMHO.
initdb forced because of internal changes in pg_stat view definitions.
but the code is basically working. Along the way, rewrite the entire
approach to processing OR index conditions, and make it work in join
cases for the first time ever. orindxpath.c is now basically obsolete,
but I left it in for the time being to allow easy comparison testing
against the old implementation.
Tom Lane
Alvaro Herrera
Robert Haas
Bruce Momjian