On 2016-01-11 20:03:18 +0100, Andres Freund wrote:
> More generally, I'm doubtful that the approach of reading data from WAL
> as proposed here is a very good idea. It seems better to "just" dump the
> entire 2pc state into *one* file at checkpoint time.
Or better: After determining the checkpoint redo location, insert a WAL
record representing the entire 2PC state as of that moment. That way it
can easily restored during WAL replay and nothing special has to be done
on a standby. This way we'll need no extra wal flushes and fsyncs.