On 09/08/2018 19:57, hubert depesz lubaczewski wrote:
I just noticed that when I called a procedure that commits and rollbacks - the xact_start in pg_stat_activity is not updated. Is it intentional?
It's an artifact of the way this is computed. The reported transaction timestamp is the timestamp of the first top-level statement of the transaction. This assumes that transactions contain statements, not the other way around, like it is now possible. I'm not sure what an appropriate improvement would be here.
That would just mean that query_start would be older than xact_start, but that's okay because the displayed query would be a CALL so we'll know what's going on.
Note that this issue already exists for other commands that start transactions internally, such as VACUUM and CREATE INDEX CONCURRENTLY. At the moment, one should interpret xact_start as referring to the top-level transaction only.
In practice, I think the value of xact_start versus query_start is to anayze idle transactions. This doesn't happen with internal transactions, so it's not a big deal in practice.
This was added as an open item by Michael[1]. When the RMT discussed,
we were able to make arguments both ways (i.e. adjusting the behavior vs.
not).
Peter, from your analysis it sounds like we should leave it, but I wanted to
