Add Serving Workload section on KEP-976.

keps/976-plain-pods/README.md

@@ -551,6 +551,24 @@ spec:
               cpu: 1m
 ```
 
+#### Serving Workload
+
+1. The Pod Group integration adds successfully completed pods to the `ReclaimablePods` list. However,
+   this is problematic for serving workloads, such as `StatefulSet`, because it prevents to ungate the replacement
+   pod. This behavior is incorrect, as recreated pods for serving workloads should continue to run
+   regardless if the pod was failed or succeeded.
+   To resolve this issue, the `kueue.x-k8s.io/pod-group-serving` annotation can be used. When this
+   annotation is set to true, the `ReclaimablePods` mechanism no longer tracks the number of
+   pods, allowing to ungate the replacement pod.
+2. The Pod Group integration waits until all pods are created. However, for serving workloads such
+   as `StatefulSets` with a `PodManagementPolicyType` of `OrderedReady`, pods are created sequentially,
+   with each subsequent pod being created only after the previous pod is fully running. This
+   sequential behavior can result in a deadlock.
+   To resolve this issue, the `kueue.x-k8s.io/pod-group-fast-admission` annotation is used.
+   When this annotation is set to true, the PodGroup can proceed with admission without requiring
+   all pods to reach the ungated state.
+
+
 ### Tracking admitted and finished Pods
 
 Pods need to have finalizers so that we can reliably track how many of them run to completion and be