On Mon, Oct 14, 2019 at 08:18:47PM +0200, Tomas Vondra wrote:
> Note: FWIW, a single snapshot of pg_stats* may be misleading, because
> it's cumulative, so it's not clear how accurately it reflects current
> state. Next time take two snapshots and subtract them.
For bonus points, capture it with timestamp and make RRD graphs.
I took me awhile to get around to following this advice, but now I have 12+
months of history at 5 minute granularity across all our customers, and I've
used my own implementation to track down inefficient queries being run
periodically from cron, and notice other radical changes in writes/reads
I recall seeing that the pgCluu project does this.
http://pgcluu.darold.net/
Justin