...

Current state: Accepted

Discussion thread: TBD https://lists.apache.org/list?dev@kafka.apache.org:2020-12:KIP-691

JIRA: 

Related JIRA:

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

Motivation

Kafka Producer supports transactional semantics since 0.11, including the following APIs:

• InitTransaction for transactional producer identity initialization
• beginTransaction to start a new transaction 
• sendOffsetsToTransaction to commit consumer offsets advanced within the current transaction
• commitTransaction commit the ongoing transaction
• abortTransaction abort the ongoing transaction

,

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

Motivation

Kafka Producer supports transactional semantics since 0.11, including the following APIs:

• InitTransaction for transactional producer identity initialization
• beginTransaction to start a new transaction 
• sendOffsetsToTransaction to commit consumer offsets advanced within the current transaction
• commitTransaction commit the ongoing transaction
• abortTransaction abort the ongoing transaction

Within a transaction session, the internal state tracks the last fatal/abortable errorWithin a transaction session, the internal state tracks the last fatal/abortable error. When the Producer hits a fatal/abortable exception, it will transit to the error state, and when next time someone uses a transactional API, it will throw the buffered exception. The caveat is that today we wrap many non-fatal exceptions as KafkaException, which does not make a clear boundary on whether the thrown exception is fatal – should fail fast, or just abortable – should catch and abort the ongoing transaction to resume. It affects the stability of upstream users as well such as Streams EOS. This KIP tries to address this gap by revamping Producer APIs to make it more robust and let exception handling coherent and consistent.

...

Now here's a full summary of all possible exceptions below:

Besides the detailed suggestions on each of the above category lines, there are also a few meta-level proposals:

1. We should have a clear distinguishment between fatal and abortable errors. Today they are all wrapped as KafkaExceptions and hence users cannot really tell the difference, so they have to just treat them all as fatal. One idea is to wrap them differently: we use KafkaException to wrap fatal errors, while use AbortableException to wrap abortable errors.we use KafkaException to wrap fatal errors, while use AbortableException to wrap abortable errors.
2. We should have consistent mechanisms on wrapping errors. More specifically:
   1. All exceptions should be wrapped at most once.
   2. We should not have a nested wrapping like KE(KE(e)). Instead we should just have KE(e).
   3. For IllegalState / Runtime errors which indicate a bug, we should not wrap but directly throw.
3. For errors that can be thrown from both scenario 1) and 2) above, we should have a clear guidance on how EOS users should handle themWe should have consistent mechanisms on wrapping errors. More specifically:
   1. All exceptions should be wrapped at most once.
   2. We should not have a nested wrapping like KE(KE(e)). Instead we should just have KE(e).
   3. For IllegalState / Runtime errors which indicate a bug, we should not wrap but directly throw.
   For errors that can be thrown from both scenario 1) and 2) above, we should have a clear guidance on how EOS users should handle them. More specifically:
   1. Users should be safely ignore the returned future since all errors that should be set inside the TransactionManager as well and hence users should just capture the function itself.
   2. The only error that's thrown only in 2) is TransactionAbortedException, and it could just be ignored since the user asked the producer to close.
4. From EOS-related exceptions thrown directly from the call, we would just check if they are abortable or fatal:
   1. For abortable exception, re-throw them as TxnCorruptedException which will be handled by aborting the txns.
   2. For fatal exception, re-throw them as TaskMigratedException which will cause us to lose all tasks and re-join the group
   3. Users should be safely ignore the returned future since all errors that should be set inside the TransactionManager as well and hence users should just capture the function itself.
   4. The only error that's thrown only in 2) is TransactionAbortedException, and it could just be ignored since the user asked the producer to close.

Proposed Changes

We are proposing a new transactional API usage template which makes EOS processing safer from handling a mix of fatal and non-fatal exceptions:

KafkaConsumer consumer = new KafkaConsumer<>(consumerConfig);
producer.initTransactions();
volatile boolean isRunning = true;

try {
    while (isRunning) {
        ConsumerRecords<String, String> records = consumer.poll(CONSUMER_POLL_TIMEOUT);
        final boolean shouldCommit;
        try {
            producer.beginTransaction();

            // Do some processing and build the records we want to produce.
            List<ProducerRecord> processed = process(consumed);

            for (ProducerRecord record : processed)
                producer.send(record, (metadata, exception) -> {
                    // not required to capture the exception here.
                });
            producer.sendOffsetsToTransaction(consumedOffsets, consumer.groupMetadata());

            shouldCommit = true;
        } catch (Exception e) {
            // Catch any exception thrown from the data transmission phase.
            shouldCommit = false;
        }

        try {
            if (shouldCommit) {
                producer.commitTransaction();
            } else {
                resetToLastCommittedPositions(consumer);
                producer.abortTransaction();
            }
        } catch (CommitFailedException e) {
            // Transaction commit failed with abortable error, user could reset
            // the application state and resume with a new transaction. The root
            // cause was wrapped in the thrown exception.
            resetToLastCommittedPositions(consumer);
            producer.abortTransaction();
        }
    }
} catch (KafkaException e) { // Recommended closing producer/consumer for fatal exceptions
    producer.close();
    consumer.close();
     throw e;
}

In the above example, we separate the transactional processing into two phases: the data transmission phase, and the commit phase. In data transmission phase, any exception thrown would be an indication of the ongoing transaction failure, so that we got a clear signal for the next stage whether to commit or abort the ongoing transaction.

In the commit phase, we should decide whether to commit or abort transaction based on the previous stage result. In new Producer API, commitTransaction() will no longer throw non-fatal exceptions in their raw formats. Instead, it would try to wrap all non-fatal exceptions as `CommitFailedException`. This means any exception other than `CommitFailedException` caught during the commit phase will be definitely fatal, so user's error handling experience could be simplified by just doing a controlled shutdown.

The only debatable case is timeout exception within commit/abort transaction. It could be treated either fatal or not, as strictly speaking producer would have already done the retrying for max.block.ms, so a timeout here may be suggesting a fatal state to a basic user's perspective. Blindly call abortTxn upon timeout could result in illegal state as well when the previous commit already writes `prepare_commit` on the broker side. Usually caller level could have more sophisticated handling to do an application level retry if necessary, but we don't do any recommendations here. It is highly recommended to increase the request timeout here instead of relying on unreliable retries. 

We also put another try-catch block outside of the whole processing loop to get a chance catching all fatal exceptions and close producer and consumer modules. It is a recommended way to handle fatal exceptions when the application still wants to proceed without any memory leak, but is optional to users.

Unify Wrapped KafkaException

As discussed in the motivation section, in KafkaProducer we have a logic to wrap all thrown exceptions as KafkaException. To make the semantic clear and for advanced users such as Kafka Streams to better understand the root cause, we shall no longer wrap any fatal exceptions, but instead only wrap non-fatal ones as KafkaException. We also detect certain cases where we did a double-wrap of KafkaException internally, which will be addressed to ensure only one layer wrapping is supported.

Callback Exception Improvement

As we have seen, there is a callback mechanism in the producer#send which carries the exception type. In EOS setup, it is not required to handle the exception, but for non-EOS cases, the current exception type mechanism is complicated as it throws raw exceptions. Although in the callback function comments we listed all the fatal and non-fatal exceptions, in users' perspective they still need to maintain an exhausting list for checking exception types and take proper actions. The application code is fragile when facing ever-changing underlying producer logic with new exception types, and we have seen the difficulty to classify populated exceptions in application level such as Streams.

On the other thread, there are proposals around making producer API adopt more modern return types than a mere Future. This is a potential good opportunity to merge these two efforts. 

To make the handling easier and consistent, we suggest to add a new exception type called ProduceFailedException which wraps the thrown exception and contains an enum field indicating failure type. Right now there would be 3 types of failures:

1. message rejected: this error associates with this specific record being produced, such as InvalidTopic or RecordTooLarge
2. delivery failed: suggests a failure to produce last record, such as NotEnoughReplicas or Timeout
3. transaction error: an exception relating specifically to transactional semantic, such as ProducerFenced or InvalidTransactionState

With additional flagging, producer users would be in a much better position interpreting the callback and take proper actions with less effort to diagnose the failures by themselves. 

This new exception type would be thrown back to the user only in the new producer API depicted in KIP-706.

Stream side change

For EOS Kafka Streams case, we would adopt these simplified exception throwing logic by catching all exceptions in the data transmission phase to decide for Streams commit. Furthermore, these changes leave to door open for us to analyze the non-fatal exceptions thrown as well by unwrapping KafkaException's cause and reading failure type through callback.

For known exceptions such as ProducerFenced, the handling shall be simplified as we no longer need to wrap them as TaskMigratedException in the send callback, since they should not crash the stream thread if thrown in raw format, once we adopt the new processing model in the send phase.

...

throw e;
}

In the above example, we separate the transactional processing into two phases: the data transmission phase, and the commit phase. In data transmission phase, any exception thrown would be an indication of the ongoing transaction failure, so that we got a clear signal for the next stage whether to commit or abort the ongoing transaction.

In the commit phase, we should decide whether to commit or abort transaction based on the previous stage result. In new Producer API, commitTransaction() will no longer throw non-fatal exceptions in their raw formats. Instead, it would try to wrap all non-fatal exceptions as `CommitFailedException`. This means any exception other than `CommitFailedException` caught during the commit phase will be definitely fatal, so user's error handling experience could be simplified by just doing a controlled shutdown.

The only debatable case is timeout exception within commit/abort transaction. It could be treated either fatal or not, as strictly speaking producer would have already done the retrying for max.block.ms, so a timeout here may be suggesting a fatal state to a basic user's perspective. Blindly call abortTxn upon timeout could result in illegal state as well when the previous commit already writes `prepare_commit` on the broker side. Usually caller level could have more sophisticated handling to do an application level retry if necessary, but we don't do any recommendations here. It is highly recommended to increase the request timeout here instead of relying on unreliable retries. 

We also put another try-catch block outside of the whole processing loop to get a chance catching all fatal exceptions and close producer and consumer modules. It is a recommended way to handle fatal exceptions when the application still wants to proceed without any memory leak, but is optional to users.

Unify Wrapped KafkaException

As discussed in the motivation section, in KafkaProducer we have a logic to wrap all thrown exceptions as KafkaException. To make the semantic clear and for advanced users such as Kafka Streams to better understand the root cause, we shall no longer wrap any fatal exceptions, but instead only wrap non-fatal ones as KafkaException. We also detect certain cases where we did a double-wrap of KafkaException internally, which will be addressed to ensure only one layer wrapping is supported.

Callback Exception Improvement

As we have seen, there is a callback mechanism in the producer#send which carries the exception type. In EOS setup, it is not required to handle the exception, but for non-EOS cases, the current exception type mechanism is complicated as it throws raw exceptions. Although in the callback function comments we listed all the fatal and non-fatal exceptions, in users' perspective they still need to maintain an exhausting list for checking exception types and take proper actions. The application code is fragile when facing ever-changing underlying producer logic with new exception types, and we have seen the difficulty to classify populated exceptions in application level such as Streams.

On the other thread, there are proposals around making producer API adopt more modern return types than a mere Future. This is a potential good opportunity to merge these two efforts. 

To make the handling easier and consistent, we suggest to add a new exception type called ProduceFailedException which wraps the thrown exception and contains an enum field indicating failure type. Right now there would be 3 types of failures:

1. message rejected: this error associates with this specific record being produced, such as InvalidTopic or RecordTooLarge
2. delivery failed: suggests a failure to produce last record, such as NotEnoughReplicas or Timeout
3. transaction error: an exception relating specifically to transactional semantic, such as ProducerFenced or InvalidTransactionState

With additional flagging, producer users would be in a much better position interpreting the callback and take proper actions with less effort to diagnose the failures by themselves. 

This new exception type would be thrown back to the user only in the new producer API depicted in KIP-706.

Streams Side Change

For EOS Kafka Streams case, we would adopt these simplified exception throwing logic by catching all exceptions in the data transmission phase to decide for Streams commit. Furthermore, these changes leave to door open for us to analyze the non-fatal exceptions thrown as well by unwrapping KafkaException's cause and reading failure type through callback.

More specifically:

• For known exceptions such as ProducerFenced, the handling shall be simplified as we no longer need to wrap them as TaskMigratedException in the send callback, since they should not crash the stream thread if thrown in raw format, once we adopt the new processing model in the send phase.

• When handling lost-all-partitions, which would trigger when 1) the rebalance listener's onPartitionsLost are called, indicating the consumer member has been kicked out of the group, 2) a task-migration exception is thrown, we should not need to reset the producer by closing the current one and re-creating a new producer any more. Instead, we should still be able to reuse the same producer after we've re-joined the consumer group. Instead we just need to re-`initTxn` on the producer to make sure the previous dandling txns have been aborted before new transactions are about to start.
• We should distinguish exceptions thrown from the send() callback v.s. from the send() / commit() / etc call directly. With EOS, the only exception that would ONLY be thrown in the callback would be TransactionAbortedException, which we can actually ignore; and hence we would only need to just capture all exceptions thrown from the calls directly. That means we would handle exceptions differently between ALOS and EOS:
  • ALOS: try to capture exceptions from the callback, handle them just as today.
  • EOS: ignore exceptions from the callback, instead handle directly from the function calls.

Public Interfaces

As mentioned in the proposed changes section, we would be doing the following public API changes:

...

The failure type in the embed exception should be helpful, for example, they could choose to ignore transactional errors since other txn APIs are already taking care of them. We expect this error code to be implemented once KIP-706 is accepted, which would provide a more user-friendly send API with CompletableFuture or similar.

Documentation change

We shall put the newly marked fatal exceptions on the public Producer API docs correspondingly, including 

...