Horton

Re: Refactor recovery conflict signaling a little

Xuneng Zhou <xunengzhou@gmail.com>

View thread View on postgr.es
From: Xuneng Zhou <xunengzhou@gmail.com>
To: Heikki Linnakangas <hlinnaka@iki.fi>
Cc: Alexander Lakhin <exclusion@gmail.com>, "pgsql-hackers@postgresql.org" <pgsql-hackers@postgresql.org>, Chao Li <li.evan.chao@gmail.com>
Date: 2026-03-09T17:05:26Z
Lists: pgsql-hackers

Attachments

On Mon, Mar 9, 2026 at 11:28 PM Heikki Linnakangas <hlinnaka@iki.fi> wrote:
>
> On 09/03/2026 17:02, Xuneng Zhou wrote:
> > Did you use Alexander’s reproducer script? I tried reproducing with a
> > 1 ms pg_usleep() added to all three functions that clear
> > MyProc->pendingRecoveryConflicts, but I still couldn’t reproduce the
> > issue.
>
> I used the attached, to be precise. With that it fails every time for
> me. I'm not sure if the "if (am_walsender)" check is necessary, I added
> it just to make the test run faster.
>
> - Heikki

I was able to reproduce the issue using a wider sleep window as you
suggested and can confirm that the flag is not cleared after applying
the patch. Below are two logs—one from a successful run and one from a
failed run. I'll look further into the patch later on.

failed run:
startup[1418915] LOG:  DBG SignalRecoveryConflict target_pid=1419118
reason=4 old_mask=0x0 new_mask=0x10
walsender[1419118] LOG:  DBG ProcArrayEndTransaction(no-xid) CLEARING
pendingRecoveryConflicts=0x10

successful run:
startup[1433218] LOG:  DBG SignalRecoveryConflict target_pid=1433406
reason=4 old_mask=0x0 new_mask=0x10
walsender[1433406] LOG:  DBG ProcessInterrupts handler fired 1
time(s), pending=0x10 -- processing
walsender[1433406] ERROR:  canceling statement due to conflict with recovery

-- 
Best,
Xuneng

Commits

  1. Don't clear pendingRecoveryConflicts at end of transaction

  2. Refactor ProcessRecoveryConflictInterrupt for readability

  3. Separate RecoveryConflictReasons from procsignals

  4. Use ProcNumber rather than pid in ReplicationSlot

  5. Don't hint that you can reconnect when the database is dropped

  6. Remove useless errdetail_abort()

  7. Teach standby conflict resolution to use SIGUSR1