Karmada
v1.15.7Orchestration & ManagementKarmada v1.15.7 patches three bugs: a cert rotation deadlock in the agent, a Helm chart rendering failure on upgrade, and a race condition that silently dropped graceful eviction tasks.
breakingCheck if cert rotation has been silently failing in your agents
If your karmada-agent has been running for a while without certificate renewal, the SignerName mismatch bug means CSRs were never approved and certs may be expired or near expiry. After upgrading to v1.15.7, verify agent certificate validity and manually trigger rotation if needed. This is especially urgent in long-running clusters where cert TTLs are short.
securityAudit agent certificates for expiry before upgrading
The cert rotation bug could have left agents with expired certificates, creating an outage risk or a gap where mTLS wasn't enforced. Before upgrading, check the NotAfter field on agent certs. Post-upgrade, confirm the cert_rotation_controller and agent_csr_approving are now operating with matching SignerNames and that pending CSRs get approved.
breakingSilent eviction failures may have left workloads on bad clusters — audit now
The race condition in graceful eviction means any cluster that was tainted or marked unhealthy while multiple controllers were active may not have had its workloads evacuated. Before upgrading, check ResourceBinding and ClusterResourceBinding objects for stale taints or pending eviction conditions that were never resolved. Post-upgrade, re-trigger eviction for any affected bindings.
Key changes (3)
- karmada-agent: Certificate rotation CSRs now auto-approve correctly — a SignerName mismatch between cert_rotation_controller and agent_csr_approving was blocking all cert renewals.
- karmada-chart: Helm upgrades no longer leave {{ ca_crt }} as a literal unrendered string, which would break TLS config silently.
- karmada-controller-manager: Race condition fixed where concurrent modifications to the same ResourceBinding or ClusterResourceBinding caused graceful eviction tasks to be silently dropped, leaving workloads stranded on tainted or failing clusters.