1) The deployment assignment is stored using a map [node ID -> number of assigned services]. However, this assignment is not very effective for cases when service configuration is (maxPerCluster = 0, maxPerNode > 0), because in this case, we can avoid assignment recalculation at all. The assignment for this case may look like (eachNode=N). In this case, the assignment does not change and we can effectively skip it during the reassign loop.
2) We store zero assignment counters, which does not make sense at all - if there are no service deployments for a node, there should be no corresponding entry in the map at all. The size of assignments for (maxPerCluster > 0) configurations is O(number of nodes in the cluster), but it should be O(maxPerCluster).
3) If an assignment did not change, we should not commit the assignment transaction - this is redundant
4) Perhaps, it also makes sense to calculate several assignments at once and do a putAll commit instead of single puts - this should also decrease the assignment calculation latency