Re: Show WAL write and fsync stats in pg_stat_io

On Mon, Dec 25, 2023 at 04:09:34PM +0300, Nazir Bilal Yavuz wrote:
> On Wed, 9 Aug 2023 at 21:52, Melanie Plageman <melanieplageman(at)gmail(dot)com> wrote:
>> If there is any combination of BackendType and IOContext which will
>> always read XLOG_BLCKSZ bytes, we could use XLOG_BLCKSZ for that row's
>> op_bytes. For other cases, we may have to consider using op_bytes 1 and
>> tracking reads and write IOOps in number of bytes (instead of number of
>> pages). I don't actually know if there is a clear separation by
>> BackendType for these different cases.
>
> Using op_bytes as 1 solves this problem but since it will be different
> from the rest of the pg_stat_io view it could be hard to understand.
> There is no clear separation by backends as it can be seen from the walsender.

I find the use of 1 in this context a bit confusing, because when
referring to a counter at N, then it can be understood as doing N
times a operation, but it would be much less than that. Another
solution would be to use NULL (as a synonym of "I don't know") and
then document that in this case all the bigint counters of pg_stat_io
track the number of bytes rather than the number of operations?

>> The other alternative I see is to use XLOG_BLCKSZ as the op_bytes and
>> treat op_bytes * number of reads as an approximation of the number of
>> bytes read. I don't actually know what makes more sense. I don't think I
>> would like having a number for bytes that is not accurate.
>
> Also, we have a similar problem in XLogPageRead() in xlogrecovery.c.
> pg_pread() call tries to read XLOG_BLCKSZ but it is not certain and we
> don't count IO if it couldn't read XLOG_BLCKSZ. IMO, this is not as
> important as the previous problem but it still is a problem.
>
> I would be glad to hear opinions on these problems.

Correctness matters a lot for monitoring, IMO.
--
Michael