- Fixed incorrect behavior of preemptor pod when preemption of the victim takes long to complete. The preemptor pod should not be circling in scheduling cycles until preemption is finished. ([#134294](https://github.com/kubernetes/kubernetes/pull/134294), [@ania-borowiec](https://github.com/ania-borowiec)) [SIG Scheduling and Testing]

  maxSurge: 100%

## Update strategy for StatefulSets
statefulSetUpdate:
  updateStrategy: RollingUpdate

## Pod priority settings
## ref: https://kubernetes.io/docs/concepts/configuration/pod-priority-preemption/
##
priorityClassName: ""

## Pod runtime class name
## ref https://kubernetes.io/docs/concepts/containers/runtime-class/
##
runtimeClassName: ""

## Set default rootUser, rootPassword

* **[Pod priority and preemption](https://kubernetes.io/docs/concepts/configuration/pod-priority-preemption/) is now enabled by default.** Note that this means that pods from *any* namespace can now request priority classes that compete with and/or cause preemption of critical system pods that are already running. If that is not desired, disable the PodPriority feature by setting `--feature-...

### Bug or Regression

- Fix a scheduler preemption issue where the victim pod was not deleted due to incorrect status patching. This issue occurred when the preemptor and victim pods had different QoS classes in their status, causing the preemption to fail entirely. ([#126694](https://github.com/kubernetes/kubernetes/pull/126694), [@Huang-Wei](https://github.com/Huang-Wei)) [SIG Scheduling]

- Added kubectl explain -r flag as a shorthand for --recursive ([#135283](https://github.com/kubernetes/kubernetes/pull/135283), [@laervn](https://github.com/laervn)) [SIG CLI]
- Align the meaning of victim metrics between async preemption and sync preemption. The definition has been standardized to refer to the number of Pods chosen as victims. ([#135955](https://github.com/kubernetes/kubernetes/pull/135955), [@utam0k](https://github.com/utam0k)) [SIG Scheduling]

- Fixed a scheduler preemption issue where the victim pod was not deleted due to incorrect status patching. This issue occurred when the preemptor and victim pods had different QoS classes in their status, causing the preemption to fail entirely. ([#126644](https://github.com/kubernetes/kubernetes/pull/126644), [@Huang-Wei](https://github.com/Huang-Wei))

- A terminating pod on a node that is not caused by preemption no longer prevents `kube-scheduler` from preempting pods on that node
  - Rename `PreemptionByKubeScheduler` to `PreemptionByScheduler` ([#114623](https://github.com/kubernetes/kubernetes/pull/114623), [@Huang-Wei](https://github.com/Huang-Wei))

// TFE_ContextUpdateServerDefWithTimeout to be simple wrappers around the same
// C++ function.
// Retries are used for CreateContext calls, which is used in
// ParameterServerStrategy initialization to be robust to worker preemption.
TF_CAPI_EXPORT extern void TFE_ContextSetServerDefWithTimeoutAndRetries(
    TFE_Context* ctx, int keep_alive_secs, const void* proto, size_t proto_len,
    int64_t init_timeout_in_ms, int retries, TF_Status* status,

Pod Priority and Preemption in Kubernetes ([#564](https://github.com/kubernetes/enhancements/issues/564))

    int64_t init_timeout_in_ms, TF_Status* status,
    bool clear_existing_contexts);

// Set server def with retries and timeout. This is helpful for fault-tolerant
// initial connection in high-preemption environments, such as
// ParameterServerStrategy training.
// This API is for experimental usage and may be subject to change.
TF_CAPI_EXPORT extern void TFE_ContextSetServerDefWithTimeoutAndRetries(