Re: Show WAL write and fsync stats in pg_stat_io

On Wed, Sep 20, 2023 at 1:28 PM Nazir Bilal Yavuz <byavuz81(at)gmail(dot)com> wrote:
>
> Hi,
>
> Thanks for the review!
>
> Current status of the patch is:
> - IOOBJECT_WAL / IOCONTEXT_NORMAL read, write and fsync stats are added.
> - IOOBJECT_WAL / IOCONTEXT_NORMAL write and fsync tests are added.
> - IOOBJECT_WAL / IOCONTEXT_INIT stats are added.
> - pg_stat_io shows different op_bytes for the IOOBJECT_WAL operations.
> - Working on which 'BackendType / IOContext / IOOp' should be banned in pg_stat_io.
> - PendingWalStats.wal_sync and PendingWalStats.wal_write_time / PendingWalStats.wal_sync_time are moved to pgstat_count_io_op_n() / pgstat_count_io_op_time() respectively.
>
> TODOs:
> - Documentation.
> - Try to set op_bytes for BackendType / IOContext.
> - Decide which 'BackendType / IOContext / IOOp' should not be tracked.
> - Add IOOBJECT_WAL / IOCONTEXT_NORMAL read tests.
> - Add IOOBJECT_WAL / IOCONTEXT_INIT tests.

This patchset currently covers:
- IOOBJECT_WAL / IOCONTEXT_NORMAL read, write and fsync.
- IOOBJECT_WAL / IOCONTEXT_INIT write and fsync.

doesn't cover:
- Streaming replication WAL IO.

Is there any plan to account for WAL read stats in the WALRead()
function which will cover walsenders i.e. WAL read by logical and
streaming replication, WAL read by pg_walinspect and so on? I see the
patch already covers WAL read stats by recovery in XLogPageRead(), but
not other page_read callbacks which will end up in WALRead()
eventually. If added, the feature at
https://www.postgresql.org/message-id/CALj2ACXKKK%3DwbiG5_t6dGao5GoecMwRkhr7GjVBM_jg54%2BNa%3DQ%40mail.gmail.com
can then extend it to cover WAL read from WAL buffer stats.

--
Bharath Rupireddy
PostgreSQL Contributors Team
RDS Open Source Databases
Amazon Web Services: https://aws.amazon.com