Status

Current state: Under Voting 

Discussion thread: here

JIRA: here 

Vote thread: here

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

Motivation

Kafka Producer supports following transactional APIs

1. initTransactions for transactional producer identity initialization

2. beginTransaction to start a new transaction

3. sendOffsetsToTransaction to commit consumer offsets advanced within the current transaction

4. commitTransaction commit the ongoing transaction

5. abortTransaction abort the ongoing transaction

In addition to above transactional APIs, idempotent producer sends records to broker using send API

Apache Kafka supports a variety of client SDKs for different programming languages:

1. Java: The official Java client library supports the producer, consumer, Streams, and Connect APIs.

2. librdkafka and derived clients:

   1. C/C++: A C/C++ client library supporting the Producer and Consumer APIs.

   2. Python: A Python client library supporting the Producer and Consumer APIs.

   3. Go: A Go client library supporting the Producer and Consumer APIs.

   4. .NET: A .NET client library supporting the Producer and Consumer APIs.

We require consistent error handling across all clients SDKs and APIs. Functionally the client can handle error by:

• Retry the operation without bothering the application ()

• Share with the application that an error has occurred, but the client can recover if the transaction is aborted by the client and under the covers, the epoch is bumped (abortable)

• Share with the application that an error has occurred, and there is no way to recover except shut down the client (fatal)

In this KIP, we want to accomplish a few goals:

• Simplify error handling, make it easy for other clients to implement protocols

• Give clarity in error messages

• Make the expected errors clear: [KAFKA-14439] Specify returned errors for various APIs and versions - ASF JIRA, [KAFKA-7787] Add error specifications to KAFKA-7609 - ASF JIRA

We have attempted error handing in accepted KIP: KIP-691: Enhance Transactional Producer Exception Handling - Apache Kafka - Apache Software Foundation which we can perhaps leverage some of the work here.

Proposed Changes

We are proposing to group exceptions into four types:

1. Producer-Retriable: mostly maps onto current retriable errors. The producer can handle the retry on its own and the failure is invisible to the application. Some of these errors may indicate that a metadata update at the producer level is required. Because of that we introduce a few subclasses to the retriable error

   1. retry only (just send the request again – after some period of backoff)

   2. refresh metadata and retry (request metadata and maybe modify the request before resending)

2. Producer-Abortable: mostly maps onto abortable errors. The error is bubbled to the application layer, and it can choose to roll back any state from the ongoing transaction and abort the transaction. The producer does not need to restart, as after aborting, it can be confident that the state was as it was before the transaction started. The application can also close the producer and react as it does for application-recoverable cases, but it doesn’t need to.

3. Application-Recoverable: maps on to some fatal errors. The error is bubbled to the application layer, and it may need to do a bit more to roll back and clean up state. The producer can not simply abort and know the current state of the partitions. Applications can handle this in different ways – streams may rebalance a task and/or close the task and restart it. Another application may read from their own checkpoint to continue. In any case, the producer must restart and will be unusable after encountering this error.

4. Invalid-Configuration: maps to some fatal errors. The error is bubbled up to the application layer. The application can decide what to do. The producer doesn’t need to restart, but the application may chose to close it.

Each error code always represent the same class and rarely rely on client state to determine how to handle.

While it is good to have a mapping, I think it is also useful to have a general strategy – ie a typical unknown error (not specified by the client to have a type) should probably be application recoverable.

Exception Table 

Below table contains list of exceptions with current and New handing. All exceptions are grouped under a header which represent Exception Name:

Yellow → Needs modification

Green → New exceptions added as part of this KIP

Check → Exception handling match as mentioned in KIP-691

Cross → Exception handling doesn't match as mentioned in KIP-691

AbortableTransactionException

ProducerRetriableTransactionException

ProducerRefreshRetriableTransactionException

ApplicationRecoverableTransactionException

InvalidConfiguationTransactionException

InvalidTxnStateException requires special handling where producer API and Transactional API has different handling:

Public Interfaces

We will add new exception types, as listed in the table below, that extend the existing exceptions.

// Producer Abortable Transaction
public class AbortableTransactionException extends ApiException {
    ...
}

//Producer Retriable 
public class ProducerRetriableTransactionException extends ApiException {
 	...
}

//Producer Refresh and Retriable
public class ProducerRefreshRetriableTransactionException extends ApiException {
	...
}

//Application-Recoverable
public class ApplicationRecoverableTransactionException extends ApiException {
    ...
}

// Invalid-Configuration
public class InvalidConfiguationTransactionException extends ApiException {
 	...
}

Client side code example

public class TransactionalClientDemo {

    private static final String CONSUMER_GROUP_ID = "my-group-id";
    private static final String OUTPUT_TOPIC = "output";
    private static final String INPUT_TOPIC = "input";
    private static KafkaConsumer<String, String> consumer;
    private static KafkaProducer<String, String> producer;

    public static void main(String[] args) {
        initializeApplication();

        boolean isRunning = true;
        // Continuously poll for records
        while (isRunning) {
            try {
                try {
                    // Poll records from Kafka for a timeout of 60 seconds
                    ConsumerRecords<String, String> records = consumer.poll(ofSeconds(60));

                    // Process records to generate word count map
                    Map<String, Integer> wordCountMap = new HashMap<>();

                    for (ConsumerRecord<String, String> record : records) {
                        String[] words = record.value().split(" ");
                        for (String word : words) {
                            wordCountMap.merge(word, 1, Integer::sum);
                        }
                    }

                    // Begin transaction
                    producer.beginTransaction();

                    // Produce word count results to output topic
                    wordCountMap.forEach((key, value) ->
                            producer.send(new ProducerRecord<>(OUTPUT_TOPIC, key, value.toString())));

                    // Determine offsets to commit
                    Map<TopicPartition, OffsetAndMetadata> offsetsToCommit = new HashMap<>();
                    for (TopicPartition partition : records.partitions()) {
                        List<ConsumerRecord<String, String>> partitionedRecords = records.records(partition);
                        long offset = partitionedRecords.get(partitionedRecords.size() - 1).offset();
                        offsetsToCommit.put(partition, new OffsetAndMetadata(offset + 1));
                    }

                    // Send offsets to transaction for atomic commit
                    producer.sendOffsetsToTransaction(offsetsToCommit, CONSUMER_GROUP_ID);

                    // Commit transaction
                    producer.commitTransaction();
                } catch (AbortableTransactionException e) {
                    // Abortable Exception: Handle Kafka exception by aborting transaction. producer.abortTransaction() should not throw abortable exception.
                    producer.abortTransaction();
                    resetToLastCommittedPositions(consumer);
                }
            } catch (InvalidConfiguationTransactionException e) {
                //  Fatal Error: The error is bubbled up to the application layer. The application can decide what to do
                closeAll();
                throw e;
            } catch (KafkaException | ApplicationRecoverableTransactionException e) {
                // Application Recoverable: The application must restart
                closeAll();
                initializeApplication();
            }
        }

    }

Full example can be accessed at: https://github.com/apache/kafka/pull/15913/files 

Compatibility, Deprecation, and Migration Plan

This KIP involves client side changes which only affects the resiliency of new Producer client and Streams. New clients have the flexibility to adjust their exception handling logic based on the specific type of exception encountered, enabling them to take appropriate actions as described in the Clientsidecodeexample section. Existing clients need not make any modifications, as they can continue to utilize their current error handling mechanisms.

Test Plan

We will add additional integration and unit tests to check if client acts as expected.

Rejected Alternatives

We discussed another approach where we encode the handling into the response as a separate field from the error code or something similar. Produce responses have space for a code and for a message, so something similar could be done to distinguish the cause/meaning of the error from the action that should be taken. The encoding of the error could be included in the definition of the response spec for other client compatibility. We rejected this approach as it requires a larger overhaul on the APIs as well as the clients.