Recreate and rolling updates
Recreate
All old instances stop, then new ones start.
- Pro: simplest; versions never run concurrently (good when they can't — e.g. an app that takes an exclusive lock or runs incompatible migrations on startup).
- Con: downtime between old stopping and new being ready; rollback also means downtime.
spec:
strategy:
type: Recreate
Use it for dev environments, singleton workers, or when brief downtime is acceptable and coexistence is impossible.
Rolling update — how Kubernetes actually does it
RollingUpdate is the default for Deployments. Changing the pod template (new image, env, config) creates a new ReplicaSet; the Deployment controller scales the new one up and the old one down in steps, governed by two knobs:
- maxSurge — how many pods above
replicasmay exist during the rollout (number or %, default 25%, rounded up). - maxUnavailable — how many pods below
replicasmay be unavailable (default 25%, rounded down).
Worked example, replicas: 4:
| Setting | Behaviour |
|---|---|
| defaults (25%/25%) | surge 1, unavailable 1: at least 3 available, at most 5 total |
maxSurge: 1, maxUnavailable: 0 |
Start one new pod, wait until it's available, remove one old — capacity never drops; slowest, safest |
maxSurge: 0, maxUnavailable: 1 |
Remove one old first, then add new — no extra capacity needed, but capacity dips |
maxSurge: 100%, maxUnavailable: 0 |
Bring up a full new set, then remove the old — fast, blue-green-like capacity cost |
Both can't be zero.
Other fields that matter:
spec:
replicas: 6
minReadySeconds: 15 # a new pod must stay Ready this long before it counts as "available"
progressDeadlineSeconds: 600 # no progress for 10 min → condition Progressing=False (ProgressDeadlineExceeded)
revisionHistoryLimit: 10 # old ReplicaSets kept for rollback
strategy:
type: RollingUpdate
rollingUpdate: { maxSurge: 25%, maxUnavailable: 0 }
- minReadySeconds catches pods that pass readiness briefly then crash.
- progressDeadlineSeconds doesn't roll back by itself — it marks the rollout failed so
kubectl rollout statusexits non-zero and your pipeline/Flux health checks can react. - A failing readiness probe stalls the rollout: new pods never become available, old pods aren't removed, users keep being served by the old version. That's the safety net working.
kubectl rollout status deploy/orders --timeout=5m # wait; non-zero exit on failure
kubectl rollout history deploy/orders
kubectl rollout history deploy/orders --revision=7
kubectl rollout undo deploy/orders # previous revision
kubectl rollout undo deploy/orders --to-revision=5
kubectl rollout pause deploy/orders # batch several changes, or freeze mid-rollout
kubectl rollout resume deploy/orders
kubectl rollout restart deploy/orders # re-create pods with the same spec
Set kubernetes.io/change-cause annotations (or rely on Git history with GitOps) so revisions are understandable.
Watch out: with Flux/GitOps,
kubectl rollout undois temporary — Flux reconciles the Deployment back to what's in Git. Roll back by reverting the commit (or suspend the Kustomization during an emergency, then revert).
Rolling updates outside Kubernetes
- EC2 Auto Scaling — Instance Refresh: replace instances in batches with a new launch template version;
MinHealthyPercentage, warm-up time, checkpoints, and automatic rollback on failure. - ECS:
minimumHealthyPercent/maximumPercent(the same idea as maxUnavailable/maxSurge), with the deployment circuit breaker to roll back automatically when tasks fail to stabilise. - Elastic Beanstalk / CodeDeploy in-place: rolling with a batch size.
Pros of rolling: no duplicate environment, gradual. Cons: both versions serve traffic simultaneously (they must be compatible), and rollback is another rolling update, not instant.