This page is meant as a template for writing a KIP. To create a KIP choose Tools->Copy on this page and modify with your content and replace the heading with the next KIP number and a description of your issue. Replace anything in italics with your own description.

Status

Current state: Under Discussion

Discussion thread: here 

JIRA: [KAFKA-19883]( KAFKA-19883 - Getting issue details... STATUS )

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

Motivation

Ensure atomicity between share‑group acknowledgments and downstream side effects (e.g., Kafka transactions) so a record is either both processed and acknowledged or neither.

1. Share consumer polls records -> broker sets them to ACQUIRED.
2. Framework processes and acknowledges records (implicit or explicit).
3. Checkpoint fails before sink outputs are committed.
4. Records are permanently ACKNOWLEDGED (not redelivered) -> data loss.
5. Note:
   1. Share records become terminal when acked: RecordState
   2. Ack path today is irreversible once ACKNOWLEDGED: SharePartition.acknowledge()

Goal: Enable exactly-once read semantics via transactional acknowledgements.

We will be following the similar existing pattern we have in Kafka Producer Transactions.

Use Cases

• Consume‑Transform‑Produce (CTP): bind acks to Kafka output transaction.
• Consume‑only frameworks: transactional acks independent of producer.

Frameworks:

- Apache Flink: Exactly-once checkpointing with Share Group sources
- Apache Spark: Structured Streaming with Share Group consumers
- Any coordinator-worker streaming framework requiring atomic acknowledgements

Public Interfaces

APIs 

• KafkaProducer.sendShareAcksToTransaction(acks, groupId) - for CTP 
  • This mirrors the existing KafkaProducer.sendOffsetsToTransaction(offsets, groupMetadata)

```

// When you have a producer and want atomic output + acks

producer.beginTransaction();

producer.send(output);

producer.sendShareAcksToTransaction(acks, groupId);

producer.commitTransaction();

```

• TransactionalShareAcknowledger - for standalone ack transactions

```

// When there's no producer (Flink source, Spark source, pure consumer)

public class TransactionalShareAcknowledger implements Closeable {

    public TransactionalShareAcknowledger(Map<String, Object> config);

    public void initTransactions();

    public void commitAcknowledgements(

        Map<TopicPartition, List<ShareAcknowledgement>> acks, String groupId);

    public void abortAcknowledgements();

    public void close();

}

```

Internally, TransactionalShareAcknowledger is a thin wrapper around a KafkaProducer (or its TransactionManager). It use the exact same RPCs - InitProducerId, AddShareAcksToTxn, TxnShareAcknowledge, EndTxn.

No new server-side infrastructure needed.

```

// TransactionalShareAcknowledger — internally just wraps a KafkaProducer 
TransactionalShareAcknowledger acknowledger = 
    new TransactionalShareAcknowledger(config);  // config has transactional.id
acknowledger.initTransactions();
acknowledger.commitAcknowledgements(acks, shareGroupId);  // begin+ack+commit in one call

```

Coordinator API

• Reuses the existing TransactionCoordinator
• new server-side component is a completeTransaction() method on ShareCoordinator, mirroring GroupCoordinator.completeTransaction().
• WriteTxnMarkers as the mechanism for the transaction coordinator to tell the group coordinator to complete transactional operations on __consumer_offsets
  • For consumer group we have => groupCoordinator.completeTransaction(partition, ...)
  • Similarly implement shareCoordinator.completeTransaction(partition, ...) for share group 

New Metrics

Metric                                Type        Description
share-transaction-active              Gauge       Active transactions
share-transaction-prepare-time-ms     Histogram   Prepare latency
share-transaction-commit-time-ms      Histogram   Commit latency
share-transaction-abort-total         Counter     Aborted transactions
share-transaction-timeout-total       Counter     Timed-out transactions

Proposed Changes

1. Two-Phase Commit Protocol

Transaction States Transitions

existing Kafka 2PC (TransactionState: ONGOING→PREPARE_COMMIT→COMPLETE_COMMIT)

- EMPTY -> ONGOING -> PREPARE -> ABORTED (abort after prepare)

- EMPTY -> ONGOING -> ABORTED (early abort)

Sequence Diagram

Abort case: 

This KIP reuses Kafka's existing two-phase commit protocol:

• EndTxn(COMMIT) transitions the transaction to PREPARE_COMMIT in __transaction_state
• WriteTxnMarkers dispatches commit markers to all partitions (data topics, __consumer_offsets, and now __share_group_state)
• When all markers are confirmed, state transitions to COMPLETE_COMMIT

Transaction metadata stays in __transaction_state.

Share acks are written transactionally to __share_group_state using CoordinatorRuntime.scheduleTransactionalWriteOperation().

Failure Handling

existing TransactionCoordinator recovery

2. Wire Protocol Details

TxnShareAcknowledgeRequest

```
ShareAcknowledgeTransactionalRequest => GroupId TransactionId [Topics]
  GroupId => STRING
  TransactionId => STRING
  Topics => TopicId [Partitions]
    Partitions => Partition [AcknowledgementBatches]
      AcknowledgementBatches => FirstOffset LastOffset AcknowledgeType
```

ShareBeginTransactionRequest

Use existing InitProducerIdRequest

SharePrepareTransactionRequest

 Use existing EndTxnRequest

ShareCommitTransactionRequest / ShareAbortTransactionRequest

Handled by existing WriteTxnMarkersRequest (inter-broker) and  EndTxnRequest(ABORT)

AddShareAcksToTxnRequest

• mirrors AddOffsetsToTxnRequest; This tells the TransactionCoordinator to add __share_group_state partitions to the producer's transaction.
• AddShareAcksToTxn computes TopicPartition(SHARE_GROUP_STATE_TOPIC_NAME, shareCoordinator.partitionFor(shareGroupId, topicPartition)) and  adds it to the producer's TransactionMetadata.topicPartitions set via TransactionCoordinator.handleAddPartitionsToTransaction().
• This is what causes WriteTxnMarkers to dispatch markers to __share_group_state partitions during commit.

```

AddShareAcksToTxnRequest => TransactionalId ProducerId ProducerEpoch GroupId [Topics]

  TransactionalId => STRING

  ProducerId => INT64

  ProducerEpoch => INT16

  GroupId => STRING

  Topics => TopicName [Partitions]

    Partitions => INT32

```

Consume-Transform-Produce Pattern

This is the exactly-once guarantee: either both output AND acks commit, or neither does: 

```

KafkaShareConsumer<K, V> consumer = new KafkaShareConsumer<>(props);  // NOT transactional
KafkaProducer<K, V> producer = new KafkaProducer<>(props);            // transactional.id

producer.initTransactions();

while (true) {
    ConsumerRecords<K, V> records = consumer.poll(...);  // ← just reads, not transactional
    
    producer.beginTransaction();
    for (var record : records) {
        producer.send(new ProducerRecord<>("output", transform(record)));
    }
    producer.sendShareAcksToTransaction(buildAcks(records), "my-share-group");  //  transactional write
    producer.commitTransaction();
    // Atomically: output records committed + share acks committed
    // If abort: output rolled back + acks discarded → records return to AVAILABLE
}

```

In the existing consumer group CTP pattern, the same thing happens (in the above code instead of KafkaShareConsumer, use KafkaConsumer and Consumer offsets to __consumer_offsets).

If abortTransaction() is called (or the transaction times out):

• Output records get ABORT markers -> consumers with read_committed skip them
• Share acks are discarded -> records stay ACQUIRED -> acquisition lock timeout -> records return to AVAILABLE -> re-delivered to another consumer

no-producer (Flink/Spark) use case 

```

// Flink ShareGroupSource connector — on checkpoint complete

class FlinkShareGroupSourceReader implements SourceReader<...> {

    private TransactionalShareAcknowledger acknowledger;

    private List<ShareAcknowledgement> pendingAcks;

    void open() {

        acknowledger = new TransactionalShareAcknowledger(config);

        acknowledger.initTransactions();

    }

    void snapshotState(long checkpointId) {

        // Save pending acks to checkpoint state

    }

    void notifyCheckpointComplete(long checkpointId) {

        // Atomically commit acks for this checkpoint

        acknowledger.commitAcknowledgements(pendingAcks, shareGroupId);

    }

}

```

Compatibility, Deprecation, and Migration Plan

Impact on Existing Users

• If older broker doesn’t support TxnShareAcknowledgeRequest, fallback to current explicit/implicit ack with warning.
• Clients must fail fast on unsupported brokers
• No breaking changes for existing Share Group users
• `transactional` mode is opt-in via `share.acknowledgement.mode` config
• Existing `implicit` and `explicit` modes continue to work unchanged

Migration Path

1. Phase 1: Add `transactional` acknowledgement mode (backward compatible)
2. Phase 2: Streaming frameworks (Flink, Spark) implement transactional sources
3. Phase 3: Documentation and best practices for exactly-once semantics

Deprecation

- No deprecation of existing modes planned
- `transactional` mode recommended for exactly-once use cases

Test Plan

Unit Tests

- Transaction state machine transitions
- Idempotency of all transaction operations
- Timeout handling and auto-abort
- Record state transitions (ACQUIRED → ACKNOWLEDGED on commit, ACQUIRED → AVAILABLE on abort/timeout)

Integration Tests

- End-to-end transaction commit/abort flows
- Multi-consumer transactions within same group
- Coordinator failure and recovery
- Network partition scenarios

System Tests

- Flink checkpoint integration with transactional Share Groups
- Exactly-once delivery verification under failures
- Performance benchmarks comparing transactional vs non-transactional modes
- Stress testing with concurrent transactions

Chaos Tests

- Broker failure during PREPARE state
- Coordinator crash before/after prepare
- Consumer crash during transaction
- Network partitions between coordinator and brokers

Include exactly‑once recovery tests similar to Flink e2e:

• • Kill after sendShareAcksToTransaction but before commit.
  • Kill after commit but before notifyCheckpointComplete.
  • Ensure acks are either replayed or visible, never lost.

Rejected Alternatives

If there are alternative ways of accomplishing the same thing, what were they? The purpose of this section is to motivate why the design is the way it is and not some other way.