Dynamic topology changes

Some changes to a Kafka Streams application are instant. Others require a restart. Some require a full migration procedure. Kafka Streams ties your topology to persistent state, and different kinds of changes carry different costs.

Why changes have different requirements

Kafka Streams ties your topology to persistent state:

• Internal topics (changelog, repartition) are created based on your topology shape
• State stores are written to local disk with specific formats
• Consumer group membership determines who processes which partitions
• Task assignments pin specific work to specific instances

Changes that affect these require care. The framework enforces safety by making expensive changes explicit.

The four tiers of change

Tier	What you can change	How to change it	Examples
Hot	Runtime configuration	Function call on running instance	Worker threads, pause/resume, listeners
Warm	Group membership	Add/remove instances	Scaling horizontally, rolling deploys
Restart	Startup configuration	Restart the process	Processing guarantee, dispatch mode, most config
Migration	Topology fundamentals	Follow a procedure	Change topology shape, application ID, serde

Hot tier: change without stopping

These changes take effect immediately without disturbing processing.

Adjust worker thread count

Add or remove processing threads within your application:

import qualified Kafka.Streams.Runtime as R

R.addStreamThread streams     -- Add a worker
R.removeStreamThread streams  -- Remove a worker

Workers share the same consumer connection. Adding one just reshuffles partition assignments within the process. No broker coordination needed.

Use this when your CPU usage is low and you want more parallelism, or when you want to reduce threads during quiet periods.

Pause and resume processing

Temporarily stop processing without leaving the consumer group:

R.pauseKafkaStreams streams   -- Stop processing
R.resumeKafkaStreams streams  -- Resume processing

The consumer keeps heartbeating, so the broker knows the instance is alive, but no records are polled or processed.

Use when you need exclusive access to state stores for maintenance, when you’re coordinating with an external batch job, or when you want to pause during a dependency outage without triggering rebalances.

Register event listeners

Add or replace listeners for observability:

-- State transitions (RUNNING, REBALANCING, ERROR)
R.setStateListener streams $ \old new ->
  publishLog ("State: " <> show old <> " -> " <> show new)

-- Rebalance events (assignments, revocations)
R.setRebalanceListener streams $ \event ->
  case event of
    Assigned partitions -> warmCaches partitions
    Revoked partitions  -> flushBuffers partitions

Listeners observe; they don’t change processing. Replacing a listener just changes who gets notified.

Use when you need to add monitoring, change alert destinations, or adjust debug logging.

Warm tier: change with one rebalance

These changes require coordination with the consumer group but don’t interrupt processing for long.

Scale by adding instances

Start a new process with the same applicationId:

-- In a new terminal or pod
my-streams-app  -- same config, same applicationId

The new instance joins the consumer group, the group coordinator computes new partition assignments, existing instances release some partitions (incremental rebalance), the new instance picks up its share, and processing resumes. Typically takes seconds; the new instance must replay changelog tails for assigned state stores.

Scale by removing instances

Stop a process cleanly:

R.closeKafkaStreams streams  -- Clean shutdown

The instance leaves the consumer group, remaining instances get reassigned its partitions, and if standbys exist they become active (fast); otherwise replay from changelog (slower). The group must rebalance to redistribute work, which is why this is warm, not hot.

Restart tier: change configuration

These settings are read once at startup and baked into the runtime:

Setting	Why it requires restart
processingGuarantee	Determines whether producer is transactional
dispatchMode	Worker pool routing function baked in
numStandbyReplicas	Standby state machine initialized at startup
commitIntervalMs	Commit cycle scheduler configured at start
stateDir	Local paths opened at initialization

To change: update configuration, drain the instance (stop processing cleanly), restart with new config, rejoin the consumer group.

For rolling restarts, update one instance at a time and wait for it to reach RUNNING state before proceeding. With standbys configured, each restart is a metadata-only promotion.

Migration tier: change fundamentals

These changes affect data layout or identity and require explicit procedures.

Change topology shape

Adding, removing, or renaming operators affects internal topics and state stores. Your old changelog topic contains state in the old format. The new topology expects a different format or different stores entirely.

Procedure:

1. Review the change in topology evolution
2. Run golden-file diff in CI to see what changes
3. For stateful operator changes: drain, deploy, accept warmup cost
4. For renames: old topics become orphans, clean up after verification

Change application ID

The applicationId is your consumer group identity and internal topic prefix. Changing it means a new consumer group (starts from scratch, no offsets), new internal topics created (old ones become orphans), and state rebuilt from upstream (queries return empty until catch-up).

Treat it as a fresh-start deployment. Run old and new in parallel during transition if possible. Clean up old internal topics after confirming the new app is stable.

Change state store serde

Changing how state is serialized affects changelog compatibility.

• Schema-compatible: Both old and new can read each other’s data. Deploy normally.
• Schema-incompatible: Use SchemaVersioned store to migrate reads forward. Or double-write and cut over.

See topology evolution for the full procedure.

What you cannot change (and why)

Some changes aren’t supported because they would undermine correctness.

Truly dynamic topology

You cannot add a new processing branch to a running topology without restarting. New nodes need state stores created atomically across the group, the rebalance protocol would need to version topology shapes, and optimizations assume static topology for correctness.

Alternatives: use TopicNameExtractor for dynamic sink topics (one topology, many destinations), use Branched.withConsumer for runtime-registered handlers (within fixed topology), or deploy a new topology version as a separate application.

Change key-group count live

You cannot scale kgcTotal without restart. Key-group routing is baked into state store sharding, and changing it requires re-partitioning state.

Alternative: drain, wipe local state, restart with new count, warm from changelog.

Summary table

Change	Tier	Downtime	Procedure
Add worker threads	Hot	None	Function call
Pause/resume	Hot	None	Function call
Add listener	Hot	None	Function call
Add instance	Warm	Seconds	Start new process
Remove instance	Warm	Seconds	closeKafkaStreams
Change processing guarantee	Restart	Minutes	Rolling restart
Change commit interval	Restart	Minutes	Rolling restart
Rename operator	Migration	Minutes	Drain + deploy
Change serde	Migration	Minutes	Schema migration or double-write
Change application ID	Migration	Hours	Fresh deploy, parallel run

Related reading

• Topology evolution: Detailed procedures for migration-tier changes
• Scaling: Details on warm-tier scaling
• Running in containers: Configuring for Kubernetes and similar environments