Re: BUG #18146: Rows reappearing in Tables after Auto-Vacuum Failure in PostgreSQL on Windows

On Thu, Oct 12, 2023 at 5:59 AM Robert Haas <robertmhaas@gmail.com> wrote:
> Suppose that RelationTruncate set both DELAY_CHKPT_START and
> DELAY_CHKPT_COMPLETE. I think that would prevent this problem. P2
> could still choose the redo LSN after P1 logged the truncate, but it
> wouldn't then be able to reach CheckPointBuffers() until after P1 had
> reached RegisterSyncRequest. [...]

Thanks for thinking about this.  Yeah, the existing _START flag does
indeed seem to be enough.  I'd been focusing on trying to control the
redo point selection, but we just need to delay ProcessSyncRequests(),
and we had a hammer for that already.  Who wants to write the patch?
It should be trivial, except for the comments.

It's interesting that we'd stun the checkpointer just before we try to
send it a request in a fixed size queue, but that's OK because we'll
perform the fsync ourselves if the queue is full.