Prepared Statement connection is active and wait for ClientRead.

Hi team,

Right now, asyncpg does prepared statements in two steps.

1. Parse, Describe, Flush.
2. 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!

从 issue 中描述的 asyncpg prepared-statement 流程以及 PostgreSQL 的 extended-query 协议文档开始。确定更改 Flush/Sync 的时序是否能在包含多个 Bind/Execute 步骤的情况下保持正确性；完成的标准是 pg_stat_activity 一致地报告连接状态，同时不破坏 prepared statements。