Hi, On 2021-05-16 16:23:02 -0400, Tom Lane wrote: > And the reason oldestXID contains that is that pg_upgrade applied > pg_resetwal, which does this: > > /* > * For the moment, just set oldestXid to a value that will force > * immediate autovacuum-for-wraparound. It's not clear whether adding > * user control of this is useful, so let's just do something that's > * reasonably safe. The magic constant here corresponds to the > * maximum allowed value of autovacuum_freeze_max_age. > */ > ControlFile.checkPointCopy.oldestXid = set_xid - 2000000000; > if (ControlFile.checkPointCopy.oldestXid < FirstNormalTransactionId) > ControlFile.checkPointCopy.oldestXid += FirstNormalTransactionId;
Yea - this is causing quite a few problems... See https://www.postgresql.org/message-id/20210423234256.hwopuftipdmp3okf%40alap3.anarazel.de > So it seems like we should do some combination of these things: > > 1. Fix FullXidRelativeTo to be a little less trusting. It'd > probably be sane to make it return FirstNormalTransactionId > when it'd otherwise produce a wrapped-around FullXid, but is > there any situation where we'd want it to throw an error instead? I'm wondering whether we should *always* make it an error, and fix the places where that causes problems. > 2. Change pg_resetwal to not do the above. It's not entirely > apparent to me what business it has trying to force > autovacuum-for-wraparound anyway, but if it does need to do that, > can we devise a less klugy method? Yes, see the above email. I think we really to transport accurate oldest xid + epoch for pg_upgrade. > It also seems like some assertions in procarray.c would be a > good idea. With the attached patch, we get through core > regression just fine, but the pg_upgrade test fails immediately > after the "Resetting WAL archives" step. Agreed. Greetings, Andres Freund
