Every scheduling decision runs through named extension points you can plug into.

Since 1.19 the scheduler is a framework. Built-in behaviour such as taints, affinity, preemption and volume binding are all plugins at these points.

QueueSort

order pending Pods

PreFilter

precompute, early reject

Filter

remove infeasible nodes

PostFilter

only if none feasible: preemption

Score

rank feasible nodes

Reserve

hold resources on the winner

Permit

approve, deny or wait: gang scheduling

PreBind

e.g. provision a volume

Bind

write nodeName

PostBind

informational

PostFilter = preemption

When filtering finds nothing, the default plugin checks whether evicting lower-priority Pods would make room.

Permit = gang scheduling

A distributed training job needs all its Pods or none. Plugins in Volcano or Kueue hold Pods here until the whole group fits.

Why you care

Specialised needs (GPU, batch, topology) are added as plugins or extra schedulers, not by forking Kubernetes.

The scheduler is a pipeline of plugins

The scheduler is not a monolith. Since Kubernetes 1.19 it is built on the scheduling framework: a fixed sequence of extension points with plugins attached. Even the built-in features you use daily, such as node affinity, taints and tolerations, topology spread and volume binding, are plugins implementing one or more of these points.

The key points to know: Filter and Score are the two phases you already met. PostFilter runs only when no node is feasible, and the default plugin there implements preemption. Reserve holds the resources on the chosen node so the next Pod's decision sees them. Permit is a final gate that can delay binding, which is exactly what gang scheduling needs so that a group of Pods is placed together or not at all.

Why does this matter to an operator? Because it explains where behaviour comes from and how to change it. You can configure plugin weights through a scheduler profile, run a second scheduler for special workloads (selected with schedulerName in the Pod spec), or use projects such as Kueue and Volcano for batch and AI workloads.

In managed Kubernetes you usually cannot change the control-plane scheduler's configuration, but GKE, EKS and AKS all support running additional schedulers as Deployments.