Horton

Re: wake up logical workers after ALTER SUBSCRIPTION

Nathan Bossart <nathandbossart@gmail.com>

View thread View on postgr.es
From: Nathan Bossart <nathandbossart@gmail.com>
To: Tom Lane <tgl@sss.pgh.pa.us>
Cc: Melih Mutlu <m.melihmutlu@gmail.com>, Thomas Munro <thomas.munro@gmail.com>, "Hayato Kuroda (Fujitsu)" <kuroda.hayato@fujitsu.com>, "pgsql-hackers@postgresql.org" <pgsql-hackers@postgresql.org>
Date: 2022-12-15T22:47:21Z
Lists: pgsql-hackers

Attachments

I tried setting wal_retrieve_retry_interval to 1ms for all TAP tests
(similar to what was done in 2710ccd), and I noticed that the recovery
tests consistently took much longer.  Upon further inspection, it looks
like the same (or a very similar) race condition described in e5d494d's
commit message [0].  With some added debug logs, I see that all of the
callers of MaybeStartWalReceiver() complete before SIGCHLD is processed, so
ServerLoop() waits for a minute before starting the WAL receiver.

A simple fix is to have DetermineSleepTime() take the WalReceiverRequested
flag into consideration.  The attached 0002 patch shortens the sleep time
to 100ms if it looks like we are waiting on a SIGCHLD.  I'm not certain
this is the best approach, but it seems to fix the tests.

On my machine, I see the following improvements in the tests (all units in
seconds):
                     HEAD  patched (v9)
    check-world -j8  165   138
    subscription     120   75
    recovery         111   108

[0] https://postgr.es/m/21344.1498494720%40sss.pgh.pa.us

-- 
Nathan Bossart
Amazon Web Services: https://aws.amazon.com

Commits

  1. Avoid type cheats for invalid dsa_handles and dshash_table_handles.

  2. Track logrep apply workers' last start times to avoid useless waits.

  3. Check for two_phase change at end of process_syncing_tables_for_apply.

  4. Wake up a subscription's replication worker processes after DDL.