4671-gang-scheduling |
scheduling |
Landmark |
implementable / beta |
v1.35 / v1.37 / v1.38 |
GenericWorkload |
引入 Workload / PodGroup,scheduler 以组为单位做准入、permit、binding 和失败回退。 |
5710-workload-aware-preemption |
scheduling |
Landmark |
implementable / beta |
v1.36 / v1.37 / v1.39 |
GenericWorkload |
抢占从 Pod 级扩展到 PodGroup / workload 级,避免抢占后仍不能整体运行。 |
6012-composite-podgroup-api |
scheduling |
Important |
implementable / alpha |
v1.37 / v1.38 / v1.40 |
CompositePodGroup |
表达多组件 workload 的组合 PodGroup,是 JobSet/LWS/训练任务接入 gang 语义的 API 补充。 |
6089-was-controller-apis |
scheduling |
Important |
implementable / alpha |
v1.37 / v1.38 / v1.39 |
WorkloadWithJob |
给 controller 提供 workload-aware scheduling building blocks,而不是强制所有 workload 使用同一顶层 API。 |
6132-prequeueing-hints |
scheduling |
Important |
implementable / beta |
- / v1.37 / v1.39 |
SchedulerPreQueueingHints |
在 Pod 入 activeQ 前判断事件是否值得唤醒,降低 DRA/gang/resize 等等待型场景的无效重试。 |
5598-opportunistic-batching |
scheduling |
Important |
implementable / beta |
- / v1.35 / v1.38 |
OpportunisticBatching |
scheduler 批量处理可调度机会,提高大规模集群吞吐。 |
4381-dra-structured-parameters |
node + scheduling + autoscaling |
Landmark |
implemented / stable |
v1.30 / v1.32 / v1.34 |
DynamicResourceAllocation, DRASchedulerFilterTimeout |
DRA 主线:ResourceSlice/ResourceClaim 结构化设备参数,让 scheduler/autoscaler 可推理设备。 |
3063-dynamic-resource-allocation |
node |
Historical |
withdrawn / alpha |
v1.26 / - / - |
DynamicResourceAllocation, DRAControlPlaneController |
早期 opaque DRA controller allocation 路线,已撤回;主要价值是解释为什么转向 structured parameters。 |
5075-dra-consumable-capacity |
scheduling |
Landmark |
implementable / beta |
v1.34 / v1.36 / v1.38 |
DRAConsumableCapacity |
设备不再只是离散实例,也可以有可消费 capacity,例如 GPU memory、带宽、license。 |
5517-dra-node-allocatable-resources |
scheduling |
Important |
implementable / alpha |
v1.36 / v1.38 / v1.39 |
DRANodeAllocatableResources |
暴露 DRA node allocatable,给调度和容量推理提供节点级库存信号。 |
5729-resourceclaim-support-for-workloads |
scheduling |
Important |
implementable / beta |
v1.36 / v1.37 / - |
DRAWorkloadResourceClaims |
ResourceClaim 支持 workload 级使用,连接 DRA 和 PodGroup/gang scheduling。 |
5941-dra-shared-consumable-capacity |
scheduling |
Important |
implementable / alpha |
v1.37 / v1.38 / v1.39 |
DRASharedConsumableCapacity |
多个 logical device 共享同一个 capacity pool,处理共享 buffer、memory 或 fabric bandwidth。 |
5963-device-compatibility-groups |
scheduling |
Important |
implementable / alpha |
v1.37 / v1.38 / v1.39 |
DRADeviceCompatibilityGroups |
表达多个设备必须兼容或来自同一组合集合,适合多卡/多 NIC 配对。 |
6080-dra-derived-attributes |
scheduling |
Watch |
provisional / alpha |
v1.37 / v1.38 / v1.40 |
DRADerivedAttributes |
从 driver 属性派生标准化匹配属性;还不是稳定实现主线。 |
4817-resource-claim-device-status |
node |
Important |
implementable / stable |
v1.32 / v1.33 / v1.37 |
DRAResourceClaimDeviceStatus |
将分配到的设备信息写入 ResourceClaim status,给用户和控制器可观测结果。 |
5304-dra-attributes-downward-api |
node |
Important |
implementable / beta |
v1.36 / v1.37 / v1.38 |
NA |
把 DRA device attributes 暴露给 Pod,方便 workload 发现自己拿到的设备属性。 |
5677-dra-resource-availability-visibility |
node |
Important |
implementable / alpha |
v1.36 / - / - |
DRAResourcePoolStatus |
暴露资源池可用性,减少用户只看到 claim pending 却不知道设备库存原因。 |
5945-dra-optional-node-preparation |
node |
Important |
implementable / alpha |
v1.37 / v1.38 / v1.39 |
DRAOptionalNodePreparation |
允许某些 DRA 场景跳过强制 NodePrepare,降低轻量设备或无需节点准备的路径成本。 |
6072-dra-standard-numanode |
node |
Important |
implementable / stable |
- / - / v1.37 |
- |
标准化 DRA numaNode 属性,连接 DRA 和 Topology Manager / NUMA-aware scheduling。 |
4951-configurable-hpa-tolerance |
autoscaling |
Important |
implementable / stable |
v1.33 / v1.35 / v1.37 |
HPAConfigurableTolerance |
把 HPA tolerance 从全局参数下放到 HPA scaleUp/scaleDown 规则。 |
1610-container-resource-autoscaling |
autoscaling |
Important |
implemented / stable |
v1.20 / v1.27 / v1.30 |
HPAContainerMetrics |
HPA 可按指定 container 的 CPU/memory 扩缩,避免 sidecar 稀释主容器指标。 |
5325-hpa-pod-selection-accuracy |
autoscaling |
Important |
implementable / alpha |
v1.35 / v1.36 / v1.37 |
HPASelectionStrategy |
HPA 可按 owner reference 过滤 Pod,避免 label selector 误选其他 workload。 |
2021-scale-from-zero |
autoscaling |
Landmark |
implementable / beta |
v1.16 / v1.37 / x.y |
HPAScaleToZero |
HPA 基于 object/external metrics 支持 0 副本恢复,用 condition 区分自动缩零与用户暂停。 |
5679-external-metric-fallback |
autoscaling |
Important |
implementable / alpha |
v1.36 / v1.37 / v1.38 |
HPAExternalMetricFallback |
external metric 连续失败时使用固定 fallback replicas,并进入 HPA 多指标 max 合并。 |
5030-attach-limit-autoscaler |
autoscaling |
Important |
implementable / beta |
v1.35 / v1.37 / v1.38 |
VolumeLimitScaling |
Cluster Autoscaler 模拟 CSI volume attach limit,避免加了节点仍不可调度。 |
1287-in-place-update-pod-resources |
node |
Landmark |
implemented / stable |
v1.27 / v1.33 / v1.35 |
InPlacePodVerticalScaling, InPlacePodVerticalScalingAllocatedStatus |
Pod /resize subresource 与 desired/allocated/actual 状态模型,支持 CPU/memory 原地调整。 |
5419-pod-level-resources-in-place-resize |
node |
Important |
implementable / beta |
v1.35 / v1.36 / - |
InPlacePodLevelResourcesVerticalScaling, PodLevelResources |
将 in-place resize 扩展到 Pod-level resources。 |
5526-pod-level-resource-managers |
node |
Important |
implementable / beta |
v1.36 / v1.37 / v1.39 |
PodLevelResources, PodLevelResourceManagers |
CPU/Memory/Topology Manager 从 container scope 扩展到 Pod scope。 |
5554-in-place-update-pod-resources-alongside-static-cpu-manager-policy |
node |
Important |
implementable / alpha |
v1.37 / v1.38 / v1.39 |
InPlacePodVerticalScalingExclusiveCPUs |
处理 static CPU Manager exclusive CPUs 下的原地 resize。 |
6122-configurable-scaling-delay-with-pod-resource-exposure |
node |
Important |
implementable / alpha |
v1.37 / v1.39 / v1.40 |
DownwardAPIAssignedResources 等 |
在收回 exclusive CPUs 前暴露资源变化并给应用准备窗口。 |
5836-scheduler-preemption-for-ippr |
scheduling |
Important |
implementable / alpha |
v1.37 / v1.38 / v1.39 |
SchedulerPreemptionForPodResize |
scheduler 为 deferred in-place resize 在已绑定节点上触发抢占。 |
2371-cri-pod-container-stats |
node |
Important |
implementable / beta |
v1.29 / v1.37 / - |
PodAndContainerStatsFromCRI |
kubelet stats 从 cAdvisor 迁往 CRI,影响 metrics-server/HPA 的输入可靠性。 |
5394-psi-node-conditions |
node |
Important |
implementable / alpha |
v1.36 / - / - |
PSINodeCondition |
将 PSI pressure 升级为 Node Conditions,便于调度、排障和自动化响应。 |
4680-add-resource-health-to-pod-status |
node |
Important |
implementable / beta |
v1.31 / v1.36 / v1.37 |
ResourceHealthStatus |
Device Plugin / DRA 资源健康进入 Pod status。 |
2033-kubelet-in-userns-aka-rootless |
node |
Important |
implementable / beta |
v1.22 / v1.37 / - |
KubeletInUserNamespace |
kubelet 以 user namespace/rootless 模式运行,降低节点控制面特权。 |
5607-hostnetwork-userns |
node |
Important |
implementable / alpha |
v1.35 / - / - |
UserNamespacesHostNetworkSupport |
HostNetwork Pod 也可使用 user namespaces,补上高权限网络场景隔离。 |
4438-container-restart-termination |
node |
Important |
implementable / alpha |
v1.37 / v1.38 / v1.39 |
SidecarsRestartableDuringPodTermination |
明确 Pod 终止期间 sidecar 是否可 restart,修补 lifecycle 边界。 |
4563-eviction-request-api |
node |
Important |
implementable / alpha |
v1.37 / - / - |
EvictionRequestAPI |
给 eviction 请求更清晰的 API 表达。 |