Skip to main content
Version: v0.38 Stable

Size for high availability and scale

A production cluster needs a control plane that survives a replica failure and has headroom for its workload. Make these decisions before production traffic arrives. Each section links to the procedure that carries it out.

Run three control plane replicas​

Run three replicas for any production cluster. A single replica can't tolerate even a brief etcd compaction stall without affecting the cluster's API availability.

More than one replica needs a backing store other than embedded SQLite, which vCluster rejects with multiple replicas. See Choose a backing store.

To understand what tenants see when a control plane or the control plane cluster fails, see What happens during control plane outages?.

To deploy, see High availability for a pod-based control plane, or Standalone high availability.

Choose a resource profile​

Resource profiles set the CPU and memory for each control plane replica. Start with P3, 4 CPU and 8 GiB per replica, which suits most production clusters.

  • Step up to P5 for operator-rich workloads, such as clusters with 20 or more controllers holding persistent watches, or for latency-sensitive SLOs under sustained churn.
  • Above 4 CPU, more CPU buys latency headroom rather than more nodes. The etcd write path becomes the limit.
  • Memory follows object count, not node count. Plan memory from the number of pods, Deployments, and Services you expect.

For all six profiles and the tuning settings, see Control plane sizing and performance. The numbers come from the control plane benchmark methodology.

Plan DNS capacity​

CoreDNS runs a single replica by default. Total DNS query volume decides when that's not enough, and AI and ML workloads with bursty pod scheduling are the most common trigger. See Why does CoreDNS become a DNS bottleneck?.

The default separate CoreDNS Deployment scales by adding replicas. See Scale CoreDNS for large node counts. Embedded CoreDNS gains capacity only through control plane resources and replicas. You can switch between the two on a live cluster, so this choice isn't fixed at creation.

Plan control plane cluster capacity​

Sizing is per cluster, not aggregate. To plan the control plane cluster's nodes, multiply the per-cluster profile by the number of clusters you expect to host. See Operational guidance for node sizes and alert thresholds.

Next steps​