"Florian G. Pflug" <fgp@xxxxxxxx> writes: > Tom Lane wrote: >> Until when? How would you synchronize the switchover? > Every snapshot would either contain the old, or the new version of > the corresponding pg_class tuple. The ones using the old version > couldn't possible be writer, only reader (TRUNCATE would still need > to acquire a lock that ensures that). New transactions started after > the commit of the truncate would see the new version, and use > the new datafile. Wrong. *All* transactions read the system catalogs with SnapshotNow. There's been some discussion of using MVCC semantics on catalog accesses, but we are very far from being ready to do that; the implications will reach into every part of the backend. To give just one example of why this isn't necessarily a hot idea, suppose that transaction A adds a constraint to a table T and commits, and then transaction B (which started before A and hasn't previously touched T) comes along and inserts a tuple into T that violates the constraint. If B uses an old snapshot for looking up T's catalog info then it will not see the constraint that it must satisfy. Locking does not help since A released its locks on T before B would have tried to take any lock. regards, tom lane
