Re: New statistics for tuning WAL buffer size

On 2020/10/01 10:50, tsunakawa(dot)takay(at)fujitsu(dot)com wrote:
> From: Kyotaro Horiguchi <horikyota(dot)ntt(at)gmail(dot)com>
>> Another reason that I mildly want to object to subdivided functions is
>> I was annoyed that a stats view makes many individual calls to
>> functions that internally share the same statistics entry. That
>> behavior required me to provide an entry-caching feature to my
>> shared-memory statistics patch.
>
> +1
> The views for troubleshooting performance problems should be as light as possible. IIRC, we saw frequently searching pg_stat_replication consume unexpectedly high CPU power, because it calls pg_stat_get_activity(null) to get all sessions and join them with the walsenders. At that time, we had hundreds of client sessions. We expected pg_stat_replication to be very lightweight because it provides information about a few walsenders.

I think that we can improve that, for example, by storing backend id
into WalSndCtl and making pg_stat_get_wal_senders() directly
get the walsender's LocalPgBackendStatus with the backend id,
rather than joining pg_stat_get_activity() and pg_stat_get_wal_senders().

Regards,

--
Fujii Masao
Advanced Computing Technology Center
Research and Development Headquarters
NTT DATA CORPORATION