kubernetes

We know Kubernetes from the inside.

Clusters, upgrades, multi-tenancy: routine work. The difference shows when the problem has no off-the-shelf solution: at that point we write the operator ourselves. It happened with Understudy, open source and Apache 2.0: it keeps single-replica workloads alive through drains, consolidations and spot interruptions, without paying for a second replica around the clock. It’s not the product: it’s the level we work at.

what that means, in practice

operators and admission webhooks written to measure, when it’s truly needed
cluster upgrades without night-time windows or crossed fingers
multi-tenancy with real boundaries: namespaces, quotas, network policies, costs
capacity and autoscaling tuned with Karpenter, spot included

In a real spot interruption, measured on EKS, Understudy lost 1 request out of 764: less than an ordinary rolling update.

The rest of the numbers, with their context, is in the README.

understudy.yaml

apiVersion: apps.understudy.sh/v1alpha1
kind: Understudy
metadata:
  name: tenant-worker
spec:
  targetRef:
    kind: Deployment
    name: tenant-worker
  holdMode: always
  minSurgeTimeSeconds: 60
  readinessDeadlineSeconds: 600

Apache 2.0 · Go · CRD and admission webhook · the repository

We know Kubernetes from the inside.

A technical call, not a demo.