Karmada
v1.16.8Orchestration & ManagementKarmada v1.16.8 is a patch release concluding the v1.16 line (v1.16.0 through v1.16.8). v1.16.0 introduces multi-component scheduling (feature-gated, off by default), the Webster replica algorithm, and eviction rate limiting; it removes three etcd configuration fields and deprecates two metric labels and a CLI flag. Patch releases v1.16.1–v1.16.8 roll up bugfixes and routine base-image updates.
breakingExternal etcd config fields removed
Three etcd configuration fields—CAData, CertData, and KeyData—have been removed in v1.16.0. If your karmada-operator configuration references these fields, you must remove them before or during the upgrade to v1.16.0.
breakingMetric labels cluster and cluster_name deprecated with v1.18 removal planned
Prometheus metric labels `cluster` and `cluster_name` are deprecated in v1.16.0 and will be removed in v1.18. Queries, alerts, and dashboards relying on these labels must migrate to `member_cluster` before v1.18 ships.
breakingControllerPriorityQueue gate alters workqueue_depth metric shape
When the ControllerPriorityQueue feature gate is enabled, the asyncWorker switches to a priority-queue-based implementation. This changes the processing order of items in the resource detector and splits the `workqueue_depth` metric into multiple series, potentially affecting monitoring queries and alerts that depend on the metric's current cardinality.
breakingFlag --etcd-init-image deprecated
The `--etcd-init-image` flag in the `karmadactl init` command is deprecated in v1.16.0 because it is no longer used and will be removed in a future release. Remove any references to this flag in your initialization workflows.
Key changes (8)
- Multi-component scheduling introduced, gated by MultiplePodTemplatesScheduling (disabled by default), with resource interpreters for FlinkDeployment, SparkApplication, Volcano Job, MPIJob, RayCluster, RayJob, TFJob, and PyTorchJob.
- Replica distribution now uses the Webster (Sainte-Laguë) algorithm for fairer, monotonic assignment across weight strategies.
- Eviction rate limiting added to taint manager via --eviction-rate and --resource-eviction-rate flags, with new failover eviction queue metrics.
- ControllerPriorityQueue expanded to all async-worker-based controllers, reducing downtime after controller-manager restart or leader transition.
- External etcd config fields CAData, CertData, KeyData removed in v1.16.0.
- Metric labels cluster and cluster_name deprecated with removal scheduled for v1.18.
- Default component images bumped: kube-apiserver and kube-controller-manager to v1.34.1, etcd to 3.6.0-0.
- Six stability fixes including Job status aggregation with kube-apiserver v1.32+, member cluster client timeout, metrics-adapter panic, scheduler divide-by-zero, backoff queue sorting, and cluster client transience handling.
A weekly email arrives when a release needs action. Like the breaking changes in this release.