-
Notifications
You must be signed in to change notification settings - Fork 435
Description
Hi team,
Right now, asyncpg does prepared statements in two steps.
- Parse, Describe, Flush.
- Bind, Execute, Sync.
Two coroutines represent these two steps. In a highly concurrent setup, the wall-clock gap between these two steps can be large. Why is it causing a problem for me? Command Flush
does not change the connection state, i.e., active
. So I observe a lot of queries show up as
wait_event_type | wait_event | state
-----------------+------------+--------
Client | ClientRead | active
in pg_stat_activity
. This causes a lot of unnecessary confusion for our database monitors/observability tools. The connection is essentially idle between these two steps, so I think it is better to report it as wait_event=ClientRead
but state=idle
? Meanwhile, the command Sync
will change the connection state to idle
, so after step 2, the connection becomes idle.
My ask is: could we use Sync
instead of Flush
in step 1?
Moreover, there could be more than one bind
step, so the sequence could be
1. Parse, Describe, Flush.
2. Bind, Execute, Sync.
3. Bind, Execute, Sync.
...
In this case, the connection is active only between step 1 and step 2, but idle for all other step gaps. I kind of feel this behavior is inconsistent.
I might have neglected some basic design about PostgreSQL extended query. Hope to hear from you soon!