Simon Riggs wrote:
> If you want to do things a different way you need to say what you want
> to do and what effects those changes will have.
I want to reduce the coupling between the primary and the master. The
less they need to communicate, the better. I want to get rid of slotid,
and as many of the other extra information carried in WAL records that I
can. I believe that will make the patch both simpler and more robust.
> Are there tradeoffs? If so what are they?
I don't think there's any big difference in user-visible behavior.
RecordKnownAssignedTransactionId now needs to be called for every xlog
record as opposed to just the first record where an xid appears, because
I eliminated the hint flag in WAL to mark those records. And it needs to
look up the recover proc by xid, instead of using the slot id. But I
don't think that will have a significant impact on performance.
-- Heikki Linnakangas EnterpriseDB http://www.enterprisedb.com
