Hi, Thanks for raising this. I did not look at the patch, but I have some high level comments.
> I would like to propose adding a last_executed timestamptz column to > pg_stat_statements. This column records when each tracked statement > was most recently executed. I do think there is value in adding a last_executed timestamp. I actually think last_executed should be the time the query started timestamp, so we should actually create an entry at ExecutorStart, along with calls_started and calls_completed. This is great for tracking cancelled queries. The issue is the extra overhead of tracking the query on EcecutorStart, but that should be less of an issue once we move pg_stat_statements to the cumulative statistics system, which will be possible once we get some prerequisite work to make this happen [0]. Another concern is the width of the current view. I think before we add any new attribute, pg_stat_statements fields should be split. This was discussed in [1]. > The motivation comes from real world experience with monitoring tools > like pgwatch that poll pg_stat_statements regularly. Currently, these > tools must fetch and store statistics for all statements, even those > that haven't executed recently. This creates significant storage > overhead. For a database with around 3400 statements polled every 3 > minutes, storing full query text requires roughly 2.5 MB per snapshot. > Over two weeks, this accumulates to about 17 GB. Even without query > text, storage reaches 10 GB. > > With a last_executed timestamptz, monitoring tools can simply filter > statements by "last_executed > NOW() - polling_interval" to fetch only > statements that have been executed since the last poll. This > eliminates the need for complex workarounds that some tools currently > use to identify changed statements > (https://github.com/cybertec-postgresql/pgwatch/blob/759df3a149cbbe973165547186068aa7b5332f9d/internal/metrics/metrics.yaml#L2605-L2766). Can pg_stat_statements.stats_since help here? for example "where stats_since > last_poll_timestamp" ? The client does have to track the last_poll_timestamp in that case. [0] https://www.postgresql.org/message-id/flat/caa5rz0s9sdou+z6veojchwk+kdetktatc-ky9fq9z6bjddu...@mail.gmail.com [1] https://www.postgresql.org/message-id/03f82e6f-66a3-4c4d-935c-ea4d93871dc1%40gmail.com -- Sami Imseih Amazon Web Services (AWS)
