On Tue, Aug 10, 2021 at 03:15:27PM +0900, Masahiko Sawada wrote:
> > This shows that the Hash wrote 1100 temp buffers, but I/O timings didn't increase at all.
> Yes, we don't collect the time spent reading and writing temp files.
And this is the source of my concern.
Not that we have this documented in one way or another, but why exactly
aren't we taking temp buffers into timing? There are cases where it is
the temp buffers IO that is using most of the time.
Best regards,
depesz
