Global Control Plane
The Global Control Plane provides one API to manage independent LiteLLM deployments across regions and clouds. Each deployment is a data plane that serves inference requests and keeps its own keys, teams, budgets, PostgreSQL database, and Redis.
This feature requires a LiteLLM Enterprise license. Start a free 30-day trial or book a demo. See what Enterprise includes.
Use cases
| Need | How the control plane helps |
|---|---|
| Isolate data for compliance | Keep each data plane's database, Redis, keys, and credentials in a separate regional environment. The control plane receives management and usage data. |
| Manage resources through one endpoint | Read and update keys, users, teams, models, budgets, and other supported resources on a selected data plane. |
| Compare usage across deployments | Aggregate spend, requests, and tokens. Group results by data plane, model, provider, team, or API key. |
| Give operators limited access | Assign control-plane roles and restrict callers to named data planes. Each data plane also checks the caller's local permissions. |
How it works
An operator or internal portal sends an authenticated request to the control plane. The control plane checks the caller's role and data-plane access, obtains a credential for the selected data plane, and calls its management API over HTTPS. The data plane checks its own permissions before it returns a result.
Applications send inference requests directly to their regional data plane. A control-plane outage stops central management and reporting. Data planes continue serving inference using their own configuration and dependencies.
| Component | Responsibility | State |
|---|---|---|
| Control plane | Authentication, access rules, management requests, runtime registry and role writes, fleet inventory, usage aggregation, and audit export | Startup registry and roles, optionally extended by runtime entries in its own dedicated PostgreSQL database and Redis stream. Recent audit records and login throttles remain in process memory. |
| Data plane | Inference, local resource management, model routing, and enforcement of keys, teams, and budgets | Its own PostgreSQL, Redis, provider credentials, and master key. |
PostgreSQL and Redis are not required for configuration-only operation. Enable runtime configuration to manage registry entries and roles through the API and propagate them across control-plane replicas.
The control plane does not copy keys, synchronize data plane configuration, enforce a shared global budget, or fail over inference between data planes. Runtime configuration sharing only covers the control plane's own registry and roles. Use Multi-Region Deployment when regional gateways must share one database and the same keys.
This implementation exposes REST APIs and interactive API documentation. It does not serve the Admin UI.
Documentation
| Topic | Contents |
|---|---|
| Reference Architectures | AWS, Azure, GCP, and hybrid cloud diagrams with regions, private networks, and management connections. |
| Configuration | Settings and defaults for the control plane, data plane authentication, and audit destinations. |
| API Reference | Endpoints, scopes, parameters, and responses. |
| UI Deployment Selector | The existing Admin UI setup using worker_registry. |