>> Hmm, that's pretty interesting ... are there any nearby autovacuums of >> pg_toast_2619?
> Several, both before and after the error message:
Well, it seems clear that somehow the vacuum deleted a toast tuple that the other statement was about to fetch, but it's not clear how this could be. The pg_statistic fetch must have come from the planner, which should always be called with a transaction snapshot established, and that ought to protect it against vacuum deleting anything that could be visible to SnapshotNow. Weird.
[ pokes around for a bit ... ] Hmm, can you say how the failing query was submitted, exactly? I'm wondering if it came in via simple Query (PQexec) or extended-query protocol (anything with parameters).
The command was sent from a python via django-1.2.1 using psycopg2-2.2.1