On Fri, Jun 11, 2004 at 12:17:57PM -0400, Greg Stark wrote:
> Steve Atkins <steve@blighty.com> writes:
>
> > So, if you take a local snapshot of the global at the beginning of
> > your transaction then the visible changes at any point are those from
> > transactions that commited before your transaction started. That's
> > well-defined, at least, and appears to be pretty much the same as the
> > standard read commited isolation level.
>
> no, read committed would see any other updates that have been committed since
> the start of your transaction.
Uhm... only updates within the current transaction. So if you merge the
global state and the local state that's exactly what you'll see.
> For some linear aggregates you could start with the initcond, apply all the
> local updates and whenever you have to read the actual value then use the
> global variable at that time. But not all aggregates can be handled that way.
> I think all the standard ones could be though, sum(), count(), stddev(), etc.
I think all the standard ones can (anything with an associative update
function, if I remember my math correctly). And my thought was not
that this would be a neato transparent optimization that the parser
would use directly in all cases, rather that it would be a hack
explicitly setup by the DBA for those specific cases where they need
it.
Cheers, Steve