A couple of bug-fixes and a fail over capability for the partitioned client #69
Conversation
… a different node if the node goes down (and we do not notice immediately due to zookeeper delays).
…s in the consistent hash load balancer.
|
At this point, this pull request is not ready for merging into the main source line. I did want to create and publish it, so that the LinkedIn team can have visibility into what's happening in my fork. |
|
One more thing which I want to mention here, since I cannot submit bug reports anywhere: graceful shutdown does not work properly in Norbert right now. I tried to play around with the shutdown code, but I was unable to fix the issue. The problem appears to be the following: on shutdown the node marks itself offline. Once that happens further writes into the netty channel fail (so any ongoing requests on the node are not properly drained). Shutdown timeout does not work in this case either (I still see failures during testing). |
|
Thanks! Nick. I will take a look at it. |
… partitioned loadbalancer allows clients to manually specify to which partition a request should be routed.
… the Norbert version to have a suffix of -pb25 (protobuf 2.5). Altered the manner in which the session id is generated (it is now generated in the filter and passed onward into all services).
…e able to fail over to multiple nodes.
2 participants
The first couple of commits are pure bug-fixes; I think they should be applied to the main project.
I also added a fail over capability to the partitioned load balancer client and a capability to work with non monotonically increasing partition identifiers (this is still being tested).
Protocol buffers was brought to a more recent version, 2.5.0 and the protocol files were re-generated.