Il 04/10/19 18:28, Alvaro Herrera ha scritto:
> I wonder if it would work to just clear that multixact with
> SELECT ... WHERE ctid='(3160,31)' FOR UPDATE
select ...what? :-) Sorry but it's totally beyond my knowledge and my
control.... after resolving the issue i'll surely go and search docs to
understand what we've done....
>
> If this was in my hands, I would scan the WAL looking for the place that
> last touched this page (and the latest FPI for this page, also). It
> might have an explanation of what went on. Maybe use the page's LSN
> (from pageinspect's page_header()) as starting point for the WAL
> location that modified the page. I hope you have a WAL archive that
> goes back to well before the previous checkpoint.
One thing I forgot to report is that this cluster is just upgraded from
a 9.1 that was crashing at least once a day (in many cases the upgrade
itself resolved the issue)
here's the log line
2019-10-03 15:11:52 CEST LOG: server process (PID 18668) was terminated
by exception 0xC0000005
In this case probably the access violation was due to a data corruption.
These are customer machines that are really badly kept and NTFS issues
are not that rare, so I won't bother investigating what's happened but
just make the customer up & running again.
Thanks for your time
Moreno
>
