Temper documentation
Everything you need to install the platform, protect your first workload, and run it in production. Start with the quickstart; it takes about 15 minutes on a real cluster.
Getting started
Getting startedQuickstarthelm install to a protected node in ~15 minutes: build images, install the chart, verify, assign a PriorityClass, know the kill switch.
Getting startedPilot runbookA 30/60/90 evaluation plan: observe-only, tiers on one pool, density, consolidation dry-run — each stage reversible, each with exit numbers.
ConceptsArchitectureThe three layers (node enforcement, cluster intelligence, management plane), what talks to what, and why everything stays in-cluster.
Capabilities
ConceptsKernel-level QoS enforcementFive tiers derived from Kubernetes PriorityClasses, mapped to kernel scheduler layers sized from real resource requests.
ConceptsWorkload thread profilesPer-thread-group scheduling inside a single pod: builtin profiles, file-based custom profiles, and the training pipeline.
PlatformSLO-safe density & overcommitThe density-aware scheduler plugin, the reversible overcommit webhook, and complement mode under Karpenter or Cast AI.
PlatformFleet management planeExplorer, logs, manifests and diffs, audit-logged actions, RBAC, the savings methodology, and the multi-cluster hub.
PlatformObservability & rightsizing/metrics, /observe, Grafana dashboards, the placement linter, kernel trace capture, and the thread-aware rightsizer.
Operations & reference
OperationsSafety & rollbackThe kernel revert contract, the safe-mode kill switch, measured failover, churn cost, and the cpu.max caveat.
OperationsOperationsCanary upgrades, the rollback runbook, monitoring the agent, and sizing guidance for protected tiers.
OperationsPlatform supportThe verified matrix for GKE, EKS, and on-prem — including the platforms that are not supported and why — kernel requirements, and ARM64 status.
ReferenceSupportChannels for each tier, what to include in a technical report, response expectations, and where security reports go.
ReferenceFAQHPA/VPA coexistence, CPU limits under sched_ext, air-gap operation, licensing and source access, and more.