Alvaro Herrera <alvherre@commandprompt.com> writes:
> Tom Lane wrote:
>> While I was looking at this I wondered whether
>> RelationCacheInitializePhase2 really needs to be inside the startup
>> transaction at all. I think it could probably be moved up before
>> that. However, if the AV launcher has to do GetTransactionSnapshot
>> then it's not clear that improves matters anyway.
> Well, the difference is that the initial transaction would be a few
> microsec shorter ... not sure if that matters.
Actually, there is a better way to do this: if we move up the
RelationCacheInitializePhase2 call, then we can have the AV launcher
case just fall out *before* the transaction start. It can do
GetTransactionSnapshot inside its own transaction that reads
pg_database. This is a better solution because it'll have a more
up-to-date version of RecentGlobalXmin while scanning pg_database.
(Indeed, I think this might be *necessary* over the very long haul.)
I think I've got the signal handling cleaned up, but need to test.
Is there any really good reason why autovacuum has its own avl_quickdie
instead of using quickdie() for SIGQUIT?
regards, tom lane