---

name: operations-and-security description: "Operate and secure Xinference deployment surfaces, including environment variables, auth, API keys, metrics, logging, Web UI static serving, and production deployment notes." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Operations and Security

Use this sub-skill for production-facing Xinference concerns: environment variables, persistent state, auth bootstrap, API keys, OIDC, audit/admin, metrics, logging, frontend static serving, IP and network exposure, Docker/Kubernetes deployment, and operational safety.

Use this for

• XINFERENCE_HOME, model cache/source, launch history, log, auth DB, and frontend dist path planning.
• Health checks, metrics exporter host/port, disabling metrics, request timeouts, concurrency, batching, and launch strategy knobs.
• Advanced auth, users, API keys, password reset/migration CLI entry points, JWT/encryption key persistence, OIDC, audit logs, and Elasticsearch-backed audit search.
• Allowed IPs, trusted proxies, permissive CORS caveats, and network exposure boundaries.
• Docker, Docker Compose, Kubernetes, and static Web UI deployment notes.

Route elsewhere

• Model launch flags, replicas, GPU placement, cache/model lifecycle CLI syntax, and distributed supervisor/worker command templates: serving-and-cli.
• Model families, optional backend extras, LoRA, virtualenv model dependencies, and custom model specs: models-and-backends.
• Python client calls, HTTP request bodies, OpenAI-compatible snippets, and streaming behavior: client-and-api.

Working pattern

1. Establish the deployment shape: local, Docker/Compose, Kubernetes, or a supervisor/worker cluster behind a proxy.
2. Choose persistent XINFERENCE_HOME and related DB/log/cache paths before enabling auth or relying on launch history.
3. Decide whether advanced auth stays enabled; if it does, plan admin bootstrap, API key issuance, token storage, and recovery procedures.
4. Configure network exposure separately from CORS. Treat allowed IPs and trusted proxies as access-control inputs.
5. Decide metrics/logging/audit policy before exposing /metrics or collecting audit records.
6. Route model-specific failures or request-body failures to sibling sub-skills.

Security rules

• Never invent or print real default credentials, JWT secrets, encryption keys, OIDC client secrets, model hub tokens, or API keys.
• Use placeholders such as <api-key>, <jwt-secret>, and <oidc-client-secret> in examples.
• Persist auth secrets and databases under stable storage before depending on users, permissions, or API keys across restarts.
• Treat permissive CORS as non-authoritative; use network boundaries, IP restrictions, and trusted proxy settings for real access control.
• Keep static environment matrices from reading live secret values unless the user explicitly asks for a local audit.

References

• Configuration for environment variables grouped by persistence, model source, health, launch, frontend, auth, and virtualenv.
• Security and auth for advanced auth, API keys, OIDC, reset/migration, and audit/admin guidance.
• Metrics and observability for Prometheus metrics, logging, and OpenTelemetry notes.
• Deployment notes for local, Docker, Compose, Kubernetes, frontend, and external-service caveats.
• Troubleshooting for auth failures, missing Web UI assets, disabled metrics, proxy/IP issues, and persistence mistakes.

Helper script

• render_env_matrix.py prints a static matrix of important environment variables without reading or revealing live values.