Open Stack

On 06/12/2015 01:17 AM, Salvatore Orlando wrote:
> It is however interesting that both "lock wait timeouts" and "missing
> savepoint" errors occur in operations pertaining the same table -
> securitygroups in this case.
> I wonder if the switch to pymysl has not actually uncovered some other
> bug in Neutron.
> 
> I have no opposition to a revert, but since this will affect most
> projects, it's probably worth finding some time to investigate what is
> triggering this failure when sqlalchemy is backed by pymysql before
> doing that.

Right, we knew that the db driver would move some bugs around because
we're no longer blocking python processes on db access (so there used to
be a pseudo synchronization point before you ever got to the database).

My feeling is this should be looked into before it is straight reverted
(are jobs failing beyond Rally?). There are a number of benefits with
the new driver, and we can't get to python3 with the old one.

Rally failing is also an indicator that just such and implicit lock was
behavior that was depended on before, because it will be sending a bunch
of similar operations all at once as a kind of stress test. It would
tend to expose issues like this first.

	-Sean

-- 
Sean Dague
http://dague.net