Hi,
On 2022-08-31 19:52:49 +0200, Matthias van de Meent wrote:
> As for having a lower granularity and preventing the
> one-syscall-per-Relation issue, can't we reuse the query_start or
> state_change timestamps that appear in pg_stat_activity (potentially
> updated immediately before this stat flush), or some other per-backend
> timestamp that is already maintained and considered accurate enough
> for this use?
The problem is that it won't change at all for a query that runs for a week -
and we'll report the timestamp from a week ago when it finally ends.
But given this is done when stats are flushed, which only happens after the
transaction ended, we can just use GetCurrentTransactionStopTimestamp() - if
we got to flushing the transaction stats we'll already have computed that.
> tabentry->numscans += lstats->t_counts.t_numscans;
> + if (pgstat_track_scans && lstats->t_counts.t_numscans)
> + tabentry->lastscan = GetCurrentTimestamp();
Besides replacing GetCurrentTimestamp() with
GetCurrentTransactionStopTimestamp(), this should then also check if
tabentry->lastscan is already newer than the new timestamp.
Greetings,
Andres Freund