This PR was opened by the Changesets release GitHub action. When you're ready to do a release, you can merge this and the packages will be published to npm automatically. If you're not ready to do a release yet, that's fine, whenever you add more changesets to main, this PR will be updated.

Releases

@kafkats/[email protected]

Minor Changes

• fe140c0: ShareConsumer: Kafka 4.2 share groups GA support (ShareFetch/ShareAcknowledge v2)
  • KIP-1206: new acquireMode config ('batch_optimized' default | 'record_limit') — record_limit strictly caps each fetch at maxRecords.
  • KIP-1222: new message.renew() to extend the acquisition lock without finalizing delivery; safe to call multiple times per message.
  • ShareAcknowledgeResponse.acquisitionLockTimeoutMs is now decoded on v2 responses.
  • AckManager dedupes same-offset RENEW + finalizing-ack pairs and collapses duplicate same-offset RENEWs so the wire never carries overlapping batches.
  • @experimental markers removed; share groups are GA in Kafka 4.2.
  • Backward compatible: against Kafka 4.1 brokers the client negotiates v1 and produces byte-identical wire output to before this change. record_limit and renew() throw with a clear "requires v2 (Kafka 4.2+)" error against older brokers.

Patch Changes

• ca7d22f: Correctness and reliability fixes across the producer, consumer, client protocol, and flow state stores.

  Producer

  • flush() now waits for deferred drains so it no longer resolves while writes are still buffered.

  • transactionalSend freezes its partition set up front to avoid double-calling the partitioner.

  • Retry TxnOffsetCommit on retriable and rebalance-in-progress errors instead of failing the transaction.

  • Fence the producer on OUT_OF_ORDER_SEQUENCE_NUMBER to surface the idempotence violation rather than silently corrupting the stream.

  • Reject orphaned in-flight send promises during disconnect so callers are not left hanging.

    Consumer

  • Always reschedule the heartbeat after non-fatal errors, preventing the consumer from dropping out of the group.

  • Surface auto-commit errors to the consumer instead of swallowing them.

  • Clear only the successfully committed partitions on a partial commit failure so the rest are retried.

  • Fence stale fetch responses across partition replacement to avoid delivering records from a revoked assignment.

    Client / protocol

  • Guard against varint shift overflow in the decoder.

  • Close a ConnectionPool.acquire race that could hand out a connection mid-teardown.

  • Clamp the SASL reauthentication delay to the setTimeout maximum.

  • Don't await a response on Produce with acks=0 (the broker sends none).

  • Always use the v0 response header for ApiVersions, matching broker behavior across versions.

    Share consumer

  • Release (rather than acknowledge) the last yielded message on shutdown so unprocessed records are redelivered.

    Flow

  • Write to the local store before the changelog (local-first ordering) to shrink the inconsistency window on crash.

  • Serialize the EOS commit triggered by a rebalance against the next message, and close the EOS rebalance offset-commit gap via an awaitable rebalance hook (shared with the consumer fix above).

  • Use stream time (not wall-clock time) for window expiry.

  • Implement hopping windows (TimeWindows.advanceBy).

  • Throw on SlidingWindows instead of silently misbehaving (not yet implemented).

  • Correct TableGroupByNode mapping ordering.

  • Apply retention to reduce and session aggregators, emitting changelog tombstones for expired state.

  • Surface checkpoint persistence errors instead of swallowing them.

    flow-state-lmdb

  • Make range() upper bound inclusive, matching the in-memory store.

  • Await fsync on checkpoint writes so a checkpoint is durable before it is reported complete.

  • Correct WindowStore time-bound math.

  • Use order-preserving signed-i64 encoding for time keys so range scans iterate in correct chronological order.

@kafkats/[email protected]

Minor Changes

• fe140c0: ShareConsumer: Kafka 4.2 share groups GA support (ShareFetch/ShareAcknowledge v2)
  • KIP-1206: new acquireMode config ('batch_optimized' default | 'record_limit') — record_limit strictly caps each fetch at maxRecords.
  • KIP-1222: new message.renew() to extend the acquisition lock without finalizing delivery; safe to call multiple times per message.
  • ShareAcknowledgeResponse.acquisitionLockTimeoutMs is now decoded on v2 responses.
  • AckManager dedupes same-offset RENEW + finalizing-ack pairs and collapses duplicate same-offset RENEWs so the wire never carries overlapping batches.
  • @experimental markers removed; share groups are GA in Kafka 4.2.
  • Backward compatible: against Kafka 4.1 brokers the client negotiates v1 and produces byte-identical wire output to before this change. record_limit and renew() throw with a clear "requires v2 (Kafka 4.2+)" error against older brokers.

Patch Changes

• ca7d22f: Correctness and reliability fixes across the producer, consumer, client protocol, and flow state stores.

  Producer

  • flush() now waits for deferred drains so it no longer resolves while writes are still buffered.

  • transactionalSend freezes its partition set up front to avoid double-calling the partitioner.

  • Retry TxnOffsetCommit on retriable and rebalance-in-progress errors instead of failing the transaction.

  • Fence the producer on OUT_OF_ORDER_SEQUENCE_NUMBER to surface the idempotence violation rather than silently corrupting the stream.

  • Reject orphaned in-flight send promises during disconnect so callers are not left hanging.

    Consumer

  • Always reschedule the heartbeat after non-fatal errors, preventing the consumer from dropping out of the group.

  • Surface auto-commit errors to the consumer instead of swallowing them.

  • Clear only the successfully committed partitions on a partial commit failure so the rest are retried.

  • Fence stale fetch responses across partition replacement to avoid delivering records from a revoked assignment.

    Client / protocol

  • Guard against varint shift overflow in the decoder.

  • Close a ConnectionPool.acquire race that could hand out a connection mid-teardown.

  • Clamp the SASL reauthentication delay to the setTimeout maximum.

  • Don't await a response on Produce with acks=0 (the broker sends none).

  • Always use the v0 response header for ApiVersions, matching broker behavior across versions.

    Share consumer

  • Release (rather than acknowledge) the last yielded message on shutdown so unprocessed records are redelivered.

    Flow

  • Write to the local store before the changelog (local-first ordering) to shrink the inconsistency window on crash.

  • Serialize the EOS commit triggered by a rebalance against the next message, and close the EOS rebalance offset-commit gap via an awaitable rebalance hook (shared with the consumer fix above).

  • Use stream time (not wall-clock time) for window expiry.

  • Implement hopping windows (TimeWindows.advanceBy).

  • Throw on SlidingWindows instead of silently misbehaving (not yet implemented).

  • Correct TableGroupByNode mapping ordering.

  • Apply retention to reduce and session aggregators, emitting changelog tombstones for expired state.

  • Surface checkpoint persistence errors instead of swallowing them.

    flow-state-lmdb

  • Make range() upper bound inclusive, matching the in-memory store.

  • Await fsync on checkpoint writes so a checkpoint is durable before it is reported complete.

  • Correct WindowStore time-bound math.

  • Use order-preserving signed-i64 encoding for time keys so range scans iterate in correct chronological order.

• Updated dependencies [ca7d22f]

• Updated dependencies [fe140c0]

  • @kafkats/[email protected]

@kafkats/[email protected]

Minor Changes

• fe140c0: ShareConsumer: Kafka 4.2 share groups GA support (ShareFetch/ShareAcknowledge v2)
  • KIP-1206: new acquireMode config ('batch_optimized' default | 'record_limit') — record_limit strictly caps each fetch at maxRecords.
  • KIP-1222: new message.renew() to extend the acquisition lock without finalizing delivery; safe to call multiple times per message.
  • ShareAcknowledgeResponse.acquisitionLockTimeoutMs is now decoded on v2 responses.
  • AckManager dedupes same-offset RENEW + finalizing-ack pairs and collapses duplicate same-offset RENEWs so the wire never carries overlapping batches.
  • @experimental markers removed; share groups are GA in Kafka 4.2.
  • Backward compatible: against Kafka 4.1 brokers the client negotiates v1 and produces byte-identical wire output to before this change. record_limit and renew() throw with a clear "requires v2 (Kafka 4.2+)" error against older brokers.

Patch Changes

• ca7d22f: Correctness and reliability fixes across the producer, consumer, client protocol, and flow state stores.

  Producer

  • flush() now waits for deferred drains so it no longer resolves while writes are still buffered.

  • transactionalSend freezes its partition set up front to avoid double-calling the partitioner.

  • Retry TxnOffsetCommit on retriable and rebalance-in-progress errors instead of failing the transaction.

  • Fence the producer on OUT_OF_ORDER_SEQUENCE_NUMBER to surface the idempotence violation rather than silently corrupting the stream.

  • Reject orphaned in-flight send promises during disconnect so callers are not left hanging.

    Consumer

  • Always reschedule the heartbeat after non-fatal errors, preventing the consumer from dropping out of the group.

  • Surface auto-commit errors to the consumer instead of swallowing them.

  • Clear only the successfully committed partitions on a partial commit failure so the rest are retried.

  • Fence stale fetch responses across partition replacement to avoid delivering records from a revoked assignment.

    Client / protocol

  • Guard against varint shift overflow in the decoder.

  • Close a ConnectionPool.acquire race that could hand out a connection mid-teardown.

  • Clamp the SASL reauthentication delay to the setTimeout maximum.

  • Don't await a response on Produce with acks=0 (the broker sends none).

  • Always use the v0 response header for ApiVersions, matching broker behavior across versions.

    Share consumer

  • Release (rather than acknowledge) the last yielded message on shutdown so unprocessed records are redelivered.

    Flow

  • Write to the local store before the changelog (local-first ordering) to shrink the inconsistency window on crash.

  • Serialize the EOS commit triggered by a rebalance against the next message, and close the EOS rebalance offset-commit gap via an awaitable rebalance hook (shared with the consumer fix above).

  • Use stream time (not wall-clock time) for window expiry.

  • Implement hopping windows (TimeWindows.advanceBy).

  • Throw on SlidingWindows instead of silently misbehaving (not yet implemented).

  • Correct TableGroupByNode mapping ordering.

  • Apply retention to reduce and session aggregators, emitting changelog tombstones for expired state.

  • Surface checkpoint persistence errors instead of swallowing them.

    flow-state-lmdb

  • Make range() upper bound inclusive, matching the in-memory store.

  • Await fsync on checkpoint writes so a checkpoint is durable before it is reported complete.

  • Correct WindowStore time-bound math.

  • Use order-preserving signed-i64 encoding for time keys so range scans iterate in correct chronological order.

• Updated dependencies [ca7d22f]

• Updated dependencies [fe140c0]

  • @kafkats/[email protected]

@kafkats/[email protected]

Minor Changes

• fe140c0: ShareConsumer: Kafka 4.2 share groups GA support (ShareFetch/ShareAcknowledge v2)
  • KIP-1206: new acquireMode config ('batch_optimized' default | 'record_limit') — record_limit strictly caps each fetch at maxRecords.
  • KIP-1222: new message.renew() to extend the acquisition lock without finalizing delivery; safe to call multiple times per message.
  • ShareAcknowledgeResponse.acquisitionLockTimeoutMs is now decoded on v2 responses.
  • AckManager dedupes same-offset RENEW + finalizing-ack pairs and collapses duplicate same-offset RENEWs so the wire never carries overlapping batches.
  • @experimental markers removed; share groups are GA in Kafka 4.2.
  • Backward compatible: against Kafka 4.1 brokers the client negotiates v1 and produces byte-identical wire output to before this change. record_limit and renew() throw with a clear "requires v2 (Kafka 4.2+)" error against older brokers.

Patch Changes

• ca7d22f: Correctness and reliability fixes across the producer, consumer, client protocol, and flow state stores.

  Producer

  • flush() now waits for deferred drains so it no longer resolves while writes are still buffered.

  • transactionalSend freezes its partition set up front to avoid double-calling the partitioner.

  • Retry TxnOffsetCommit on retriable and rebalance-in-progress errors instead of failing the transaction.

  • Fence the producer on OUT_OF_ORDER_SEQUENCE_NUMBER to surface the idempotence violation rather than silently corrupting the stream.

  • Reject orphaned in-flight send promises during disconnect so callers are not left hanging.

    Consumer

  • Always reschedule the heartbeat after non-fatal errors, preventing the consumer from dropping out of the group.

  • Surface auto-commit errors to the consumer instead of swallowing them.

  • Clear only the successfully committed partitions on a partial commit failure so the rest are retried.

  • Fence stale fetch responses across partition replacement to avoid delivering records from a revoked assignment.

    Client / protocol

  • Guard against varint shift overflow in the decoder.

  • Close a ConnectionPool.acquire race that could hand out a connection mid-teardown.

  • Clamp the SASL reauthentication delay to the setTimeout maximum.

  • Don't await a response on Produce with acks=0 (the broker sends none).

  • Always use the v0 response header for ApiVersions, matching broker behavior across versions.

    Share consumer

  • Release (rather than acknowledge) the last yielded message on shutdown so unprocessed records are redelivered.

    Flow

  • Write to the local store before the changelog (local-first ordering) to shrink the inconsistency window on crash.

  • Serialize the EOS commit triggered by a rebalance against the next message, and close the EOS rebalance offset-commit gap via an awaitable rebalance hook (shared with the consumer fix above).

  • Use stream time (not wall-clock time) for window expiry.

  • Implement hopping windows (TimeWindows.advanceBy).

  • Throw on SlidingWindows instead of silently misbehaving (not yet implemented).

  • Correct TableGroupByNode mapping ordering.

  • Apply retention to reduce and session aggregators, emitting changelog tombstones for expired state.

  • Surface checkpoint persistence errors instead of swallowing them.

    flow-state-lmdb

  • Make range() upper bound inclusive, matching the in-memory store.

  • Await fsync on checkpoint writes so a checkpoint is durable before it is reported complete.

  • Correct WindowStore time-bound math.

  • Use order-preserving signed-i64 encoding for time keys so range scans iterate in correct chronological order.

• Updated dependencies [ca7d22f]

• Updated dependencies [fe140c0]

  • @kafkats/[email protected]