El 10/02/16 a las 21:46, Tom Lane escribió:
>
> Think you're outta luck on that. If we logged the duration before
> commit, it would be entirely misleading for short commands, because
> the time needed to commit wouldn't be included. So we log it after,
> when there's no longer any active transaction.
Any other way to log the DDL and DML statements (maybe delay writing the
log until the xid is available)?
Yes, I know the complexity of this, but I'm surprised this hasn't come
up in the list before (maybe it did and i missed the mail).
> We could maybe fix this by redefining %x as "the current or most recent
> xid", so that it'd still be valid for messages issued post-commit.
> But I'm afraid that would add about as many bad behaviors as it would
> remove. In your example above, that would result in a pretty misleading
> xid attached to the "begin" statement, since at that point we have
> started a new transaction but not assigned it any xid.
This really gives little use for recovery_target_xid. :(
Regards,
--
Martín Marqués http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services