Kubernetes Deployments, Autoscaling, & Self-Healing

Kubernetes operates on a declarative Reconciliation Loop (Control Loop) model:

Loop: Observe Current State \longrightarrow Compute Diff vs Desired State \longrightarrow Execute Corrective Action

Self-Healing Scenarios:

1. Container Crash: If an application crashes (Exit Code 1 or panic), the kubelet restarts the container according to its restartPolicy (e.g., Always), applying exponential backoff delay (CrashLoopBackOff) if crashes persist.
2. Node Hardware Failure: Worker nodes maintain periodic heartbeats with the API Server via Lease objects (every 10 seconds). If a node fails to report for node-monitor-grace-period (default 40s), it is marked NotReady. After pod-eviction-timeout (default 5m), the controller evicts the pods and schedules replacements on surviving healthy nodes.
3. Pod Desired Count Invariant: A Deployment manages a ReplicaSet. If the user declares replicas: 10 and 2 pods are terminated, the ReplicaSet controller detects the diff (Desired = 10, Actual = 8) and immediately creates 2 replacement pods.