A single replica can be a deliberate choice for a service with limited operating hours. It also makes the disruption policy easy to misunderstand: a minAvailable: 1 PodDisruptionBudget cannot turn one pod into an always-available service.

At SmartDoctor, I used schedule-aware scaling for suitable workloads and investigated how PriorityClass, PDBs and rollout controls interacted. The useful result was a clearer boundary for each mechanism, rather than treating their presence as evidence of availability.

First identify who is removing the pod

A normal eviction-aware node drain asks the Eviction API for permission. A matching PDB can reject that request, so drain may wait until a replacement is available or an operator changes the conditions.

Scheduler preemption is different. A pending higher-priority pod may require lower-priority pods to leave. Kubernetes considers PDBs when selecting victims, but respecting them is best effort; a PDB violation can remain necessary to schedule the higher-priority workload.

Direct pod deletion, node failure and node-pressure eviction have different boundaries again. A PDB does not prevent all of those disruptions.

A Deployment rollout has its own contract

Deployment rolling updates are governed by the workload’s rollout strategy. maxUnavailable, maxSurge, readiness and rollout deadlines determine the replacement sequence. A PDB is not the control that stops a Deployment from updating its pods.

That matters when a service has one replica. An allowed voluntary eviction may take the service down; a PDB requiring that sole pod to remain can instead block a node drain. Either outcome needs an explicit operational decision.

I separated these questions during investigation:

  • Is the workload expected to serve requests at this time?
  • Can a replacement become ready before the current instance leaves?
  • Is deletion using the Eviction API or bypassing it?
  • Is higher-priority work causing scheduler preemption?

Priority expresses scheduling importance

The retained charts define critical, service and background priority classes, including classes configured not to initiate preemption. A class with preemptionPolicy: Never can still be preempted by another eligible higher-priority pod; it controls whether that class initiates preemption.

Priority does not create spare capacity, repair an unhealthy dependency or preserve a connection during process shutdown. It belongs alongside placement, replicas and graceful termination, with each serving a different purpose.

Make the tradeoff visible

Some reviewed KEDA templates request one daytime replica in development/release and permit zero outside the schedule. Hospital-group KEDA is disabled in the snapshot. Those are service-specific configuration choices, not a claim that the whole production fleet used the same policy.

For a request-serving workload, I would review the operating window, readiness dependency, replacement capacity and disruption path together. For a queued worker, I would also inspect acknowledgment and replay behavior. The same replica count can carry very different consequences.

The readiness, deadline and rollback lab explores one controlled release failure. Its local observations do not establish historical hospital-service uptime. The Istio routing case explains the group and gateway boundaries behind the platform configuration.

References: Kubernetes disruptions, Pod priority and preemption.