RATATOSKRATATOSK
ログイン

CRI-O

v1.37.0Kubernetes Core
2026年9月15日

ACTION 2CHECK 4OTHER 32

オペレーター向けの機能と互換性に関わる変更を含み、デフォルトやインターフェースの見直し、正確性と性能の改善が行われています。明示されたセキュリティ修正に加え、依存関係の更新と保守も含まれます。

要対応 (2)

影響確認 (4)

  • breaking非推奨のレジストリー設定とCLIフラグの削除

    insecure_registries を設定している場合、または --insecure-registry を使用している場合に適用されます。

    非推奨だった insecure_registries 設定オプションと --insecure-registry CLI フラグを削除しました。代わりに registries.confinsecure オプションを使用します。

  • breakingチェックポイントとリストアのデフォルト動作を変更

    container_level_enabled を設定していない場合に適用されます。

    checkpoint_restore の設定フィールド container_level_enabled のデフォルトを checkpoint_only に変更し、デフォルトではリストアを無効にしました。

  • breakinggRPCメッセージサイズのデフォルトを縮小

    grpc_max_send_msg_size または grpc_max_recv_msg_size を設定していない場合に適用されます。

    grpc_max_send_msg_sizegrpc_max_recv_msg_size のデフォルトを、従来の 80 MiB から 16 MiB に変更しました。

  • breakingCRIメトリクスの name ラベルを削除

    CRI メトリクスを使用している場合に適用されます。

    CRI メトリクスから name ラベルを削除しました。

その他の記録済み変更 32 件すべてfixes 14 · value changes 7 · additions 6 · constraints 3 · defaults 1 · renames 1

fixes (14)

  • CRI-O now keeps logging blocked syscalls detected by the seccomp notifier instead of only reporting the first event when the notifier action is not "stop".
  • Deduplicate concurrent image volume MountImage calls to avoid containers/storage lock serialization when many pods mount the same image volume simultaneously.
  • Fix CRI-O crash (nil pointer dereference) in ListPodSandboxStats when a container's cgroup is cleaned up during teardown.
  • Fix SELinux labeling so systemd/init containers receive container_init_t again instead of remaining container_t and failing to create cgroups.
  • Fix a bug in high performance hooks where an init container would continue to hold a cpuset that could be used by a regular container in the same pod.
  • Fix a latent crash in pinns caused by a missing zero terminator on the getopt_long long options array.
  • Fix artifact names from externally created OCI layouts (e.g. via skopeo) being incorrectly normalized to wrong docker references.
  • Fix network metrics collection to use the pod's network namespace.
  • Fixed a bug where ImageRef in container status changed from a repo@digest to a raw image ID hash after CRI-O restart.
  • Fixed privileged sandboxes inheriting generate.New()'s default-deny seccomp filter when privileged_seccomp_profile is unset, so they now run unconfined as intended.
  • NRI container lifecycle events now correctly include the container name and pod sandbox ID.
  • Reduced the verbosity of debug logs for List* RPC calls to improve performance
  • Restores pre-1.34 behavior handling envvar values containing non-utf8 data
  • Fix user namespace mappings persistence across CRI-O restarts

value changes (7)

  • Add conmon-rs v1.0.1 support, drop backward compatibility with previous versions.
  • Added namespace, pod, and container labels to CRI metrics
  • Optimized environment variable merging during container creation by eliminating redundant string operations, improving container startup performance.
  • Remove experimental markers from additional_artifact_stores configuration option.
  • Bumped go.podman.io/common to v0.68.0.
  • Because of this change, the Configuration File lookup behavior has changed.
  • Update gomaxprocs hook to ignore workload partitioning when considering whether to inject, as well as update the calculation to ensure containers get at least double the requested number of CPUs, to reduce potential go scheduler throttling.

additions (6)

  • Add enable_cni_status_monitoring config option (default false) and cni_status_grace_period (default 60s) to gate continuous CNI STATUS monitoring and tolerate brief plugin disruptions during upgrades.
  • Add a new config field to allow node admins to only allow checkpointing, and disallowing restore.
  • Added container memory metrics container_memory_active_anon_bytes and container_memory_inactive_anon_bytes (cgroup v1 and v2), and container_memory_anon_thp_bytes, container_memory_shmem_thp_bytes and container_memory_file_thp_bytes (cgroup v2 only).
  • CRI-O now reports runtimeFeatures.UserNamespacesHostNetwork as enabled on Linux and uses a recursively read-only /sys bind mount for containers running with both host network and user namespaces.
  • Support for encrypted container images in Confidential Containers use case: cri-o will now ignore the container image layers, allowing the underlying Confidential Container runtime to manage the image pull management.
  • Support gVisor containerd-shim-runsc-v1

constraints (3)

  • Add btrfs storage driver support to static builds
  • These are gated behind a new memoryExtra value for included_pod_metrics, separate from memory.
  • Added support for per-container targeting of cpu-load-balancing.crio.io, irq-load-balancing.crio.io, and cpu-quota.crio.io annotations.

defaults (1)

  • The default pause image is now registry.k8s.io/pause:3.10.2.

renames (1)

  • Renamed --checkpoint-restore-container-level-enabled flag to --checkpoint-restore-level
CRI-Oをスタックに追加

対応が必要なリリースが出たときに、週次メールでお知らせします。 今回のセキュリティパッチと破壊的変更も、その一例です。

スタックに追加