Proposed Changes

1) Changes in the controller for handling topic creation and the existing topic znodes

- When the topic znode is created, controller should update the topic znode with the initial_partition_count. Controller should always write the initial_partition_count for existing topic znodes whose initial_partition_count is not specified.

initial_partition_count will be used by producer to map message key to partition.

2) Changes in the controller for handling partition expansion

- User executes kafka-topics.sh to update the topic znode with the new assignment. This triggers the topic znode listener in controller.

- Controller updates the topic znode with the new undeleted_partition_count.

- For those partitions of this topic which already have the partition znode, controller increments their leaderEpoch by 1 in the partition znode. Controller sends LeaderAndIsrRequest and wait for LeaderAndIsrResponse. The LeaderAndIsrRequest should include the new leaderEpoch for each partition and the updated undeleted_partition_count of the topic.

- For each partition of this topic which does not have the partition znode, controller creates the partition znode, such that the leaderEpochAfterCreation field in the znode data maps partition of this topic to the corresponding leaderEpoch (recorded before controller increments the leaderEpoch)

- Controller propagates the UpdateMetadataRequest with the latest undeleted_partition_count per topic.

- Controller continue the existing logic of partition expansion.

Note that this procedure is fault tolerant. If controller fails in any of these step, the new controller can continue creating partition znode following the same procedure.

The motivation of this change is that, for each existing partition before the partition expansion, there exists a leaderEpoch such that all messags after this leaderEpoch will be produced with the producers using the incremented undeleted_partition_count. If for all existing partitions, a consumer group has delivered all messages up to such leaderEpoch, then it should be safe to consume messages from the newly created partition without worrying about out-of-order delivery. Consumers can get such leaderEpoch from the partition znode, convert leaderEpoch to lowerOffsetThreshold using OffsetsForLeaderEpochRequest, and make sure that it only starts consuming new partition after messages before the lowerOffsetThreshold have all been consumed.

3) Changes in the controller for handling partition deletion

- User executes "kafka-topics.sh --alter --partition partitionNum --topic topic" to alter the partition count of the given topic. If the given undeleted_partition_count is smaller than the current undeleted_partition_count of the given topic AND no smaller than the initial_partition_count, the script writes the given undeleted_partition_count in the topic znode. This triggers the topic znode listener in controller.

- For those partitions P1 whose partitionId >= undeleted_partition_count, if the leaderEpochBeforeDeletion in the partition znode empty or it is not specified, controller updates partition znode of P1 such that for every undeleted partitions P2, leaderEpochBeforeDeletion maps the partition P2 to its current leaderEpoch. Controller also increments the leaderEpoch by 1 for partition znode of every P2. Then controller sends LeaderAndIsrRequest with the updated undeleted_partition_count for this topic and the updated leaderEpoch for P2.

- Controller propagates the UpdateMetadataRequest with the latest undeleted_partition_count per topic.

- Controlelr schedules a task that periodically which periodically sens ListOffsetRequest to check the size of those partitions which have been marked for deletion, i.e. with non-empty leaderEpochBeforeDeletion map in the parittion znode. If the size of a yet-to-be-deleted partition is 0, i.e. its logStartOffset == logEndOffset, controller sends StopReplicaRequest with delete=true, removes the partition from the assignment in the topic znode, and removes the partition znode.

Note that this procedure is fault tolerant. If controller fails in any of these step, the new controller can continue to read the undeleted_partition_count from the topic znode, update partition znode and sends LeaderAndIsrRequest.

The motivation of this change is that, for each undeleted partition after the partition deletion, there exists a leaderEpoch such that all messages up to this leaderEpoch are produced with the producers using the old undeleted_partition_count prior to the partition deletion. If the consumer group has delivered all messages in a partition that has been marked for deletion, it should be safe for this consumer group to consume any message in undeleted partition. If the consumer group has not delivered some messages in a partition that has been marked for deletion, it should still be safe for this consumer group to delivery messages up to the corresponding leaderEpoch in the undeleted partitions.

4) Changes in how broker handles ProduceRequest

- When broker receives LeaderAndIsrRequest, in addition to the existing procedure (e.g. updating the leaderEpochCache for the new leaderEpoch), the broker should also record in memory the undeleted_partition_count for each topic.

- When broker receives ProduceRequest, for each partition in the request, broker checks whether its undeleted_partition_count equals the undeleted_partition_count from the most recent LeaderAndIsrRequest. If yes, broker handles the produce request in the current way. If no, broker rejects this partition with InvalidPartitionMetadataException. This error extends InvalidMetadaException and should trigger producer to update its metadata and retry.

The motivation of this change is that all messags after the incremented leaderEpoch (read from the LeaderAndIsrRequest) will be produced by producers using the latest undeleted_partition_count (also read from the same LeaderAndIsrRequest). Note that the leaderEpoch and the undeleted_partition_count are updated atomically using the same LeaderAndIsrRequest.

5) Changes in how producer constructs ProduceRequest

- Producer should include the undeleted_partition_count for each topic in the ProduceRequest.

- Producer should determine the partition for both keyed and unkeyed message using Linear Hashing (see https://en.wikipedia.org/wiki/Linear_hashing), where N = initial_partition_count, L is the largest integer that satisfies N*2^L <= undeleted_partition_count, and S can similariy be derived using N, L and undeleted_partition_count.

- Producer will update metadata and retry ProduceRequest if ProduceResponse shows InvalidPartitionMetadataException, which happens if producer's undeleted_partition_count is different (maybe newer or older) than the undeleted_partition_count in the broker.

The motivation of these changes are:

- Producer always use the undeleted_partition_count in the leader broker to determine the partition of each message
- The benefits of using linear hashing as compared to existing hashing algorithm include:
   - Each newly added partition will be blocked on exactly one existing partition before it can be consumed.
   - Each newly deleted partition will block the consumption of at most O(log(n)) remaing partitions.
   - There is no movement of keys between existing partitions so that the existing partitions do not block on each other.
   - The movement of keys between partitions is much less and thus less state needs to be loaded/stored for stream processing.

Note: the extra steps introduced in here is skipped for partitions of topics whose enable.ordered.delivery is set to false.

6) Changes in the leader of the consumer group

- Leader of the consumer group splits the partition list to those partitions that have not been marked for deletion and those partitions that have been marked for deletion. It should apply the user-defined assignment algorithm to these two lists seperately to determine the partition distribution across consumers in the group, so that partitions which have not been marked for deletion can also be evenly distributed across consumers. This is to prevent load imbalance across consumers because there will be no new data to those partitions which have been marked for deletion.

The motivation of this change is to evenly distribute the byte-in-rate of undeleted partitions across consumers of the consumer group. Note that even though there is still remaning data in the partitions which have been marked for deletion, there is no new data into these partitions. Thus a consumer will likely be under-utilized if it is assigned few undeleted partitions but many partitons which have been marked for deletion.

7) Changes in how consumer queries position of partitions of the consumer group using ConsumerGroupPositionRequest and ConsumerGroupPositionResponse

- Periodically, after every heartbeat.interval.ms, consumer sends ConsumerGroupPositionRequest to the consumer group coordinator.

- ConsumerGroupPositionRequest includes the current position for each partition requested by the coordinator from the previous ConsumerGroupPositionResponse. It also includes the list of partitions for which it wants to know the position (of the consumer that is consuming this partitoin). See "Changes in how consumer consumes partition" below for how consumer determines the list of partitions for which it wants to know the position.

- Group coordinator keeps a map that maps the partition to the the list of consumer ids that are interested in the position of this partition. Group coordinator also remembers the positions for those partitions which are interesting to some consumers of the given group. Both information may be updated when coordinator receives ConsumerGroupPositionRequest.

- ConsumerGroupPositionResponse includes the list of partitions that this consumer is interested in. For each partition P1 that is assigned to this consumer AND whose position is interesting to some consumers of the consumer group, if the following following condition is satisfied , the consumer should execute the callback PartitionKeyRebalanceListener.onPartitionKeyMaybeRevoked(Collection<TopicPartition>) and include the position of the partition P1 in the ConsumerGroupPositionResponse.

Here is the condition for including partition P1. This condition optimizes the overhead of onPartitionKeyMaybeRevoked() callback such that consumer will not trigger callback and will not include the position of the partition in the ConsumerGroupPositionResponse when it knows for sure that this position will not unblock other consumers from starting to consume some partitions.

(The position of partition P1 has not been included in ConsumerGroupPositionResponse since this consumer starts) ||
(The position of partition P1 has increased since the last ConsumerGroupPositionResponse AND it equals the high watermark of P1 in the last FetchReponse) ||
(The largest leaderEpoch of those messages before position P1 has increased since the last ConsumerGroupPositionResponse)

The motivation of this change is to allow consumers of the same consumer group to communicate the position of partitions with each other. This allows a consumer to start consuming beyond a given offset of a partition only after a certain condition is satisfied for the position of other partitions in the consumer group.

Note: the extra steps introduced in here is skipped for partitions of topics whose enable.ordered.delivery is set to false.

8) Changes in how consumer consumes partition

1. Consumer receives SyncGroupResponse, which contains its assigned partitions

2. Consumer gets the startPosition, i.e.the committedOffset, for its assigned partitions.

3. Consumer sends ListOffsetRequest to get the earliest offset for its assigned partitions.

4. For each partition P1 whose startPosition is not available, or whose startPosition equals the earliest offset, if the partition index >= initial_partition_count of the topic, consumer does the following before consuming the partition P1:

    4.1 Consumer sends PartitionLeaderEpochsForPartitionsRequest to the coordinator to get the leaderEpochAfterCreation map for the partition P1, which can be read by broker from the corresponding partition znode.

    4.2 There will be exactly one parition P3 that blocks this consumer from consuming the partition P1. Since producer uses linear hashing, partition P3 can be determined based on the set of partitions prior to partition expansion (i.e. keys of the leaderEpochAfterCreation map), the partition index of P1 and the initial_partition_count of the topic.

    4.3 Then the consumer gets the oldLeaderEpoch for P3 from leaderEpochAfterCreation and sends OffsetsForLeaderEpochRequest to convert the oldLeaderEpoch to lowerOffsetThreshold, where lowerOffsetThreshold should be the last offset of messages published under the oldLeaderEpoch for partition P3.

    4.4 Consumer includes the partition P3 in the ConsumerGroupPositionRequest and gets the corresponding position of partition P3 of the consumer group in the ConsumerGroupPositionResponse. If the position of P3 > lowerOffsetThreshold of the priorParition, then the consumer executes the callback PartitionKeyRebalanceListener.onPartitionKeyAssigned(Collection<TopicPartition> partitions) for P1 and starts to consume partition P1.

5. For each parrtition P1 assigned to this consumer, consumer queries the metadata to see if there any partition of this topic has been marked for deletion. If so, consumer does the following before deliverying message with offset T from this partition P1:

    5.1 For every partition P3 that has been marked for deletion but not yet removed from this topic, consumer can determine whether partition P1 should be blocked on consuming all messages from P3. There will be at most O(logn) such partitions, where n is the maximum index of any partition of this topic. Since producer uses linear hashing, this can be determined based on index of partition P3, index of partition P1 and the initial_partition_count of the topic.

    5.2 For every partition P3 that may block the consumption of partition P1, consumer sends PartitionLeaderEpochsForPartitionsRequest to coordinator to get the leaderEpochBeforeDeletion map for P3. Note that a partition is marked for deletion if and only if the leaderEpochBeforeDeletion map in its partition znode is not empty. Then the consumer gets the oldLeaderEpoch for P1 from the leaderEpochBeforeDeletion map of P3. Next, consumer sends OffsetsForLeaderEpochRequest to convert the oldLeaderEpoch to upperOffsetThreshold for P1. This results in a map from P3 -> upperOffsetThreshold for every partition P3 which blocks the consumption of partition P1.

    5.2 For every partition P3 which may block the consumption of partition P1, consumer includes this partition in the ConsumerGroupPositionRequest and gets the corresponding position of this partition of the consumer group in the ConsumerGroupPositionResponse. Consumer also sends ListOffsetRequest to get the LogEndOffset for this partition.

    5.3 For every partition P3 which may block the consumption of partition P1, if either the position of P3 of the consumer group has reached the LogEndOffset of P3 or the offset T <= upperOffsetThreshold, then the consumer will execute the callback PartitionKeyRebalanceListener.onPartitionKeyAssigned(Collection<TopicPartition> partitions) for P1 and start to deliver messages with offset >= T from partition P1. Otherwise, the consumer can not delievery message with offset >= T from partition P1.

6. Additionally, consumer should include the undeleted_partitiont_count in the FetchRequest. If the undeleted partition count is different from what broker expects for the given topic, broker will specify InvalidPartitionMetadataException for the given partition in the FetchRequest. This should trigger metadata update in consumer and cause consumer to rejoin the group. The purpose of this change is that, when we shrink partition of a given topic, consumers in the given consumer group can perform step 5 above to ensure that messages are delivered in order. 

This change, combined with the controller change 1) and 2) described above, ensures in-order message delivery for keyed messages in the presence of partition creation and deletion.

Note: the extra steps introduced in here is skipped for partitions of topics whose enable.ordered.delivery is set to false.

9) Changes in how producer produced produces keyed messages to log compacted topics

If enable.ordered.delivery is true, message has non-null key, the topic is log compacted, and the initial_partition_count < undeleted_partition_count, producer may additionally send a message as described below:

- Say the original message to be sent by user has key K1 and value V1. Suppose this message is mapped to partition P1 by linear hashing where P1 >= initial_partition_count. And P2 is the partition which was split into {P1, P2} after partition expansion.

- Producer sends PartitionLeaderEpochsForPartitionsRequest to the coordinator to get the leaderEpochAfterCreation map for the partition P1.

- Producer gets the oldLeaderEpoch for partition P2 from the leaderEpochAfterCreation map. Then it sends OffsetsForLeaderEpochRequest to convert the oldLeaderEpoch to lowerOffsetThreshold, where lowerOffsetThreshold should be the last offset of messages published under the oldLeaderEpoch for partition P2.

- Producer sends ListOffsetRequest to get the earliest offset of P2. If the earliest offset > lowerOffsetThreshold of P2, producer sends message with key=K1 and value=null to P2 with ack=0 in adddition to sending the message key=K1 and value=V1 to P1.

The motivation of this change is to allow messages in the log compacted topic to be removed after a new message with the same key is produced after the partiton expansion.