286659a38b
In case of a broker restart/failover a reply queue can be unreachable for short period the IncomingMessage.send_reply will block for 60 seconds in this case or until rabbit recovers. But in case of the reply queue is unreachable because the rpc client is really gone, we can have a ton of reply to send waiting 60 seconds. This leads to a starvation of connection of the pool The rpc server take to much time to send reply, other rpc client will raise TimeoutError because their don't receive their replies in time. This changes introduces an object cache that stores already known gone client to not wait 60 seconds and hold a connection of the pool Keeping 200 last gone rpc client for 1 minute is enough and doesn't hold to much memory. This also don't raise anymore a frightening exception when we can't send reply to the rpc client. But just logging a info about missing exchange and a warning about unsend reply. Closes-bug: #1460652 Change-Id: I928b30c9b5f9ee007532ff703e136640b0e8aaf4
Oslo Messaging Library
The Oslo messaging API supports RPC and notifications over a number of different messaging transports.
- License: Apache License, Version 2.0
- Documentation: http://docs.openstack.org/developer/oslo.messaging
- Source: http://git.openstack.org/cgit/openstack/oslo.messaging
- Bugs: http://bugs.launchpad.net/oslo.messaging