Claudio Freire <klaussfreire@gmail.com> wrote:
> On the other hand, in order to benefit from synchro scans and
> stuff like that, one has to increase concurrency beyond what is
> normally considered optimal.
That's a good example of why any general configuration advice should
just be used as a starting point. There's no substitute for
benchmarking your own real workload on your own hardware.
On the third hand, though, you have to be very careful about
interpreting these results -- if you used a configuration with a
small effective_cache_size so you could get a lot of benefit from
the synchro scans, you might have suppressed choice of an index
which would have allowed them to run faster with lower concurrency.
Or you might have had to cut your work_mem to a small enough size
(to avoid OOM errors) to force a totally different plan. So to get
a meaningful comparison, you have to change a number of variables at
once.
Good benchmarking is really hard.
-Kevin