Geoff Winkless <pg@ukku.uk> writes:
> In our application we have a situation where once a day one process
> CREATEs an UNLOGGED table and INSERTs several hundred records using
> individual queries (no explicit transactions) all of which return
> successfully. We then send the ID of the table that we have created
> over a TCP socket to a second process, which runs a query that JOINs
> against that new table.
> Unfortunately quite often the second process is getting a
> PGRES_FATAL_ERROR with
> Primary: relation "qreftmp750" does not exist
> Now (and this is very important) this appears to be a race condition,
> because when that process immediately retries the same query (which we
> do when we get FATAL_ERROR) it sometimes works on the second or third
> (or even 11th) attempt.
Perhaps the second process is querying a standby server rather than
the primary? Replication lag could explain this. I don't really
believe that it's possible within a single PG server, though.
We take very substantial pains to avoid the race condition you're
positing.
If there is anything in your client software stack capable of issuing
implicit begin/commit, that'd deserve a second/third/fourth look...
regards, tom lane
