On 12/06/2019 02:24, Andres Freund wrote:
But anyway, to go forward, I think we should replace 'lat' with a
'txtime' (or similar) column that is not affected by -R. And then, under
-R only, add a new 'txlat' column, that shows the 'current' meaning of
lat under -R. Not convinced the names are right, but you get the gist.
I'm OK with that.
For testing the server under full load, like during that catch up period,
testing without -R seems better.
One area in which postgres is pretty weak, although less bad than we
used to be, is in is predicatable latency. Most production servers
aren't run under the highest possible throughput, therefore optimizing
jitter under loaded but not breakneck speeds is important.
And to be able to localize where such latency is introduced, it's
important to see the precise moment things got slower / where
performance recovered.
I agree with all that. I'm still not convinced the changes you're
proposing will help much, but if you would find it useful, I can't argue
with that.
- Heikki
