Managed Kubernetes takes the control plane off your hands: the API server, scheduler and etcd are someone else’s problem. That is a genuine reduction in work. It is also commonly mistaken for the whole of the work.
What remains yours
- Node pool sizing, scaling behaviour and upgrade timing.
- Workload manifests: resource requests, limits, probes, disruption budgets.
- Ingress, certificates and DNS.
- Storage classes and what happens to volumes when pods move.
- Secrets handling and workload identity.
- Observability: metrics, logs and traces, and alerts worth waking up for.
Set resource requests properly
Requests and limits are how the scheduler reasons about your cluster. Set them too low and pods get evicted or throttled under load; too high and you pay for capacity nothing uses. Measure actual consumption over a representative period and set requests near the observed steady state, with limits providing headroom rather than aspiration.
| Symptom | Usual cause |
|---|---|
| Pods pending indefinitely | Requests exceed what any node can offer. |
| Frequent restarts under load | Memory limit set below real usage. |
| Slow responses, low CPU shown | CPU limit throttling, or a downstream wait. |
| Node pool never scales down | Pods without disruption budgets blocking drain. |
| Costs rising without more traffic | Over-generous requests reserving idle capacity. |
Upgrades are a scheduled activity
- Track the versions your provider supports and when yours stops being supported.
- Read the changelog for removed APIs before upgrading, not after.
- Upgrade a non-production cluster first, with the same manifests.
- Upgrade the control plane, then node pools, watching workloads drain.
- Keep the whole thing inside a maintenance window you have announced.
Know when you do not need it
If you run a handful of services with steady traffic and no need for sophisticated scheduling, a managed container service or a couple of hosts will serve you with far less to learn. Kubernetes earns its keep when you have enough services, teams or variability that manual placement genuinely stops working.
Summary
Let the provider run the control plane, then treat requests, upgrades, ingress and observability as the real work – because they are. Clusters that stay manageable are the ones where those four are handled deliberately.
