[openstack-dev] [Trove] Guest RPC API improvements. Messages, topics, queues, consumption.

Tim Simpson tim.simpson at rackspace.com
Fri Oct 31 17:49:13 UTC 2014


Hi Denis,

It seems like the issue you're trying to solve is that these 'prepare' messages can't be consumed by the guest.
So, if the guest never actually comes online and therefore can't consume the prepare call, then you'll be left with the message in the queue forever.

If you use a ping-pong message, you'll still be left with a stray message in the queue if it fails.

I think the best fix is if we delete the queue when deleting an instance. This way you'll never have more queues in rabbit than are needed.

Thanks,

Tim


________________________________
From: Denis Makogon [dmakogon at mirantis.com]
Sent: Friday, October 31, 2014 4:32 AM
To: OpenStack Development Mailing List
Subject: [openstack-dev] [Trove] Guest RPC API improvements. Messages, topics, queues, consumption.


Hello, Stackers/Trovers.



I’d like to start discussion about how do we use guestagent API that will eventually be evaluated as a spec. For most of you who well-known with Trove’s codebase knows how do Trove acts when provisioning new instance.

I’d like to point out next moments:

  1.  When we provision new instance we expect that guest will create its topic/queue for RPC messaging needs.

  2.  Taskmanager doesn’t validate that guest is really up before sending ‘prepare’ call.

And here comes the problem, what if guest wasn’t able to start properly and consume ‘prepare’ message due to certain circumstances? In this case ‘prepare’ message would never be consumed.


Me and Sergey Gotliv were looking for proper solution for this case. And we end up with next requirements for provisioning workflow:

  1.  We must be sure that ‘prepare’ message will be consumed by guest.

  2.  Taskmanager should handle topic/queue management for guest.

  3.  Guest just need to consume income messages for already existing topic/queue.

As concrete proposal (or at least topic for discussions) i’d like to discuss next improvements:

We need to add new guest RPC API that will represent “ping-pong” action. So before sending any cast- or call-type messages we need to make sure that guest is really running.


Pros/Cons for such solution:

  1.  Guest will do only consuming.

  2.  Guest would not manage its topics/queues.

  3.  We’ll be 100% sure that no messages would be lost.

  4.  Fast-fail during provisioning.

  5.  Other minor/major improvements.



Thoughts?


P.S.: I’d like to discuss this topic during upcoming Paris summit (during contribution meetup at Friday).



Best regards,

Denis Makogon

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openstack.org/pipermail/openstack-dev/attachments/20141031/1fe6f924/attachment.html>


More information about the OpenStack-dev mailing list