On Tue, Aug 17, 2021 at 04:47:44PM +0900, Michael Paquier wrote:
> One argument
> against this approach is that if pg_rewind fails in the middle of its
> operation then we would have done a set of fsync() for nothing, with
> the data folder still unusable.
I was skimming through the patch this morning, and that argument does
not hold much water as the flushes happen in the same place. Seems
like I got confused, sorry about that.
> I would be curious to see some
> numbers to see how much it matters with many physical files (say cases
> with thousands of small relations?).
For this one, one simple idea would be to create a lot of fake
relation files with a pre-determined size and check how things
change.
--
Michael