jiajunwang commented on a change in pull request #639: Refine the WAGED
rebalancer to minimize the partial rebalance workload.
URL: https://github.com/apache/helix/pull/639#discussion_r359111212
##########
File path:
helix-core/src/main/java/org/apache/helix/controller/rebalancer/waged/model/ClusterModelProvider.java
##########
@@ -93,42 +184,126 @@ public static ClusterModel
generateClusterModel(ResourceControllerDataProvider d
return new ClusterModel(context, toBeAssignedReplicas, assignableNodes);
}
+ // Filter the replicas map so only the replicas that have been allocated in
the existing
+ // assignmentMap remain in the map.
+ private static void retainExistingReplicas(Map<String,
Set<AssignableReplica>> replicaMap,
+ Map<String, ResourceAssignment> assignmentMap) {
+ replicaMap.entrySet().parallelStream().forEach(replicaSetEntry -> {
+ // <partition, <state, instances set>>
+ Map<String, Map<String, Set<String>>> stateInstanceMap =
+ getStateInstanceMap(assignmentMap.get(replicaSetEntry.getKey()));
+ // Iterate the replicas of the resource to find the ones that require
reallocating.
+ Iterator<AssignableReplica> replicaIter =
replicaSetEntry.getValue().iterator();
+ while (replicaIter.hasNext()) {
+ AssignableReplica replica = replicaIter.next();
+ Set<String> validInstances =
+ stateInstanceMap.getOrDefault(replica.getPartitionName(),
Collections.emptyMap())
+ .getOrDefault(replica.getReplicaState(),
Collections.emptySet());
+ if (validInstances.isEmpty()) {
+ // Removing by comparing with the baseline assignment.
+ replicaIter.remove();
+ } else {
+ // Remove the instance from the state map record, so it won't be
picked up again for
+ // the other replica checkup.
+ validInstances.remove(validInstances.iterator().next());
+ }
+ }
+ });
+ }
+
/**
- * Generate a cluster model based on the current state output and data cache.
- * @param dataProvider The controller's data cache.
- * @param resourceMap The full list of the resources to be
rebalanced. Note that any
- * resources that are not in this list will be
removed from the
- * final assignment.
- * @param currentStateAssignment The resource assignment built from current
state output.
- * @return A cluster model based on the current state and data cache.
+ * Find the minimum set of replicas that need to be reassigned by comparing
the Best
+ * possible assignment with the Baseline assignment.
+ * A replica needs to be reassigned if either of the following conditions is
true:
+ * 1. The partition allocation in the Baseline and the Best possible
assignment are different.
+ * And the allocation in the Baseline is valid. So it is worthwhile to move
it.
+ * 2. The partition allocation is not in the Baseline or the Best possible
assignment.
+ * Otherwise, the rebalancer just keeps the current Best possible assignment
allocation.
Review comment:
That logic is overall correct. But that is the combined logic of
findToBeAssignedReplicasByComparingWithBaseline and retainExistingReplicas. For
this method, the logic is covered with this description. I made some
improvement for this. Please check back later after I update.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
