/gke-cluster-autoscaler

Cluster Autoscaler: Capacity Buffers (Pre-warm)

Launch stage: Preview (Pre-GA Offerings Terms). Version gates: active buffers require GKE 1.35.2-gke.1842000+; standby buffers require GKE 1.36.0-gke.2253000+.

CapacityBuffer (CRD)

• API: autoscaling.x-k8s.io/v1beta1, kind: CapacityBuffer. Namespace-scoped.
• Targets a specific ComputeClass via nodeSelector in the podTemplateRef; works with node pool auto-creation (buffers honor the class's priorities[]).
• Two provisioningStrategy values (do not confuse the domains):
  • Active: buffer.x-k8s.io/active-capacity (default) — placeholder pods hold fully running warm nodes; real workloads evict them instantly. You pay full node price while idle.
  • Standby: buffer.gke.io/standby-capacity (June 2026) — nodes are fully initialized (DaemonSets started, images pulled), then suspended.

Standby buffers (buffer.gke.io/standby-capacity)

• Cost model: while suspended you pay only persistent disk + IP address (node state is stored to disk; compute/memory charges stop). Resume takes ~30s — 2-3x faster than provisioning a fresh node.
• Requires: GKE 1.36.0-gke.2253000+; Standard clusters with node auto-provisioning / node pool auto-creation enabled. ComputeClass targeting works as with active buffers.
• Annotations (on the CapacityBuffer):
  • buffer.gke.io/standby-capacity-init-time (default "5m"): how long a (re)created buffer node stays active before suspending.
  • buffer.gke.io/standby-capacity-refresh-frequency (default "24h"; "never" disables): how often standby nodes are recreated.
• Detection: suspended nodes carry node condition Suspended (True = suspended, False = resumed, absent = never suspended).
• Hybrid pattern: small active buffer + larger standby buffer — GKE refills the active buffer by resuming standby nodes (~30s) and backfills standby in the background.
• Exclusions: no GPU/TPU nodes, Local SSD, CMEK, or Confidential GKE Nodes; Compute Engine suspend/resume limits apply (e.g., >208 GB memory nodes). Accelerator pre-warming must use active buffers.

Sizing Modes

• Fixed: replicas: 3. Always keep N units warm.
• Dynamic: percentage: 20 + scalableRef: <Deployment>. Headroom scales with workload.
• Resource limits: limits: {cpu, memory} caps total buffered capacity.

Why use Buffers?

• Bursty Serving: Pod-pending SLOs can't tolerate 60-120s node pool auto-creation delay.
• HPA outpaces cluster autoscaler: Workload scales faster than nodes can arrive.
• Pre-warming: Warm GPUs/TPUs before known traffic windows (active buffers only).
• Standby = cheap headroom: overprovisioned warm capacity at disk+IP cost instead of full node cost, if workloads tolerate the ~30s resume delay.

Note: Replaces the "dumb" floor of --min-nodes with shape-aware, class-targeted warm capacity.

Related: node pool auto-creation itself is up to 85% faster from GKE 1.34.1-gke.1829001 (multiple node pools created concurrently; no configuration needed).