On Mon, Aug 8, 2022 at 7:12 AM Noah Misch <noah@leadboat.com> wrote:
> On Sun, Aug 07, 2022 at 03:18:52PM +0200, Tomas Vondra wrote:
> > I'd bet it's about WAL prefetching, not the revert, and the bisect was a
> > bit incorrect, because the commits are close and the failures happen to
> > be rare. (Presumably you first did the bisect and then wrote the patch
> > that reproduces this, right?)
>
> No. I wrote the patch, then used the patch to drive the bisect. With ten
> iterations, commit 2c7ea57 passes 0/10, while 2c7ea57^ passes 10/10. I've now
> tried recovery_prefetch=off. With that, the test passes 10/10 at 2c7ea57.
> Given your observation of a failure at 5dc0418fab2, I agree with your
> conclusion. Whatever the role of 2c7ea57 in exposing the failure on my
> machine, a root cause in WAL prefetching looks more likely.
>
> > Adding Thomas Munro to the thread, he's the WAL prefetching expert ;-)
Thanks for the repro patch and bisection work. Looking...
