On Mon, Sep 12, 2022 at 1:42 PM Justin Pryzby <pryzby@telsasoft.com> wrote:
> On Mon, Sep 12, 2022 at 10:44:38AM +1200, Thomas Munro wrote:
> > On Sat, Sep 10, 2022 at 5:44 PM Justin Pryzby <pryzby@telsasoft.com> wrote:
> > > < 2022-09-09 19:37:25.835 CDT telsasoft >ERROR: MultiXactId 133553154 has not been created yet -- apparent
wraparound
> >
> > I guess what happened here is that after one of your (apparently
> > several?) OOM crashes, crash recovery didn't run all the way to the
> > true end of the WAL due to the maintenance_io_concurrency=0 bug. In
> > the case you reported, it couldn't complete an end-of-recovery
> > checkpoint until you disabled recovery_prefetch, but that's only
> > because of the somewhat unusual way that vismap pages work. In
> > another case it might have been able to (bogusly) complete a
> > checkpoint, leaving things in an inconsistent state.
>
> I think you're saying is that this can be explained by the
> io_concurrency bug in recovery_prefetch, if run under 15b3.
Well I don't know, but it's one way I could think of that you could
have a data page referring to a multixact that isn't on disk after
recovery (because the data page happens to have been flushed, but we
didn't replay the WAL that would create the multixact).
> But yesterday I started from initdb and restored this cluster from backup, and
> started up sqlsmith, and sent some kill -9, and now got more corruption.
> Looks like it took ~10 induced crashes before this happened.
$SUBJECT says 15b4, which doesn't have the fix. Are you still using
maintainance_io_concurrent=0?