Operations

Every change to cluster topology is an operation. Planning is a dry run: the server computes the steps and, for data moves, the before and after distribution. Nothing changes until you confirm; irreversible operations ask for the cluster name. Steps run one at a time and are stored in PostgreSQL, so an operation continues after a server restart. You can pause, resume or cancel; cleanup steps (for example removing replication throttles) run in every case. Only one operation runs per cluster. Each step and its result are in the audit log.

Safety rules

• A health gate runs before the first step and after every batch or node: no offline partitions, nothing under min.insync.replicas, no under-replicated partitions except those being moved, a quorum leader, no lagging voters.
• Reassignments carry full target replica lists, run in batches, and are throttled on the brokers and topics involved; throttles are always removed afterwards.
• Cordoned and offline brokers never receive replicas; racks are respected.
• One quorum change at a time: a new controller catches up as an observer before it becomes a voter; the last voter is never removed.
• Storage is formatted with the Kafka release that will run the node.

Types

Operation	What it does
Rebalance	Plans moves with the built-in balancer (hard goals: replication factor, racks, capacity, excluded brokers; soft goals: replicas, leaders, data, traffic and per-topic spread), then executes them throttled in batches and elects preferred leaders. Modes: even, replicas only, leaders only.
Move partitions	Your own target replica lists, with the same preview and execution.
Elect preferred leaders	Moves leadership back to the first replica.
Cordon broker	Excludes a broker from new replicas (KIP-1066, Kafka 4.3+).
Drain / decommission broker	Cordons, moves every replica away, then (decommission) stops the node and unregisters the broker.
Unregister broker	Removes a dead broker’s registration.
Add brokers	Provisions brokers on agents or Docker, waits for them, rebalances onto them.
Replace broker	For a lost broker: the same id on an empty disk (followers re-replicate), or a new id with its replicas moved over and the old id unregistered.
Add / remove controller	Grows or shrinks a dynamic quorum one voter at a time.
Migrate quorum	Moves a static quorum to a dynamic one (kraft.version 1, Kafka 4.1+). Irreversible.
Rolling restart	Restarts nodes one at a time, controllers first and the quorum leader last.
Rolling upgrade	Downloads the release on every node, then restarts them one at a time on it. Warns about known defects (KAFKA-20979 in 4.1.0 to 4.3.1) and checks Java 17+.
Finalize upgrade	Raises metadata.version and other features. Irreversible.
Scale node pool	Strimzi broker pools, see clusters.
Create / delete cluster nodes	Docker lab clusters and agent hosts.

The runbooks in docs/runbooks show the same procedures with the Kafka CLI, and test/e2e runs every scenario against the lab.

---

Source: docs/guide/operations.md at v0.1.0.