> The checkpointer is entirely incapable of either detecting the problem
> (it doesn't have enough infrastructure to examine pg_database in a
> reasonable way) or preventing backends from doing anything if it did
> know there was a problem.
Well, I guess I meant 'some regularly running process'...
>>I think people'd rather their db just stopped accepting new transactions
>>rather than just losing data...
>
> Not being able to issue new transactions *is* data loss --- how are you
> going to get the system out of that state?
Not allowing any transactions except a vacuum...
> autovacuum is the correct long-term solution to this, not some kind of
> automatic hara-kiri.
Yeah, seems like it should really happen soon...
Chris