Status

Current state: "Under Discussion"

Discussion thread: here

JIRA: Unable to render Jira issues macro, execution error.

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

Motivation

The Connect framework allows sink connector tasks to do their own offset tracking in case they want to do asynchronous processing (for instance, buffering records sent by the framework to be flushed to the sink system at some later time). The SinkTask::preCommit method allows sink task implementations to provide the framework with the consumer offsets to be committed for each topic partition. There's currently an incompatibility between Sink connectors overriding the SinkTask::preCommit method and SMTs that mutate the topic field.

The problem was present since the SinkTask::preCommit method's inception and is rooted in a mismatch between the topic/partition that is passed to open/preCommit (the original topic and partition before applying any transformations) and the topic partition that is present in the SinkRecord that the SinkTask::put method receives (after transformations are applied). Since that's all the information the connector has to implement any kind of internal offset tracking, the topic/partitions it can return in preCommit will correspond to the transformed topic, when the framework actually expects it to be the original topic.

In Unable to render Jira issues macro, execution error. we fixed the problem for connectors that don't override  SinkTask::preCommit. For the others, we acknowledge that "broader API changes are required". 

Public Interfaces

This KIP proposes to introduce a new overloaded public void put(Collection<SinkRecord> records, Map<SinkRecord, TopicPartition> originalTopicPartitions method in addition to the existing public abstract void put(Collection<SinkRecord> records) method in the org.apache.kafka.connect.sink.SinkTask abstract class. The new method has an additional parameter which provides a mapping from SinkRecord objects to their corresponding ConsumerRecord's original topic partition (i.e. before it has been potentially mutated by an SMT). SinkTask implementations can use this mapping to obtain the original topic partition for each record in order to do their own offset tracking and will be able to provide accurate offsets back to the framework through the SinkTask::preCommit method

org.apache.kafka.connect.sink.SinkTask

    /**
     * Put the records in the sink.
     *
     * If this operation fails, the SinkTask may throw a {@link org.apache.kafka.connect.errors.RetriableException} to
     * indicate that the framework should attempt to retry the same call again. Other exceptions will cause the task to
     * be stopped immediately. {@link SinkTaskContext#timeout(long)} can be used to set the maximum time before the
     * batch will be retried.
     *
     * In case {@link #put(Collection, Map)} has been overridden in order to obtain original topic partitions for records
     * (i.e. before SMTs are applied), this method can be implemented to simply call {@link #put(Collection, Map)} in this manner:
     *
     * <pre>
     * {@code
     * put(records, Collections.emptyMap())
     * }
     * </pre>
     *
     * @param records the set of records to send
     */
    public abstract void put(Collection<SinkRecord> records);

    /**
     * Put the records in the sink. This method should be overridden by sink tasks that do their own offset tracking, in
     * order to obtain the original topic and partition for each record (before any potentially topic mutating SMTs are applied).
     * @param records the set of records to send
     * @param originalTopicPartitions The original topic and partition for each record, before SMTs have been applied.
     */
    public void put(Collection<SinkRecord> records, Map<SinkRecord, TopicPartition> originalTopicPartitions) {
        put(records);
    }

Proposed Changes

Add a new method in org.apache.kafka.connect.sink.SinkTask which allows the Connect framework to pass on the original topic partitions (i.e. before topic mutating SMTs have been applied) for records to sink task implementations in order to allow them to do their own offset tracking accurately. The WorkerSinkTask class will also be updated to maintain a new data structure Map<SinkRecord, TopicPartition> containing the original topic partition for each record (i.e. before any potentially topic mutating SMTs are applied) in a message batch. WorkerSinkTask will also be updated to call the new SinkTask::put method instead of the existing one.

Compatibility, Deprecation, and Migration Plan

Backwards Compatibility

This proposal is backward compatible such that existing sink connector implementations will continue to work as before with newer Connect runtimes with this change since the new SinkTask::put method has a default implementation which calls the older SinkTask::put method. This proposal is also backward compatible such that updated connectors using the new functionality from this KIP will continue working with older Connect runtimes because the existing SinkTask::put method is abstract and has to be implemented (the Javadoc will be updated to reflect how this method should be implemented in case the newer SinkTask::put method is overridden - shown in the Public Interfaces section above). While these updated connectors will continue working with older Connect runtimes, they should document that topic mutating SMTs cannot be used with the connector if deployed on older Connect runtimes.

Rejected Alternatives

1. Address the problem entirely within the framework, doing some kind of mapping from the transformed topic back to the original topic.

   • This would only work in the cases where there’s no overlap between the transformed topic names, but would break for the rest of the transformations (e.g. static transformation, topic = “a”).

   • Even if we wanted to limit the support to those cases, it would require considerable bookkeeping to add a validation to verify that the transformation chain adheres to that expectation (and fail fast if it doesn’t).

2. Expose a new method in SinkRecord which returns the original topic and partition
   • This will require connector developers to use a try-catch around the new method calls to ensure that the connector is backward compatible with older Connect runtimes (where the new method won't be available). This is a pretty non-standard approach for most developers and we should try to keep this as a last resort.

3. Expose the entire original record through SinkRecord instead of only topic/partition (e.g. originalSinkRecord)

   • We should not expose the original value/key, transformations might be editing them for security reasons.

4. Create a method in the SinkTaskContext to get this information, instead of updating SinkRecord (e.g. SinkTaskContext::getOriginalTopic(SinkRecord sr) / SinkTaskContext::getOriginalKafkaPartition(SinkRecord sr)

   • This has the same limitations w.r.t. requiring the use of ugly try-catch logic around methods which may or may not exist to ensure backward compatibility of connectors with older Connect runtimes.