Status

Current state: Under discussionAccepted

Discussion thread: here

JIRA: : KAFKA-1997

Released: 0.8.3

Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).

...

The Mirror Maker has a potential dataloss data loss issue as explained below:

1. Mirror Maker consumer consume some messages and called producer.send(). The messages and put them into dataChannelsit in producer accumulator and haven't been sent yet.

2. Mirror Maker consumer commits the offsets

3. Mirror Maker somehow dies with some messages left in the data channel or producer cache accumulator not yet successfully delivered by producer.

4. When Mirror Maker restarts, all the messages that has been consumed and committed by consumer but not successfully delivered by producer yet will be lost.

...

The new mirror maker design is trying to solve the above issue as well as enhance the the mirror maker with the following features.

1.Currently Mirror Maker does not ensure the topics in source and target clusters have same partitions, which is required in many casesFiner control over the target partition.

2.Sometimes it is useful for Mirror Maker to Allow user do some simple process on messages (filter, message format change, etc.)of messages with a wired in message handler.

The new mirror maker is designed based on new consumer.

Public Interfaces

The no data loss solution uses the callback of new producer to determine the offsets to be committed.For Mirror Maker, an following interface change will be made to mirror maker.

• --consumer.rebalance.listener - to allow user to wire in a custom rebalance listener which implements ConusmerRebalanceCallback interface. An internal rebalance listener which implements ConsumerRebalanceCallback is wired in by default to avoid duplicates on consumer rebalance. User could still specify a custom listener class in command line argument. The internal rebalance listener will call that custom listener after it finishes the default logic.

Proposed Changes

Proposal 1

Image Removed

• A single consumer thread distribute messages into different data channel queues based on source target partition hash.
• multiple producer threads take responsibility of handling message and produce them to target cluster.
• Scale by creating more mirror maker instances or processing threads.

Pros: Decouple consume and process of messages. Allow adjustment on processing/consuming thread ratio.

Cons: More complicated structure than proposal 2.

Proposal 2

Image Removed

• --consumer.rebalance.listener.args - to provide arguments to custom rebalance listener constructor.
• --message.handler - to allow user do some simple process on the messages (e.g. filtering, reformatting, etc)
• --message.handler.args - to provide arguments for message handler init() function.
• --abort.on.send.fail - define whether mirror maker should abort when a send failed(turned on by default)

Proposed Changes

Image Added

• Each No data channel, each mirror maker thread consume messages, process them and then send them.
• Scale by creating more mirror maker thread.

Pros: Simple structure.

Cons: consumer and producer are coupled, which is less flexible. More TCP connection and memory footprint.

Common in both proposals

The only source of truth for offsets in consume-then-send pattern is target cluster. The offsets should only be committed after having received the confirmation from the target cluster. For the new producer, the confirmation is the ACK from target cluster.

Each mirror maker thread has a consumer instance associated with it, the thread will be responsible for decompression, message handling, compression and offset commit. All mirror maker threads share a producer.

The In that sense the consumer's offsets auto-commits should be turned off, and the offsets will be committed after a future associated with a message is received.

Offset commit thread

...

periodically following a flush() call on producer.

The mirror maker thread logic is as below.

while (!shutdown) {
	val records = consumer.poll(timeout)
	for (record <- records) {
		val handledRecords = messageHandler.handle(record)
		handledRecords.forEach(producer.send)
	}
    if (System.currentMilliSeconds - lastOffsetCommitMs > offsetCommitIntervalMs) {
		producer.flush()
		consumer.commitOffsets()
	}
}

Offset commit

Each mirror maker thread periodically flush() all the messages in producer and then commit offsets.

No data loss option

A no data loss option is provided with the following settings, and these are also default settings:

1. For consumer: auto.commit.enabled=false
2. For producer:
   1. max.in.flight.requests.per.connection=1
   2. retries=Int.MaxValue
   3. acks=-1
   4. block.on.buffer.full=true
3. set --abortOnSendFail

The following actions will be taken by mirror maker:

• Mirror maker will only send one request to a broker at any given point.
• If any exception is caught in mirror maker thread, mirror maker will try to commit the acked offsets then exit immediately.
• For RetriableException in producer, producer will retry indefinitely. If retry did not work, eventually the entire mirror maker will block on producer buffer full.
• For None-retriable exception, if --abort.on.send.fail is specified, stop the mirror maker. Otherwise producer callback will record the message that was not successfully sent but let the mirror maker move on. In this case, that message will be lost in target cluster.

On consumer rebalance

A new consumerRebalanceListener needs to be added to make sure there is no duplicates when consumer rebalance occurs.

/**
 * This listener is used for execution of tasks defined by user when a consumer rebalance
 * occurs in {@link kafka.consumer.ZookeeperConsumerConnector}
 */
public interface ConsumerRebalanceListener {
    /**
     * This method is called after all the fetcher threads are stopped but before the
     * ownership of partitions are released. Depending on whether auto offset commit is
     * enabled or not, offsets may or may not have been committed.
     * This listener is initially added to prevent duplicate messages on consumer rebalance
     * in mirror maker, where offset auto commit is disabled to prevent data loss. It could
     * also be used in more general cases.
     */
    public void beforeReleasingPartitions(Map<String, Set<Integer>> partitionOwnership);
  /**
   * This method is called after the new partition assignment is finished but before fetcher
   * threads start. A map of new global partition assignment is passed in as parameter.
   * @param consumerIdString The consumer instance invoking this rebalance callback
   * @param partitionAssignment The partition assignment result of current rebalance
   */
    public void beforeStartingFetchers(String consumerIdString, Map<String, Map<Integer, ConsumerThreadId>> partitionAssignment);
}

The callback beforeReleasingPartitions will be invoked  on consumer rebalance. It should

1. clear the dataChannel queue (as those messages will be re-consumed after rebalance)

2. wait until the producer drains all remaining messages

call producer.flush

23. commit the offsets

Ensure same partition number in source and target topics

Each consumer will have a local topic cache that records the topics it has ever seen.

When new topic is created in source cluster, consumer rebalance will be triggered. In the ConsumerRebalanceCallback:onPartitionAssigned:

For consumer that consumes partition 0 of a topic:

1. 1. Check the topic in both source and target cluster to see if it is a new topic.
   2. If new topic is detected, create the topic with the same partition number in target cluster as in source cluster

For consumers that saw topics that does not exist in local topic cache:

1. 1. Wait until the topic is seen on target cluster.

When a partition expansion occurs, consumer rebalance will also be triggered. Above approach does not take care of this case.

The default rebalance listener does not do anything in beforeStartingFetchers callback. This callback could be useful in some cases, for example, to create a topic in the target cluster with the same number of partitions as in source cluster.

Message handler in consumer thread

...

...

The changes are backward compatible.

Rejected Alternatives

NoneImage Added