On Mon, Mar 3, 2008 at 8:46 AM, Heikki Linnakangas
<heikki@enterprisedb.com> wrote:
> Tom Lane wrote:
> > Do you want to write up a flag-based patch, or shall I?
>
> I can do that.
BTW, I found a easier way of reproducing this (see attached 2pc.sql).
It might help with debugging or verifying a fix/regression.
Thanks for the attention so far, but I have some questions about the issue:
[1] The data file is reported missing in the second transaction only
if the first transaction was ended using PREPARE TRANSACTION. The
error does not show up if a direct COMMIT is performed (commit.sql)
instead of PREPARE TRANSACTION + COMMIT PREPARED. Why is that so?
[2] From all of the discussion here since my first post, I understand
that there are complications for session-level TEMP tables. But is it
easier to support PREPARE TRANSACTION for transactions that create and
drop their TEMP tables, i.e., so that the tables are not session-level
but just transaction-level?
[3] I am not certain how widespread they might be, but I think there
may be some backward compatibility concerns with the patch you are
proposing. On the one hand, the documentation says, "It is not
currently allowed to PREPARE a transaction that has executed any
operations involving temporary tables or created any cursors WITH
HOLD." But temporary tables that are created ON COMMIT DROP are more
like cursors that do not have WITH HOLD specified. So it does not seem
clear from the documentation that PREPARE TRANSACTION is not
supported, and indeed due to the lack of a check in Postgres today, it
seems as though it is supported. Do you think there is a risk in
breaking applications?
Thanks.
- John