On 2021-08-31 23:31:15 +0000, Bossart, Nathan wrote:
> On 8/31/21, 1:30 PM, "Andres Freund" <andres@anarazel.de> wrote:
> > On 2021-08-31 18:09:36 +0000, Bossart, Nathan wrote:
> >> What appears to happen in this case is that bgwriter eventually creates a
> >> xl_running_xacts record and nudges walwriter to flush it to disk, at which
> >> point the .ready file(s) will be created. That's admittedly a bit fragile.
> >
> > That's not guaranteed to happen. If e.g. the partial record is a checkpoint or
> > a xl_running_xacts, we'll not trigger further WAL writes in the background,
> > unless autovacuum ends up doing something.
>
> Right. Per the attached patch, a simple way to handle that could be
> to teach XLogBackgroundFlush() to flush to the "earliest" segment
> boundary if it doesn't find anything else to do. I think you could
> still miss creating a .ready file for the previous segment in single-
> user mode, though.
Maybe, but this is getting uglier and uglier.
I think patch should be reverted. It's not in a state that's appropriate for
the backbranches.