Skip to content

Quotas and limits

Why a workload can be accepted and still not start, and where to look when it does not.

Every namespace your project owns carries a quota: how much CPU and memory the workloads in it may request and use in total, and how many pods may exist.

The quota is enforced by Kubernetes itself, not by a periodic check, so it is exact and immediate.

The failure this causes

A workload is accepted when you deploy it, and then does not start.

That is not a bug, and it is the single most common confusing outcome on the platform. Accepting a deployment means the request was valid. Starting a pod means the cluster found room for it inside the quota. A workload asking for more than the namespace has left will sit unscheduled rather than evicting something that is already running.

The workload’s page says so. Look there before looking at your image or your application — a pod that never started has no logs to explain itself.

Getting room

Stop something. A stopped workload releases its share of the quota immediately, which is often the fastest answer during development.

Deploy a smaller template. Templates carry their resource requests; a smaller one may fit where the one you chose does not.

Ask for more. Quota is an allocation decision, so project admins cannot raise their own — a platform administrator sets it. That is the same reason storage quotas are not self-service.

Storage quota is separate

Object storage has its own per-project limit, unrelated to the compute quota above, and it fails differently: a write past the limit is refused at the moment it happens rather than leaving something unscheduled. See Object storage.

Network policy

A namespace can also be isolated at the network level, which is chosen when the namespace is bound to your project:

  • Off — the cluster’s own default applies, usually allowing all traffic.
  • Isolated — nothing may open a connection into the namespace except workloads already inside it.
  • Project-internal — as isolated, plus other namespaces belonging to the same project.

If a workload cannot reach another one it should be able to, this is worth checking before debugging the application. Public endpoints are unaffected — the gateway is not subject to these rules.