ADR-0215: Cluster health is formed per system, on one scale

• Status: accepted
• Date: 2026-10-28
• Deciders: Mahdi Amirabdollahi

Context

HaStateEvaluator forms every cluster's verdict from Artemis HA pairs: live nodes, split brain, replication behind. A Kafka cluster has no pairs; its health is its brokers serving, its partitions led and in sync, and its controller quorum's majority, and a ZooKeeper-mode cluster's quorum cannot be read at all. The estate must still compare clusters of both on one scale.

Decision

A ClusterHealthEvaluator system extension (ADR-0208) forms the verdict from the driver's node conditions, on the shared scale OK, DEGRADED and CRITICAL, with every cause in notes. Artemis wraps HaStateEvaluator unchanged. Kafka is CRITICAL on offline partitions, a lost quorum majority or no broker reachable, and DEGRADED on a fenced or unreachable broker, under-replicated partitions or a lagging voter. What cannot be read is stated in a note and never counted as healthy.

Consequences

• A new system brings its own evaluator; the estate needs no change.
• A ZooKeeper-mode Kafka cluster can be OK with a note that its quorum is not visible.

Alternatives considered

• One generic evaluator over conditions: loses Artemis's split-brain rule, which is critical and system-specific.