Re: Monitoring time of fsyncing WALs

On 1 July 2018 at 11:29, Michael Paquier <michael@paquier.xyz> wrote:


>
> So at the end, I would like to use the proposed patch and call it a
> day.  Thoughts?
>
>
Patch looks good.

I'll hijack the thread to add a few more perf/dtrace tracepoints in the WAL
code, as they were also missing. Proposed rider patch attached.

I've updated
https://wiki.postgresql.org/wiki/Profiling_with_perf#PostgreSQL_pre-defined_tracepoint_events
to document how to add tracepoints.

It's not that hard to trace duration of a given function call with dtrace
without such annotations, but they make it much easier to discover where in
a large codebase to look, providing a form of documentation. With perf they
make life much easier. I should add some more to make it easier to analyse
relation extension contention on indexes and the heap, instrument btree
index ops like page splits, instrument heavyweight locking (beyond
LOCK_WAIT_START), etc. They're also handy for gdb -
https://sourceware.org/gdb/onlinedocs/gdb/Static-Probe-Points.html - also
handy for gdb's "info probes".

BTW, we might want to instrument the pgstat_ counter calls
and pgstat_report_wait_start / pgstat_report_wait_end, but it's easy enough
to use dynamic tracepoints for those so I haven't yet. Maybe even just
document them as points of interest.

-- 
 Craig Ringer                   http://www.2ndQuadrant.com/
 PostgreSQL Development, 24x7 Support, Training & Services