On 7/2/15 9:15 AM, Etsuro Fujita wrote:
> While working on the foreign-join-pushdown issue, I noticed that in READ
> COMMITTED isolation level it's possible that the result of SELECT ...
> ORDER BY ... FOR UPDATE is not sorted correctly due to concurrent
> updates that replaced the sort key columns with new values as shown in
> the below example. That seems odd to me. So, I'd like to propose
> raising an error rather than returning a possibly-incorrect result for
> cases where the sorted tuples to be locked were modified by concurrent
> updates.
I don't like the idea of READ COMMITTED suddenly throwing errors due to
concurrency problems. Using FOR UPDATE correctly is really tricky, and
this is just one example. And a documented one, at that, too.
.m