Re: pg15b3: recovery fails with wal prefetch enabled

On Thu, Sep 1, 2022 at 2:01 AM Justin Pryzby <pryzby@telsasoft.com> wrote:
> < 2022-08-31 08:44:10.495 CDT  >LOG:  checkpoint starting: end-of-recovery immediate wait
> < 2022-08-31 08:44:10.609 CDT  >LOG:  request to flush past end of generated WAL; request 1201/1CAF84F0, current
position1201/1CADB730
 
> < 2022-08-31 08:44:10.609 CDT  >CONTEXT:  writing block 0 of relation base/16881/2840_vm
> < 2022-08-31 08:44:10.609 CDT  >ERROR:  xlog flush request 1201/1CAF84F0 is not satisfied --- flushed only to
1201/1CADB730
> < 2022-08-31 08:44:10.609 CDT  >CONTEXT:  writing block 0 of relation base/16881/2840_vm
> < 2022-08-31 08:44:10.609 CDT  >FATAL:  checkpoint request failed
>
> I was able to start it with -c recovery_prefetch=no, so it seems like
> prefetch tried to do too much.  The VM runs centos7 under qemu.
> I'm making a copy of the data dir in cases it's needed.

Hmm, a page with an LSN set 118208 bytes past the end of WAL.  It's a
vm fork page (which recovery prefetch should ignore completely).  Did
you happen to get a copy before the successful recovery?  After the
successful recovery, what LSN does that page have, and can you find
the references to it in the WAL with eg pg_waldump -R 1663/16681/2840
-F vm?  Have you turned FPW off (perhaps this is on ZFS?)?