Hi,
On 2017-07-18 21:49:27 +0900, Yugo Nagata wrote:
> After a DML is perfomed, the statistics is sent to pgstat by pgsent_report_sent().
> However, the mininum time between stas file update is PGSTAT_STAT_INTERVAL, so if
> a few DMLs are performed with short interval, some statistics could not be sent
> until the backend is shutdown.
>
> This is not a problem in usual cases, but in case that a session is remained in
> idle for a long time, for example when using connection pooling, statistics of
> a huge change of a table is not sent for a long time, and as a result, starting
> autovacuum might be delayed.
Hm.
> An idea to resolve this is call pgsent_report_sent() again with a delay
> after the last DML to make sure to send the statistics. The attached patch
> implements this.
That seems like it'd add a good number of new wakeups, or at least
scheduling of wakeups. Now a timer would be armed whenever a query is
run outside of a transaction - something happening quite regularly. And
it'd do so even if a) there are no outstanding stat reports b) there's
already a timer running. Additionally it'd, unless I mis-skimmed,
trigger pgstat_report_stat() from within
CHECK_FOR_INTERRUPTS/ProcessInterrupts even when inside a transaction,
if that transaction is started after the enable_timeout_after().
I'm not entirely sure what a good solution would be. One option would
be to have a regular wakeup setting a flag (considerably longer than
500ms, that's way too many additional wakeups), rearm the timer in the
handler, but only process the flag when outside of a transaction.
Or we could do nothing - I actually think that's a viable option.
- Andres
