I wrote:
> Andres Freund <andres@anarazel.de> writes:
>> What about adding a pg_fatal() that's pg_log_fatal() + exit()? That keeps
>> pg_log_* stuff "log only", but adds something adjacent enough to hopefully
>> reduce future misunderstandings?
> I'd be okay with that, except that pg_upgrade already has a pg_fatal
> (because it has its *own* logging system, just in case you thought
> this wasn't enough of a mess yet).
Wait a moment. After looking closer, I realize that pg_upgrade's
pg_fatal could trivially be turned into a macro; and the other two
existing definitions already are macros. That would remove the risk
of link-time symbol collisions that I was worried about. As a bonus,
it'd substantially reduce the number of changes needed to make
pg_upgrade use logging.c, whenever somebody wants to make that happen.
So I now propose modifying yesterday's patch thus:
* Reinstantiate the PG_LOG_FATAL enum value, add support macros
pg_log_fatal, pg_log_fatal_hint, pg_log_fatal_detail.
* Define pg_fatal as pg_log_fatal + exit(1). (This would essentially
move pg_rewind's definition into logging.h. pg_upgrade will
define it slightly differently, but the semantics end up the same.)
* Adjust call sites to match.
I do like this idea because it would not break any existing code
that expects pg_log_fatal to return. There is likely to be some
of that in outstanding patches, and this approach would merely
render it less-than-idiomatic rather than outright broken.
Updating the patch is going to be a bit tedious, so I'm not
going to do it without buy-in that this solution would be
okay to commit. Any objections?
regards, tom lane