Open Stack

On 07/07/2014 04:17 PM, Mike Bayer wrote:
>
> On 7/7/14, 3:57 PM, Matt Riedemann wrote:
>>
>>
>>
>> Regarding the eventlet + mysql sadness, I remembered this [1] in the
>> nova.db.api code.
>>
>> I'm not sure if that's just nova-specific right now, I'm a bit too
>> lazy at the moment to check if it's in other projects, but I'm not
>> seeing it in neutron, for example, and makes me wonder if it could
>> help with the neutron db lock timeouts we see in the gate [2].  Don't
>> let the bug status fool you, that thing is still showing up, or a
>> variant of it is.
>>
>> There are at least 6 lock-related neutron bugs hitting the gate [3].
>>
>> [1] https://review.openstack.org/59760
>> [2] https://bugs.launchpad.net/neutron/+bug/1283522
>> [3] http://status.openstack.org/elastic-recheck/
>
>
> yeah, tpool, correct me if I'm misunderstanding, we take some API code
> that is 90% fetching from the database, we have it all under eventlet,
> the purpose of which is, IO can be shoveled out to an arbitrary degree,
> e.g. 500 concurrent connections type of thing, but then we take all the
> IO (MySQL access) and put it into a thread pool anyway.

Yep. It makes no sense to do that, IMO.

The solution is to use a non-blocking MySQLdb library which will yield 
appropriately for evented solutions like gevent and eventlet.

Best,
-jay