On 2024/5/16 00:58, Robert Haas wrote:
On Tue, Apr 16, 2024 at 3:16 AM Quan Zongliang <quanzongli...@yeah.net> wrote:
According to the discussion with Jian He. Use the guc hook to check if
the xid needs to be output. If needed, the statement log can be delayed
to be output.
I appreciate the work that both of you have put into this, but I think
we should reject this patch and remove the TODO item. We currently
have some facilities (like log_statement) that log the statement
before parsing it or executing it, and others (like
log_min_duration_statement) that log it afterward. That is probably
not documented as clearly as it should be, but it's long-established
behavior.
What this patch does is change the behavior of log_statement so that
log_statement sometimes logs the statement before it's executed, and
sometimes after the statement. I think that's going to be confusing
and unhelpful. In particular, right now you can assume that if you set
log_statement=all and there's a statement running, it's already been
logged. With this change, that would sometimes be true and sometimes
false.
For example, suppose that at 9am sharp, I run an UPDATE command that
takes ten seconds to complete. Right now, the log_statement message
will appear at 9am. With this change, it will run at 9am if I do it
inside a transaction block that has an XID already, and at 9:00:10am
if I do it in a transaction block that does not yet have an XID, or if
I do it outside of a transaction. I don't think the benefit of getting
the XID in the log message is nearly enough to justify such a strange
behavior.
I thought about writing statement log when xid assigned. But it's seemed
too complicated.
I'm inclined to keep it for a while. Until we find a good way or give
up. It's a reasonable request, after all.
