On Thu, 2003-10-16 at 11:03, Magnus Sundberg wrote:
> Chris Mason wrote:
> > On Thu, 2003-10-16 at 10:15, Matthew T. O'Connor wrote:
> >
> >>On Thu, 2003-10-16 at 09:57, Chris Mason wrote:
> >>
> >>>Ok, here's a patch for dbmail 1.2 cvs that implements the transaction
> >>>begin/commit for mysql during pipe.c:insert_messages, and the imap
> >>>copy/append message commands.
> >>
> >>Great, dbmail should be using transactions, why use a database if you
> >>don't use transactions.
> >>
> >
>
> I beleive this is a non issue, suppose the following function
>
> start_transaction(){
> if ( transactions_supported) {
> /* Issue SQL start transaction code */
>
> };
> };
There's a lot more to using transactions than BEGIN/COMMIT. The current
code orders operations such that transactions are not purely required.
When inserting messages it sets the unique_id last (as a kind of
informal commit), does rollback manually when some part of the insert
fails, etc.
A real transactional system would let the database do all of that. My
patch only addresses a performance aspect with mysql transactional
tables...every write triggers a commit. Enclosing them in a transaction
does one commit for the whole group instead.
Inserting a huge message as an atomic unit might not be the best idea,
I'm not sure how well mysql will cope with it.
-chris
