Self-Healing Deployments: ReplicaSets, Probes, and PDBs

Three replicas. Kubernetes reports 3/3 Running. Everything looks fine, until a bug causes your process to enter a deadlock. It accepts connections but hangs every request. Pod count stays at 3. Traffic routes to all three. Nothing restarts. The ReplicaSet counts pods, not health.

This is the gap between pod count and application availability. Three primitives close it. ReplicaSets reconcile desired replica count against actual. Probes teach the kubelet what "healthy" means for your specific container. PodDisruptionBudgets cap how many pods can be voluntarily disrupted at once. None of them overlap. Each operates at a different layer. Together they cover the failure modes that a plain pod count misses.