On Fri, Jul 05, 2024 at 01:52:31PM +0900, Michael Paquier wrote:
> On Sat, Jun 29, 2024 at 11:13:04PM +0200, Tomas Vondra wrote:
>> I think those are two independent issues - knowing that the snapshot is
>> from the last checkpoint, and knowing that it's correct (not corrupted).
>> And yeah, we should be careful about fsync/durable_rename.
>
> Yeah, that's bugging me as well. I don't really get why we would not
> want durability at shutdown for this data. So how about switching the
> end of pgstat_write_statsfile() to use durable_rename()? Sounds like
> an independent change, worth on its own.
Please find attached a rebased patch set with the durability point
addressed in 0001. There were also some conflicts.
Note that I have applied the previous 0002 adding an assert in
pgstat_write_statsfile() as 734c057a8935, as I've managed to break
again this assumption while hacking more on this area..
--
Michael