Versions Compared

Old Version 152

changes.mady.by.user Sophie Blee-Goldman

Saved on

compared with

New Version 153

changes.mady.by.user guozhang Wang

Saved on

Key

  • This line was added.
  • This line was removed.
  • Formatting was changed.

...

With the existing built-in Assignor implementations, they will be updated accordingly:


Highest VersionSupported StrategyNotes
RangeAssignor0EagerCurrent default value.
RoundRobinAssignor0Eager
StickyAssignor0Eager
CooperativeStickyAssignor0Eager, CooperativeTo be default value in 3.0
StreamsAssignor4Eager, Cooperative

The reason we make "range" and "round-robin" to not support cooperative rebalance is that, this protocol implicitly relies on the assignor to be somewhat sticky to make benefits by trading an extra rebalance. However, for these two assignors, they would not be sticky (although sometimes range may luckily reassign partitions back to old owners, it is not best-effort) and hence we've decided to not make them be selected for cooperative protocol. The existing StickyAssignor was not made to support Cooperative to ensure users follow the smooth upgrade path outlined below, and avoid running into trouble if they already use the StickyAssignor and blindly upgrade.

...

The specific upgrade path is described below. Note that this will be different depending on whether you have a plain consumer app or a Streams app, and make sure to follow the appropriate one.

...

From the user's perspective, the upgrade path of leveraging new protocols is just the same as switching to a new assignor. For example, assuming the current version of Kafka consumer is 2.2 and "range" assignor is specified in the config. The upgrade path would be:

...

  • The first rolling bounce is to replace the byte code (i.e. swap the jars): set the assignors to "range, cooperative-sticky" (or round-robin/sticky if you are using a different assignor already). At this stage, the new versioned byte code will still choose EAGER as the protocol and then sends both assignors in their join-group request, since there are at least one member who's not bounced yet and therefor will only send with "range", "range" assignor will be selected to assign partitions while everyone is following the EAGER protocol. This rolling bounce is safe.
  • The second rolling bounce is to remove the "range" (or round-robin/sticky) assignor, i.e. only leave the "cooperative-sticky" assignor in the config. At this stage, whoever have been bounced will then choose COOPERATIVE protocol and not revoke partitions while others not-yet-bounced will still go with EAGER and revoke everything. However the "cooperative-sticky" assignor will be chosen since at least one member who's already bounced will not have "range" any more. The "cooperative-sticky" assignor works even when there are some members in EAGER and some members in COOPERATIVE: it is fine as long as the leader can recognize them and make assignment choice accordingly, and for EAGER members, they've revoked everything and hence did not have any pre-assigned-partitions anymore in the subscription information, hence it is safe just to move those partitions to other members immediately based on the assignor's output.

    • The key point behind this two rolling bounce is that, we want to avoid the situation where leader is on old byte-code and only recognize "eager", but due to compatibility would still be able to deserialize the new protocol data from newer versioned members, and hence just go ahead and do the assignment while new versioned members did not revoke their partitions before joining the group. Note the difference with KIP-415 here: since on consumer we do not have the luxury to leverage on list of built-in assignors since it is user-customizable and hence would be black box to the consumer coordinator, we'd need two rolling bounces instead of one rolling bounce to complete the upgrade, whereas Connect only need one rolling bounce.

    ...

  • The first rolling bounce is to replace the byte code (i.e. swap the jars): set the UPGRADE_FROM config to 2.3 (or whatever version you are upgrading from) and then bounce each instance to upgrade it to 2.4. The UPGRADE_FROM config will turn off cooperative rebalancing in the cluster until everyone is on the new byte code, and we can be sure that the leader will be able to safely complete a rebalance. 
  • The second rolling bounce is to remove the UPGRADE_FROM config: simply remove this and bounce each instance for it to begin using the cooperative protocol. Note that unlike plain consumer apps, this means you will have some members on COOPERATIVE while others may still be on EAGER – as long as everyone is on version 2.4 or later, this is safe as the Streams assignor knows how to handle the assignment with either protocol in use. 

    • ...

    These changes should all be fairly transparent to Streams apps, as there are no semantics only improved rebalancing performance. However, users using Interactive Queries (IQ) or implementing a StateListener will notice that Streams spends less time in the REBALANCING state, as we will not transition to that until the end of the rebalance. This means all owned stores will remain open for IQ while the rebalance is in progress, and Streams will continue to restore active tasks if there are any that are not yet running, and will process standbys if there aren't. 

    Looking into the Future: Allow Consumer to Return Records in Rebalance

    As summarized in

    Jira
    serverASF JIRA
    serverId5aa69414-a9e9-3523-82ec-879b028fb15b
    keyKAFKA-8421
    , an optimization we can consider is letting consumers to still return messages that still belong to its owned partitions even when it is within a rebalance.

    In order to do this, we'd need to allow the consumer#commit API to throw RebalanceInProgressException if it is committing offset while a rebalance is undergoing.

    Code Block
    languagejava
        /**
     * ...
     *
     * @throws org.apache.kafka.common.errors.RebalanceInProgressException if the consumer instance is in the middle of a rebalance
     *            so it is not yet determined which partitions would be assigned to the consumer. In such cases you can first
     *            complete the rebalance by calling {@link #poll(Duration)} and commit can be reconsidered afterwards.
     *            NOTE when you reconsider committing after the rebalance, the assigned partitions may have changed,
     *            and also for those partitions that are still assigned their fetch positions may have changed too
     *            if more records are returned from the {@link #poll(Duration)} call.
     * ...
     */
    @Override
    public void commitSync() {
        commitSync(Duration.ofMillis(defaultApiTimeoutMs));
    }


    Looking into the Future: Heartbeat Communicated Protocol

    ...

    The existing built-in Assignor implementations will then be updated to:


    		Highest VersionSupported StrategyNotes
    RangeAssignor0EagerCurrent default.
    RoundRobinAssignor0Eager
    StickAssignor (old)0Eager
    StickAssignor (new)0CooperativeWill be new default in 3.0
    StreamsAssignor (old)4Eager
    StreamsAssignor (new)4Cooperative


    Although it makes the upgrade path simpler since we would no longer need the "rebalance.protocol" config on consumer anymore, while just encoding multiple assignors during the first rolling bounce of the upgrade path, it requires duplicated assignor class (of course, the new class could just be extending from the old one and there's not much LOC duplicated) which is a bit cumbersome.

    ...