Skip to content

Runbooks

Operational procedures, incident playbooks, and governance runbooks for Private AI Platform Kit. Start with the incident-response index during an incident, or upgrade for the most common day-2 change.

Setup & day-2 operations

Runbook Purpose
Validation toolchain Install and verify the local/strict validation CLIs
Upgrade & rollback Promote a new release tag, test before promote, and roll back
GPU capacity GPU scheduling diagnostics and capacity validation
Policy-blocked deploy Diagnose and resolve Kyverno-blocked deployments
Agent workspaces Design and operate locked-down coding-agent workspaces
Tenant labs Onboard a tenant lab with quotas, egress, and isolation
API access API-key access model and rotation
External / managed stores Swap the bundled Redis, Qdrant, and Loki footprints for external/HA services

Runtime, RAG & guardrails

Runbook Purpose
RAG service Operate the RAG retrieval service
Vector RAG The Qdrant vector-store retrieval profile
Qdrant migration Collection migration with dry-run and rollback
Guardrails Prompt and secret-detection guardrails
Traceability sandbox Sandbox trace contract and request correlation
Budget controls Redis-backed sandbox spend/abuse budgets
Evaluation harness Run and interpret eval suites
End-user chat UI Put a self-hosted Open WebUI chat frontend in front of the gateway
Async batch API Operate the async Files + Batch API and the batch-processor worker

Governance & evidence

Runbook Purpose
Audit chain & SIEM forwarding Verify the tamper-evident audit hash chain, anchor its head, and forward receipts to a SIEM
Model governance Model lifecycle, promotion requests, and approval
Model provenance Artifact provenance and reproducible digest verification
Model drift monitoring Detect production model-quality drift via metrics and scheduled evals
Evidence pack Customer-facing evidence pack generation
Release gates Release-gate thresholds and strict evidence
SLO & error budget SLO objectives and error-budget review
Quota & chargeback Quota plans and chargeback labelling
Data retention Retention and privacy controls
Egress governance Approved-egress catalog and enforcement
Scorecard triage OpenSSF Scorecard finding triage

Incidents & resilience

Runbook Purpose
Incident response Severity tiers, escalation, and the incident index
Disaster recovery Single-cluster DR: RPO/RTO, restore order, operator-owned scope
Failure modes & degradation Consolidated dependency failure-mode and graceful-degradation matrix
Inference runtime incident Gateway / vLLM / Ollama runtime outages
Chaos drills Rollout/recovery and fault-injection drills
Restore drill Backup restore-tooling smoke and real data-recovery drill
Runtime threat detection Optional Falco/Tetragon detective layer for hijacked agents
OIDC / JWKS rotation Identity-provider key rotation

For the documentation map (setup, customer handoff, contracts), see docs/README.md.