On Wed, Nov 8, 2017 at 8:46 AM, Robert Haas wrote:
> On Tue, Nov 7, 2017 at 4:31 AM, Haribabu Kommi
> wrote:
> >> Updated patch attached.
> > Patch rebased.
>
> I think the earlier concerns about the performance impact of this are
> probably very valid concerns, and I don't see how the new version of
> the patch gets us much closer to solving them.
>
I will check the performance with the changes of removing the stats
collector
usage and provide the details.
> I am also not sure I understand how the backend_write_blocks column is
> intended to work. The only call to pgstat_send_walwrites() is in
> WalWriterMain, so where do the other backends report anything?
>
With the current patch, All the backends update the stats in shared memory
structure and only WAL writer process gathers the stats and share with the
stats collector.
> Also, if there's only ever one global set of counters (as opposed to
> one per table, say) then why use the stats collector machinery for
> this at all, vs. having a structure in shared memory that can be
> updated directly? It seems like adding a lot of overhead for no
> functional benefit.
>
Yes, I agree that using stats collector for these stats is an overhead.
I change the patch to use just the shared memory structure and
gather the performance results and post it to the next commitfest.
Currently I marked the patch as "returned with feedback" in the
ongoing commitfest.
Regards,
Hari Babu
Fujitsu Australia