It's possible for a node request to disappear while we are trying to fulfill it. That will cause an unhandled exception and the fake nodepool will stop. Log all exceptions and continue running. Also add an exception handler for the specific case encountered. Also, add a canceled flag to node requests to aid in some race condiations. The _updateNodeRequest method can be called out-of-band by ZK. It can race with the scheduler calling cancelRequest, which is significant for deleting the request from Nodepool.requests as well as handing off a completed request to the scheduler. The canceled flag on requests aids in that by allowing us to only delete from Nodepool.requests within the _updateNodeRequest method. It also helps us detect when a request was canceled after it was fulfilled and handed off to the scheduler. In those cases, the cancel method will have already deleted the request, but the scheduler would then procces it and perhaps try to delete it again. We check the canceled flag inside of the scheduler main loop. Since the scheduler main loop is also the only thread which can call the cancel method, we can be assured that it won't change while we evaluate it. If we find that we ended up with a completed request which was canceled, we just ignore it since we know that the cancel method would have deleted the request. Change-Id: I30f854a869d7690ab50340ba4c02067c4ae9fe2bchanges/02/466302/3
parent
ff97edb615
commit
cbbce0d12c
Loading…
Reference in new issue