Scaling and growth

Plan additional capacity to accommodate workload growth, maintenance operations, and high-availability requirements. Capacity planning should account for projected workload growth, node failures, maintenance operations, and service expansion. Review capacity utilization regularly and expand compute or storage resources before resource exhaustion affects service availability or performance.

Compute

Production deployments should reserve sufficient CPU and memory capacity to allow workloads to restart after a compute worker failure. When available reserve capacity becomes insufficient, add additional compute worker nodes.

Virtuozzo Storage

Let the cluster's storage alerts drive capacity planning rather than a fixed target. Each storage tier raises:

• A warning alert ("Cluster is running out of physical space on tier") when the tier drops below 20 percent free.
• A critical alert ("Storage tier is out of space") when the tier has less than 10 percent remaining. Write operations to the tier are blocked once it fills.

Add storage capacity (disks or nodes) when these alerts appear, and do not wait beyond the critical alert. Keeping at least 20 percent of each tier free also leaves the cluster room to rebuild data after a disk or node failure. For EC tiers, RAM capacity planning must include the per-node storage-client reservation: 5 percent of the node's total RAM plus 2 GiB.

Quantity of servers

The number of nodes determines the redundancy the cluster can provide:

• One node: evaluation only; a single metadata service and single replica mean a failure stops the cluster or loses data. In this configuration, use the disk failure domain instead of the host failure domain.
• Two nodes: same redundancy as a single node; VM high availability is not supported.
• Three nodes: minimum for production and for testing all product features; survives the failure of one node.
• Five or more nodes: survives the failure of two nodes, provides resilient metadata with five metadata services, and enables a wider choice of erasure-coding schemes.
• Ten or more nodes: wide EC schemes, such as 7+2, at low overhead, faster rebuilds, and better balance. To keep performance, keep at least 20 percent of capacity free.