mjsax commented on code in PR #23803: URL: https://github.com/apache/kafka/pull/23803#discussion_r4561131936
########## docs/streams/developer-guide/streams-rebalance-protocol.md: ########## @@ -39,7 +39,7 @@ The following features are available in the current release: * **Core Streams Group Rebalance Protocol**: The `group.protocol=streams` configuration enables the dedicated streams rebalance protocol. This separates streams groups from consumer groups and provides a streams-specific group membership lifecycle and metadata management on the broker. -* **Sticky Task Assignor**: A basic task assignment strategy that minimizes task movement during rebalances is included. It is registered under the name `sticky` and is the default assignor. +* **Built-in Task Assignors**: Two task assignment strategies are included. The `sticky` assignor minimizes task movement during rebalances and is the default assignor. The `balanced` assignor spreads the tasks of each subtopology across processes and evens out the per-thread task load, at the price of moving more tasks when the group membership changes; it is the counterpart of the `HighAvailabilityTaskAssignor` of the classic protocol, without warmup tasks, which are not an assignor feature in the streams protocol, and without rack-aware placement, which neither built-in assignor implements yet. A group selects an assignor with the group configuration `streams.assignor.name`. Review Comment: > evens out the per-thread task load So does the sticky assignor. "balanced" does not mean "even load" (I know, it easily misunderstood, but new could never come up with a better name...), but it's about "balancing the task of subtopology across all processed/members", ie, ie avoid "clustering". For the sticky assignor, because we try to avoid task movements, we would end up with a single process hosting all tasks of a sub-topology, what implies if this process fails, we lose all tasks across all partitions, and processing stops entirely. -- The balances assignor avoids this, and thus ensure that we only lose a subset of tasks from a subtoplogy, avoiding a full "horizontal" outage. -- 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]
