Redpanda Release Notes

A broker no longer aborts with an out-of-memory error after forwarding controller join and snapshot requests to itself following a controller-leader step-down.

Cluster membership no longer locks when a broker is decommissioned while it is being added as a controller (raft0) learner.

Brokers no longer report the old leader of a partition after its topic is deleted and created again. Previously, some Kafka clients then stopped producing to that partition.

DNS lookups no longer hang indefinitely while waiting for the resolver mutex.

A rare crash no longer occurs on ARM (aarch64) when the CPU profiler or stall detector samples a thread in the middle of a stack switch, for example while collecting a debug bundle.

Iceberg producer throttling now uses max_kafka_throttle_delay_ms as the throttle ceiling and quota_manager_gc_sec as the producer GC window. These two settings were previously flipped at the point of use.

Iceberg translation now supports backing the UUID logical type with the fixed(16) physical type in Avro schemas.

Iceberg commits now correctly count status=deleted manifest entries when they create a new manifest list for a snapshot. Previously, the catalog could reject commits when another writer to the table, such as a compaction process, deleted data files and wrote manifests with status=deleted entries.

The Kafka quota manager’s garbage collection no longer causes a use-after-free during shutdown.

The internal Kafka client (used by Schema Registry, HTTP Proxy, and Shadow Link) no longer leaves a broker connection permanently unable to complete API-version negotiation, which previously caused repeated broker_not_available failures until the socket dropped for an unrelated reason.

The internal Kafka client no longer desyncs a connection’s protocol framing when a response arrives after its request has timed out; previously this caused cascading disconnects and stalled requests on that connection.

Schema Registry no longer aborts the broker when a request fails before its deferred authorization check with schema_registry_enable_authorization enabled. Such requests now return an error response.

Checking JSON Schema compatibility when a required property has a boolean subschema no longer aborts the broker.

JSON Schema compatibility checks for schemas that use fragment-only or fragment-bearing $id values no longer resolve `$ref`s against the wrong subschema. Compatibility verdicts for affected schemas may change.

JSON Schema compatibility checks no longer crash when a schema’s $ref targets a boolean or a non-schema value.

Produce requests are now rejected if any schema-ID-validated key or value in the batch is invalid. Previously, a batch could be accepted when its last validated component was valid, persisting the invalid records.

Schema Registry no longer returns HTTP 500 on its first request after a broker restart while controller leadership is still being established.

Describing or deleting ACLs on clusters with many distinct ACL resource patterns no longer makes oversized memory allocations.

OpenSSL is upgraded to 3.0.22, which addresses CVE-2026-75803, CVE-2026-63072, CVE-2026-63076, and CVE-2026-63074.

c-ares is upgraded to 1.34.8, which addresses CVE-2026-33630.

Broker restarts after a crash are no longer slower than normal. Previously, on every later restart, the broker re-read the segment that each partition was writing at the time of the crash, and logged Recovered: for it. The first restart after upgrading fixes these segments.

Changing cloud_storage_segment_max_upload_interval_sec now reports that a broker restart is required. Previously, the change was accepted with no restart prompt, but brokers kept uploading on the old interval until they were restarted.

Reconnection is more reliable when inter-broker heartbeat requests time out.

Redpanda logs a warning when a DNS lookup is stuck in the resolver.

Schema Registry now logs whether a 403 response on GET /schemas/ids/{id} comes from a schema ID that does not exist or from missing ACLs. The response to clients is unchanged.

A new redpanda_schema_registry_loaded_offset metric surfaces the last offset from the internal _schemas topic applied to the local store.