> > It is intuitive. The bug was iirc, that you saw 2 versions of the same row
> > in the second select statement (= 2 rows returned by second select).
>
> I think we should be extremely wary of assuming that we have a clear
> characterization of "what the bug is", let alone "how to fix it".
> The real issue here is that SELECT has different MVCC visibility rules
> from UPDATE and SELECT FOR UPDATE. I suspect that that *must* be so
> in any mode that allows more concurrency than full serializable mode.
Yes, definitely.
> Thus, the question we are really facing is how we might alter the
> visibility rules in a way that will make the results more intuitive
> and/or useful while still allowing concurrency.
>
> This will take thought, research and discussion. A quick fix is the
> last thing that should be on our minds.
From my latest tests( see following post), I tend to agree, that this is
extremely sensitive :-(
I do however think that Vadim's patch description was the correct thing to do.
The problem case seems to be when the function is not executed inside a txn.
I was not able to reproduce any failure, when inside txns, since the first update
or select for update blocks the rest.
Andreas