> So yeah, I think this is due to confusion with two snapshots and > failing > to consider both of them when calculating TransactionXmin. > > But I think removing one of the snapshots (as the v2 does it) is rather > strange too. I very much doubt having both the transaction and active > snapshots in the parallel worker is not intentional, and Pavel may very > well be right this breaks isolation levels that use the xact snapshot > (i.e. REPEATABLE READ and SERIALIZABLE). I haven't checked, though. > > So I think we need to keep both snapshots, but fix TransactionXmin to > consider both of them - I suppose ParallelWorkerMain could simply look > at the two snapshots, and use the minimum. Which is what [1] (already > linked by Pavel) talks about, although that's related to concerns about > one of the processes dying, which is not what's happening here. > > > I'm wondering what consequences this may have on production systems, > though. We've only seen this failing because of the assert, so what > happens when the build has asserts disabled?
> Looking at SubTransGetTopmostTransaction, it seems the while loop ends > immediately (because it's pretty much the opposite of the assert), so > we > just return the "xid" as topmost XID. But will that produce incorrect > query results, or what?
I haven't seen anything incorrect on production systems with asserts turned off, by I should note this assertion is not immediately reproduced so it is not that easy to catch the possible logical inconsequences of parallel scan results. As I've noticed that subxids cache is used in most cases and even this assertion is also rare in parallel scans. Maybe that is why we lived with this bug unnoticed so long.