Scaling

Ten times the traffic. The same Tuesday afternoon.

Scaling happens on infrastructure that is already provisioned and already warm. There is no new environment to stand up and no migration window — the cluster grows inside the account it was always in, and nobody gets paged for it.

Signal
Requests, queue depth, CPU, memory
Bounds
A floor and a ceiling per service
Reaction
Seconds, on warm capacity
Scope
Per service, not per cluster
Requests / sec1,240p95118 msflat
Load balancerHTTP/3×1apinode 20×2workerjobs×1postgres15.4×1cache7.2×1

Steady at api ×2. Nobody was paged.

A twelve-second loop: load arrives, replicas follow, latency holds.

How it behaves

Bounded, warm and per service.

Autoscaling fails in three predictable ways — it reacts to the wrong number, it reacts too late, and it has no upper bound. These are the three.

  • A floor and a ceiling

    Set the minimum you always want running and the maximum you are willing to pay for. Kapten stays between them, which is what keeps an autoscaler from becoming a billing incident.

  • Scales on the signal that matters

    Requests per second for a web service, queue depth for a worker, CPU or memory where those are what actually saturate. Picking the wrong signal is the usual reason autoscaling does nothing.

  • Warm, not cold

    New replicas land on nodes that are already in the cluster, so there is no instance boot sitting in the critical path of a traffic spike.

  • The node pool follows

    When the services outgrow the nodes underneath them, the pool grows too — and shrinks back once the spike has passed, without leaving the cluster permanently sized for its worst day.

  • A pinned ceiling is a signal

    A service sitting at its maximum is not scaling any more, it is degrading quietly. Kapten treats that as something to tell you about rather than something to absorb.

  • Idle can go to zero

    Preview environments and batch workers can sleep between uses and come back on the next request, instead of holding capacity for a branch nobody has touched in a week.

Ready when you are

Your cloud. Our ergonomics.

Connect a cloud account and watch a production-grade environment come up in your own VPC. If it isn’t serving traffic in fifteen minutes, we want to hear why.