On Sat, Dec 24, 2016 at 7:46 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Stephen Frost <sfrost@snowman.net> writes:
>> * Tom Lane (tgl@sss.pgh.pa.us) wrote:
>>> The difficulty with that is it'd require a gettimeofday() call for
>>> every wait start. Even on platforms where those are relatively cheap,
>>> the overhead would be nasty --- and on some platforms, it'd be
>>> astonishingly bad. We sweated quite a lot to get the overhead of
>>> pg_stat_activity wait monitoring down to the point where it would be
>>> tolerable for non-heavyweight locks, but I'm afraid this would push
>>> it back into the not-tolerable range.
>
>> Could we handle this like log_lock_waits..?
>
> Well, that only applies to heavyweight locks, which do a gettimeofday
> anyway in order to schedule the deadlock-check timeout. If you were
> willing to populate this new column only for heavyweight locks, maybe it
> could be done for minimal overhead. But that would be backsliding
> quite a lot compared to what we just did to extend pg_stat_activity's
> coverage of lock types.
>
Can we think of introducing new guc trace_system_waits or something
like that which will indicate that the sessions will report the value
of wait_start in pg_stat_activity? The default value of such a
parameter can be false which means wait_start will be shown as NULL in
pg_stat_activity and when it is enabled the wait_start can show the
time as proposed in this thread.
--
With Regards,
Amit Kapila.
EnterpriseDB: http://www.enterprisedb.com