Re: New statistics for tuning WAL buffer size - Mailing list pgsql-hackers

On 2020/09/30 20:21, Amit Kapila wrote:
> On Tue, Sep 29, 2020 at 9:23 PM Fujii Masao <masao.fujii@oss.nttdata.com> wrote:
>>
>> On 2020/09/29 11:51, Masahiro Ikeda wrote:
>>> On 2020-09-29 11:43, Amit Kapila wrote:
>>>> On Tue, Sep 29, 2020 at 7:39 AM Masahiro Ikeda <ikedamsh@oss.nttdata.com> wrote:
>>>>>
>>>>> On 2020-09-28 12:43, Amit Kapila wrote:
>>>>>> On Mon, Sep 28, 2020 at 8:24 AM Kyotaro Horiguchi
>>>>>> <horikyota.ntt@gmail.com> wrote:
>>>>>>>
>>>>>>> At Mon, 28 Sep 2020 08:11:23 +0530, Amit Kapila
>>>>>>> <amit.kapila16@gmail.com> wrote in
>>>>>>>> One other thing that occurred to me today is can't we keep this as
>>>>>>>> part of PgStat_GlobalStats? We can use pg_stat_reset_shared('wal'); to
>>>>>>>> reset it. It seems to me this is a cluster-wide stats and somewhat
>>>>>>>> similar to some of the other stats we maintain there.
>>>>>>>
>>>>>>> I like that direction, but PgStat_GlobalStats is actually
>>>>>>> PgStat_BgWriterStats and cleard by a RESET_BGWRITER message.
>>>>>>>
>>>>>>
>>>>>> Yeah, I think if we want to pursue this direction then we probably
>>>>>> need to have a separate message to set/reset WAL-related stuff. I
>>>>>> guess we probably need to have a separate reset timestamp for WAL. I
>>>>>> think the difference would be that we can have one structure to refer
>>>>>> to global_stats instead of referring to multiple structures and we
>>>>>> don't need to issue separate read/write calls but OTOH I don't see
>>>>>> many disadvantages of the current approach as well.
>>>>>
>>>>> IIUC, if we keep wal stats as part of PgStat_GlobalStats,
>>>>> don't we need to add PgStat_ArchiverStats and PgStat_SLRUStats
>>>>> to PgStat_GlobalStats too?
>>>>>
>>>>
>>>> I have given the idea for wal_stats because there is just one counter
>>>> in that. I think you can just try to evaluate the merits of each
>>>> approach and choose whichever you feel is good. This is just a
>>>> suggestion, if you don't like it feel free to proceed with the current
>>>> approach.
>>>
>>> Thanks for your suggestion.
>>> I understood that the point is that WAL-related stats have just one counter now.
>>>
>>> Since we may add some WAL-related stats like pgWalUsage.(bytes, records, fpi),
>>> I think that the current approach is good.
>>
>> +1
>>
> 
> Okay, it makes sense to keep it in the current form if we have a plan
> to extend this view with additional stats. However, why don't we
> expose it with a function similar to pg_stat_get_archiver() instead of
> providing individual functions like pg_stat_get_wal_buffers_full() and
> pg_stat_get_wal_stat_reset_time?

We can adopt either of those approaches for pg_stat_wal. I think that
the former is a bit more flexible because we can collect only one of
WAL information even when pg_stat_wal will contain many information
in the future, by using the function. But you thought there are some
reasons that the latter is better for pg_stat_wal?

Regards,

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