2026.11.0 — 2026-11-01

Breaking

• List a cluster only the views its broker system has (43829167)

  A Kafka cluster showed Artemis-shaped screens: a topology with "Management URL: None set", "Not polled: no management URL" and an "Add a management URL" button, plus routes, request tracing, dead letters, the SQL console, flow, transfers and a Connections page that mean nothing for it or lead to a not-found page. A Kafka cluster is reached by bootstrap servers alone.

  The navigation, the command palette and the g shortcuts now list a view only on a cluster whose system it is for, or whose kinds include the resource class it lists. Row actions, selection bars and metric panels of one system are no longer offered on another. A cluster's address opens the first view its system has: the topology for Artemis, the brokers for Kafka. The command palette no longer carries an Artemis-only view over to a Kafka cluster.

Added

• Support Kafka 2.8 and later, not 3.0 (651039f9)

  A cluster running ZooKeeper-mode Kafka 2.8 was refused: the driver's floor was 3.0, taken from the metadata.version its Admin API reports, and a 2.8 cluster reports none. The cluster itself was fine — SCRAM login succeeded and 66 topics, their offsets, configs, groups and records were readable — so the refusal came from how the release was named, not from anything the cluster lacked.

  Name such a cluster from the API versions its brokers serve, in the v0 shape of ApiVersions that the protocol has fixed since Kafka 0.10 and that Kafka answers before it asks for credentials. Only keys whose release is a protocol fact are used: DescribeProducers (61) arrived in 2.8, ListTransactions (66) in 3.0. No internal Kafka class is involved, which is what ADR-0031 ruled out and ADR-0032 records as superseded in that part.

  A probe that gets no answer leaves the release unknown and does not refuse the cluster; only a readable table lacking both keys is refused as before 2.8. What a 2.8 cluster does not have is declared, never emulated: kafka.transaction-listing is unavailable and the kafka.hanging-transaction kind is not declared on it.

  The release a cluster without metadata.version is named at is a lower bound: the table says where a release starts, not its patch level.

Fixed

• Stop calling a Kafka cluster standalone (43e97a0b)

  The cluster's line said "3 nodes · standalone · reachable" for a three-broker Kafka cluster. Standalone and replication describe Artemis live and backup pairs; a Kafka cluster replicates per partition, so its line now says only how many nodes it has and whether they were reached.

• List a cluster only the views its broker system has (43829167)

  A Kafka cluster showed Artemis-shaped screens: a topology with "Management URL: None set", "Not polled: no management URL" and an "Add a management URL" button, plus routes, request tracing, dead letters, the SQL console, flow, transfers and a Connections page that mean nothing for it or lead to a not-found page. A Kafka cluster is reached by bootstrap servers alone.

  The navigation, the command palette and the g shortcuts now list a view only on a cluster whose system it is for, or whose kinds include the resource class it lists. Row actions, selection bars and metric panels of one system are no longer offered on another. A cluster's address opens the first view its system has: the topology for Artemis, the brokers for Kafka. The command palette no longer carries an Artemis-only view over to a Kafka cluster.

• Keep a cluster's connection options across a reconfiguration (113e8e06)

  Every further call on a registered cluster put it back to the driver's defaults. The gateway kept the bare node list for its change guard but handed the driver the completed spec — the node list plus the cluster's bootstrap endpoints and options — so the two never matched. Each call then took the mismatch branch and reconfigured the connection with the bare spec, which the driver takes wholesale, and the next client was built without the cluster's security settings.

  On a Kafka cluster registered with SASL_PLAINTEXT and SCRAM that meant a plaintext client: the broker closed the connection as soon as it read a Metadata request with no SASL handshake, and the cluster read as unreachable — every call timed out waiting for a node assignment while the brokers answered a raw ApiVersions request from the same container in milliseconds.

  Keep the spec the driver was given, so the guard compares like with like. The same shape was in ProbeSession, which reconfigured with a bare node list too; it now completes the spec as well.

  BrokerGatewayOptionsTest asserts that the security options survive both a reconfiguration after a new node and one after a probed node, and that an unchanged node set is not reconfigured at all.