Here's an updated patch. I added some warning messages to autovacuum.
One thing I learned trying to debug this situation in production is
that it's nigh impossible to find the pid of the session using a
temporary schema. The number in the schema refers to the backendId in
the sinval stuff for which there's no external way to look up the
corresponding pid. It would have been very helpful if autovacuum had
just told me which backend pid to kill.
I still have the regression test in the patch and as before I think
it's probably not worth committing. I'm leaning to reverting that
section of the patch before comitting.
Incidentally there's still a hole here where a new session can attach
to an existing temporary schema where a table was never truncated due
to a session dieing abnormally. That new session could be a long-lived
session but never use the temporary schema causing the old table to
just sit there. Autovacuum has no way to tell it's not actually in
use. I tend to think the optimization to defer cleaning the temporary
schema until it's used might not really be an optimization. It still
needs to be cleaned someday so it's just moving the work around. Just
removing that optimization might be the easiest way to close this
hole. The only alternative I see is adding a flag to PROC or somewhere
where autovacuum can see if the backend has actually initialized the
temporary schema yet or not.
