Ron Snyder <snyder@roguewave.com> writes:
> We've been getting errors similar to the following (the specific large
> object that is "missing" is different every time) during our nightly
> pg_dump:
>
> pg_dump: dumpBlobs(): could not open large object: ERROR:
> inv_open: large object 48217896 not found
> After doing a bunch of testing and experimenting, we're pretty sure that the
> problem we were having is due to the large objects being deleted while the
> pg_dump was running.
Sounds plausible. I proposed years ago that we ought to fix large
objects to be MVCC-compliant:
http://archives.postgresql.org/pgsql-hackers/2002-05/msg00875.php
but the issue seems to have fallen through the cracks. (It was
hard to fix at the time because there wasn't any easy way for the
LO functions to lay their hands on a suitable snapshot, but as of
8.0 I think ActiveSnapshot would work.)
> Is this a problem that is handled differently in PostgreSQL 8?
Nope. I'm feeling a strong urge to go fix it for 8.1 though.
The question from the previous mail still stands: would anybody's
applications be broken if we change the MVCC behavior of large objects?
regards, tom lane