On 10/30/2013 10:42 AM, Clint Byrum wrote:
So, recently we've had quite a long thread in gerrit regarding locking
in Heat:
https://review.openstack.org/#/c/49440/
In the patch, there are two distributed lock drivers. One uses SQL,
and suffers from all the problems you might imagine a SQL based locking
system would. It is extremely hard to detect dead lock holders, so we
end up with really long timeouts. The other is ZooKeeper.
I'm on record as saying we're not using ZooKeeper. It is a little
embarrassing to have taken such a position without really thinking things
through. The main reason I feel this way though, is not because ZooKeeper
wouldn't work for locking, but because I think locking is a mistake.
The current multi-engine paradigm has a race condition. If you have a
stack action going on, the state is held in the engine itself, and not
in the database, so if another engine starts working on another action,
they will conflict.
The locking paradigm is meant to prevent this. But I think this is a
huge mistake.
The engine should store _all_ of its state in a distributed data store
of some kind. Any engine should be aware of what is already happening
with the stack from this state and act accordingly. That includes the
engine currently working on actions. When viewed through this lense,
to me, locking is a poor excuse for serializing the state of the engine
scheduler.
It feels like TaskFlow is the answer, with an eye for making sure
TaskFlow can be made to work with distributed state. I am not well
versed on TaskFlow's details though, so I may be wrong. It worries me
that TaskFlow has existed a while and doesn't seem to be solving real
problems, but maybe I'm wrong and it is actually in use already.
Anyway, as a band-aid, we may _have_ to do locking. For that, ZooKeeper
has some real advantages over using the database. But there is hesitance
because it is not widely supported in OpenStack. What say you, OpenStack
community? Should we keep ZooKeeper out of our.. zoo?
I will -2 any patch that adds zookeeper as a dependency to Heat.
The rest of the idea sounds good though. I spoke with Joshua about
TaskFlow Friday as a possibility for solving this problem, but TaskFlow
presently does not implement a distributed task flow. Joshua indicated
there was a celerity review at https://review.openstack.org/#/c/47609/,
but this would introduce a different server dependency which suffers
from the same issues as Zookeeper, not to mention incomplete AMQP server
support for various AMQP implementations. Joshua indicated using a pure
AMQP implementation would be possible for this job but is not implemented.
I did get into a discussion with him about the subject of breaking the
tasks in the flow into "jobs", which led to the suggestion that the
parser should be part of the API server process (then the engine could
be responsible for handling the various jobs Heat needs). Sounds like
poor abstraction, not to mention serious rework required.
My take from our IRC discussion was that TaskFlow is not a job
distribution system (like Gearman) but an in-process workflow manager.
These two things are different. I was unclear if Taskflow could be made
to do both, while also operating under already supported AMQP server
infrastructure that all of OpenStack relies on currently. If it could,
that would be fantastic, as we would only have to introduce a library
dependency vs a full on server dependency with documentation, HA and
scalability concerns.
Regards
-steve
_______________________________________________
OpenStack-dev mailing list
OpenStack-dev@lists.openstack.org
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
_______________________________________________
OpenStack-dev mailing list
OpenStack-dev@lists.openstack.org
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
