From: Tom Lane [mailto:t...@sss.pgh.pa.us] > So the correct fix is to improve autovacuum's check to discover whether > an old temp table is orphaned, so that it isn't fooled by putative owner > processes that are connected to some other DB. Step 2 of the proposed patch > tries to do that, but it's only half right: we need a change near line 2264 > not only line 2090. (Evidently, we need either a comment that the logic > is repeated, or else refactor so that there's only one copy.) > > Now, you can argue that autovacuum's check can be fooled by an "owner" > backend that is connected to the current DB but hasn't actually taken > possession of its assigned temp schema (and hence the table in question > really is orphaned after all). This edge case could be taken care of by > having backends clear their temp schema immediately, as in step 1 of the > patch. But I still think that that is an expensive way to catch what would > be a really infrequent case. Perhaps a better idea is to have backends > advertise in PGPROC whether they have taken possession of their assigned > temp schema, and then autovacuum could ignore putative owners that aren't > showing that flag. Or we could just do nothing about that, on the grounds > that nobody has shown the case to be worth worrying about. > Temp schemas that are sufficiently low-numbered to be likely to have an > apparent owner are also likely to get cleaned out by actual temp table > creation reasonably often, so I'm not at all convinced that we need a > mechanism that covers this case. We do need something for the cross-DB > case though, because even a very low-numbered temp schema can be problematic > if it's in a seldom-used DB, which seems to be the case that was complained > of originally. > > On the whole, my vote is to fix and apply step 2, and leave it at that.
Done. It seems to work well. Regards Takayuki Tsunakawa
reset_temp_schema_v2.patch
Description: reset_temp_schema_v2.patch
