Page History

External Connect workers (separate JVM)

Integrated into Kafka brokers using native replication protocol

Manage Connect cluster lifecycle independently

Unified with broker operations

Separate Connect metrics and dashboards

Standard Kafka JMX metrics

Offset Translation

Lossy, requires remapping, causes reprocessing overhead

None needed - offsets preserved exactly

Metadata Sync

Requires separate connector configuration (MirrorSourceConnector, MirrorCheckpointConnector)

Automatic (topics, configs, consumer groups, ACLs)

Transactional Topics

Markers copied as regular records, incomplete transactions possible during replication

Markers mirrored, LSO truncation ensures consistency before failover

Topic Write Protection

Not supported - mirrored topics always writable

Read-only enforcement during mirroring, writable only after explicit failover

Tiered Storage

Fetches from broker (which reads from remote storage)

Not initially supported (future work)

Active-Active

Supported via topic prefixing and cycle detection

Not supported (read-only enforcement prevents cycles)

Share Groups

Not supported

Failback

Full re-mirror from offset 0

Delta sync using LastMirroredOffset

Topic Name Preservation

No - destination topics prefixed with source cluster alias (e.g., source.topic-name)

Yes - same topic name as source

Topic ID Preservation

No - destination gets new topic ID

Yes - same topic ID as source

Compression Overhead

Decompress + recompress records

Preserve source compression (zero overhead)

Failover Time

Offset translation + consumer group sync + bootstrap reconfiguration

Consumer group sync + bootstrap reconfiguration only

Bandwidth Control (Source)

None (unless client quotas manually configured)

Client quotas (current), dedicated mirror throttling (future)

Bandwidth Control (Destination)

None (unless client quotas manually)

Replica-level throttling (mirror.replication.throttled.rate)

Resource Isolation (Source)

Shares network bandwidth and disk I/O with all consumers

Shares network bandwidth and disk I/O with all consumers (same - current), Isolated on replica-level (future)

Resource Isolation (Destination

Separate JVM heap (memory isolated from brokers) but still consume resources from the broker as any client

Shares network, disk I/O; dedicated thread pool within broker JVM

Malformed Batch Handling

Crashes Connect task, all partitions in task restart

Fails individual partition (FAILED state), others continue

Blast Radius(Failure)

All partitions in task affected

Single partition affected

Catch-Up Surge Protection

Connect worker heap may OOM, affects all tasks

Throttling on destination replica + source quota prevent memory spikes

Unclean Leader Election

No protection against log divergence

Explicitly unsupported with warning logs