lianetm commented on code in PR #15215:
URL: https://github.com/apache/kafka/pull/15215#discussion_r1462378625


##########
clients/src/main/java/org/apache/kafka/clients/consumer/internals/MembershipManagerImpl.java:
##########
@@ -1134,9 +1134,38 @@ private CompletableFuture<Void> assignPartitions(
         // Make assignment effective on the client by updating the 
subscription state.
         updateSubscription(assignedPartitions, false);
 
+        // Mark assigned partitions as pendingOnAssignedCallback to 
temporarily stop fetching or
+        // initializing positions for them. Passing the full set of assigned 
partitions
+        // (previously owned and newly added), given that they are all 
provided to the user in the
+        // callback, so we could expect offsets updates for any of them.
+        Set<TopicPartition> assignedTopicPartition = 
assignedPartitions.stream().map(tIdp -> 
tIdp.topicPartition()).collect(Collectors.toSet());
+        subscriptions.markPendingOnAssignedCallback(assignedTopicPartition, 
true);
+
         // Invoke user call back.
         CompletableFuture<Void> result = 
invokeOnPartitionsAssignedCallback(addedPartitions);
 
+        // Resume partitions only if the callback succeeded.
+        result.whenComplete((error, callbackResult) -> {
+            if (error == null) {
+                // Remove pendingOnAssignedCallback flag from the assigned 
partitions, so we can
+                // start fetching, and updating positions for them if needed.
+                
subscriptions.markPendingOnAssignedCallback(assignedTopicPartition, false);
+            } else {
+                // Remove pendingOnAssignedCallback flag from the previously 
owned only so that

Review Comment:
   @dajac I've been debating about the right behaviour on this failure 
scenario. Thinking about what is expected to happen after the callback fails, 
it seemed sensible to ensure that the partitions that were previously owned are 
back to normal (fetchable and initializing positions if needed, in case the 
user just ignores the exception and continues polling), and the partitions that 
were trying to be added are not (since the consumer won't ack them to the 
broker, and will remove them from the subscription, or try again based on the 
next assignment received. Either way we don't want any progress on them). 
Thoughts? Am I missing something maybe?



-- 
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.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]

Reply via email to