Re: Fwd: Re: A new look at old NFS readdir() problems?

On 1/6/25 19:08, Tom Lane wrote:
> David Steele <david(at)pgbackrest(dot)org> writes:
>> On 1/4/25 11:07, Thomas Munro wrote:
>>> As for CIFS, there are lots of reports of this sort of thing from
>>> Linux CIFS clients.
>
>> There may be users running Postgres on CIFS but my guess is that is rare
>> -- at least I have never seen anyone doing it.
>
> It'd be news to me too. I wondered if I could test it locally, but
> while my NAS knows half a dozen such protocols it's never heard of
> CIFS.

You may also know it as SMB or Samba. pgBackRest skips directory fsyncs
if the repository type is set to CIFS so I think we'd know if anybody
was running on CIFS as I'm fairly certain a directory fsync will return
a hard error. That may be implementation dependent, though.

>> I'm more concerned about the report we saw on SUSE/NFS [1]. If that
>> report is accurate it indicates this may not be something we can just
>> document and move on from -- unless we are willing to entirely drop
>> support for NFS.
>> [1] https://github.com/pgbackrest/pgbackrest/issues/1423
>
> I installed an up-to-date OpenSUSE image (Leap 15.6) and it passes
> my "rmtree" test just fine with my NAS. The report you cite
> doesn't have any details on what the NFS server was, but I'd be
> inclined to guess that that server's filesystem lacked support
> for stable NFS cookies.

The internal report we received might have had a similar cause. Sure
seems like a minefield for any user trying to figure out if their setup
is compliant, though. In many setups (especially production) a drop
database is rare.

Regards,
-David