Kubernetes - Medium

Workloads

Not every workload fits the Deployment model of "identical, interchangeable replicas."

• StatefulSet - for workloads where identity matters: each replica gets a stable, predictable name (db-0, db-1, db-2, not random suffixes), its own stable PersistentVolumeClaim that follows it across rescheduling, and ordered, sequential startup/shutdown (db-0 fully ready before db-1 starts). This is what real databases and other clustered stateful systems need - a plain Deployment's interchangeable, arbitrarily-ordered Pods are actively wrong for them.
• DaemonSet - runs exactly one Pod on every node (or every matching node) automatically, including new nodes as they join the cluster - the standard shape for node-level agents: log collectors, monitoring agents, CNI plugins themselves.
• Job - runs a Pod (or several) to completion for a one-off task, tracking success/failure, and retrying on failure up to a configurable limit - the right primitive for a batch task, not a Deployment (which would just restart a "completed" Pod forever, since it doesn't understand success).
• CronJob - a Job that runs on a schedule (standard cron syntax) - nightly backups, scheduled reports.

Headless Services (clusterIP: None) are what StatefulSets rely on for stable per-Pod DNS names (db-0.db-service.namespace.svc.cluster.local) rather than one load-balanced VIP - each Pod needs to be individually addressable, not hidden behind aggregate load-balancing.

kubectl get statefulsets
kubectl get daemonsets
kubectl get jobs
kubectl get cronjobs
kubectl create job manual-run --from=cronjob/nightly-backup     # trigger a CronJob's Job manually, e.g. to test it

---

Scaling

Horizontal Pod Autoscaler (HPA) adjusts a Deployment/StatefulSet's replica count automatically based on observed metrics - most commonly CPU or memory utilization percentage against the Pod's own requests, but it can also scale on custom or external metrics (queue depth, requests-per-second) once a metrics adapter is wired in.

kubectl autoscale deployment web --cpu-percent=70 --min=2 --max=10
kubectl get hpa                                      # current vs target metric, current replica count

Vertical Pod Autoscaler (VPA) takes the opposite axis - instead of changing replica count, it adjusts a Pod's resource requests/limits based on observed real usage. Conceptually useful for right-sizing workloads whose actual CPU/memory needs are hard to guess up front, though it typically requires restarting Pods to apply a new size (a real operational tradeoff against HPA, which scales without disrupting existing Pods).

Cluster Autoscaler operates one layer up from both: it adds or removes nodes from the cluster based on whether Pods are stuck Pending for lack of capacity, or nodes are sitting significantly underutilized. HPA/VPA answer "how much should this workload consume," Cluster Autoscaler answers "does the cluster have enough machines to fit what's being asked of it."

---

Rolling Updates

Revisiting the Deployment rollout mechanics with the knobs that actually matter operationally:

• maxSurge - how many Pods above the desired replica count are allowed to exist temporarily during a rollout (absolute number or percentage).
• maxUnavailable - how many Pods below the desired count are tolerated during a rollout.

Setting maxUnavailable: 0 guarantees full capacity is always available during a rollout (strictly surge-only), at the cost of needing enough spare cluster capacity to briefly run more Pods than the steady-state count - the real tradeoff between guaranteed availability and resource headroom.

A rollout that gets stuck (new Pods crash-looping, never becoming Ready) doesn't roll back on its own by default - it just stalls, with kubectl rollout status hanging and old Pods still serving traffic. kubectl rollout undo is a manual, deliberate action; progressDeadlineSeconds is what marks the Deployment as failed after a timeout, but even then, nothing auto-reverts unless something else (a CI/CD pipeline, an operator) is watching for that condition and triggers the rollback itself.

---