On Sat, Oct 14, 2023 at 11:33 AM David Steele <david@pgmasters.net> wrote:
> All of this is fixable in HEAD, but seems incredibly dangerous to back
> patch. Even so, I have attached the patch in case somebody sees an
> opportunity that I do not.
I really do not think we should be even thinking about back-patching
something like this. It's clearly not a bug fix, although I'm sure
that someone can try to characterize it that way, if they want to make
the well-worn argument that any behavior they don't like is a bug. But
that's a pretty lame argument. Usage errors on the part of users are
not bugs, even if we've coded the software in such a way as to make
those errors more likely.
I think what we ought to be talking about is whether a change like
this is a good idea even in master. I don't think it's a terrible
idea, but I'm also not sure that it's a good idea. The problem is that
if you're doing the right thing with your backup_label, then this is
unnecessary, and if you're doing the wrong thing, then why should you
do the right thing about this? I mean, admittedly you can't just
ignore a fatal error, but I think people will just run pg_resetwal,
which is even worse than starting from the wrong checkpoint. I feel
like in cases where a customer I'm working with has a bad backup,
their entire focus is on doing something to that backup to get a
running system back, whatever it takes. It's already too late at that
point to fix the backup procedure - they only have the backups they
have. You could hope people would do test restores before disaster
strikes, but people who are that prepared are probably running a real
backup tool and will never have this problem in the first place.
Perhaps that's all too pessimistic. I don't know. Certainly, other
people can have experiences that are different than mine. But I feel
like I struggle to think of a case where this would have prevented a
bad outcome, and that makes me wonder whether it's really a good idea
to complicate the system.
--
Robert Haas
EDB: http://www.enterprisedb.com