On 2011-04-11 23:30, Alvaro Herrera wrote:
> Excerpts from Jesper Krogh's message of lun abr 11 17:07:33 -0300 2011:
>
>> But when the locking is done "row-level" then it is correct
>> to do it that way. It would allthough be nice with a weaker
>> locklevel for that kind of updates (I have no clue if that is
>> a hard problem).
> http://www.commandprompt.com/blogs/alvaro_herrera/2010/11/fixing_foreign_key_deadlocks/
>
That looks exactly what I have been seeing.
Naive suggestion (at least to part of the problem):
Would it be possible to identify updates that never
can violate any constraints and not do any verification
of foreign keys on the update and only pick a lock
that block concurrent updates of the same tuple?
UPDATE table set <something which is neither referenced or a reference>;
would all be of that type.
Would allthough require the database to examine
the UPDATE statement and in comparison with the
table definition figure out which of the column are
"safe" to update.
There might actually be a potential speedup since the update
would require to go visit the foreign table at all.
Jesper
--
Jesper
--
Jesper
