Re: Add statistics to pg_stat_wal view for wal related parameter tuning

From: Masahiro Ikeda <ikedamsh(at)oss(dot)nttdata(dot)com>
To: Kyotaro Horiguchi <horikyota(dot)ntt(at)gmail(dot)com>
Cc: amit(dot)kapila16(at)gmail(dot)com, pgsql-hackers(at)postgresql(dot)org
Subject: Re: Add statistics to pg_stat_wal view for wal related parameter tuning
Date: 2020-10-22 01:44:53
Message-ID: 729d83f87da2cfbca4343e577f2e20a1@oss.nttdata.com
Views: Raw Message | Whole Thread | Download mbox | Resend email
Thread:
Lists: pgsql-hackers

On 2020-10-21 18:03, Kyotaro Horiguchi wrote:
> At Tue, 20 Oct 2020 16:11:29 +0900, Masahiro Ikeda
> <ikedamsh(at)oss(dot)nttdata(dot)com> wrote in
>> On 2020-10-20 12:46, Amit Kapila wrote:
>> > I see that we also need to add extra code to capture these stats (some
>> > of which is in performance-critical path especially in
>> > XLogInsertRecord) which again makes me a bit uncomfortable. It might
>> > be that it is all fine as it is very important to collect these stats
>> > at cluster-level in spite that the same information can be gathered at
>> > statement-level to help customers but I don't see a very strong case
>> > for that in your proposal.
>
> We should avoid that duplication as possible even if the both number
> are important.
>
>> Also about performance, I thought there are few impacts because it
>> increments stats in memory. If I can implement to reuse pgWalUsage's
>> value which already collects these stats, there is no impact in
>> XLogInsertRecord.
>> For example, how about pg_stat_wal() calculates the accumulated
>> value of wal_records, wal_fpi, and wal_bytes to use pgWalUsage's
>> value?
>
> I don't think that works, but it would work that pgstat_send_wal()
> takes the difference of that values between two successive calls.
>
> WalUsage prevWalUsage;
> ...
> pgstat_send_wal()
> {
> ..
> /* fill in some values using pgWalUsage */
> WalStats.m_wal_bytes = pgWalUsage.wal_bytes -
> prevWalUsage.wal_bytes;
> WalStats.m_wal_records = pgWalUsage.wal_records -
> prevWalUsage.wal_records;
> WalStats.m_wal_wal_fpi = pgWalUsage.wal_fpi -
> prevWalUsage.wal_fpi;
> ...
> pgstat_send(&WalStats, sizeof(WalStats));
>
> /* remember the current numbers */
> prevWalUsage = pgWalUsage;

Thanks for your advice. This code can avoid the performance impact of
critical code.

By the way, what do you think to add these statistics to the pg_stat_wal
view?
I thought to remove the above statistics because there is advice that
PG13's features,
for example, pg_stat_statement view, vacuum log, and so on can cover
use-cases.

regards,
--
Masahiro Ikeda
NTT DATA CORPORATION

In response to

Responses

Browse pgsql-hackers by date

  From Date Subject
Next Message tsunakawa.takay@fujitsu.com 2020-10-22 01:53:07 RE: [Patch] Optimize dropping of relation buffers using dlist
Previous Message Michael Paquier 2020-10-22 01:41:53 Re: Online verification of checksums