On Tue, Nov 16, 2021 at 10:26:11PM -0500, Tom Lane wrote:
> I feel like doing an immediate exit() for these errors and not other
> ones is a pretty terrible idea, mainly because it doesn't account for
> the question of what to do with a failure that prevents us from getting
> to the fsync() call in the first place. So I'd like to see a better
> design rather than more quick hacking. I confess I don't have a
> clear idea of what "a better design" would look like.
[ .. thinks .. ]
We cannot really have something equivalent to data_sync_retry in the
frontends. But I'd like to think that it would be fine for
pg_basebackup to just exit() on this failure so as callers would be
able to retry a base backup. pg_receivewal is more tricky though. An
exit() would allow for flush retries of a previous WAL segment where
things failed, but that stands when using --no-loop (still the code
path triggered by this option would not be used). When not using
--no-loop, it may be better to actually just retry streaming from the
previous point so as the error should be reported from walmethods.c to
the upper stack anyway.
> However, that's largely orthogonal to any of the things my proposed
> patches are trying to fix. If you want to review the patches without
> considering the fsync-error-handling problem, that'd be great.
I have looked at them upthread, FWIW:
https://www.postgresql.org/message-id/YYtSj5vlWp5faVXz@paquier.xyz
Your proposals still look rather sane to me, after a second look.
--
Michael