Architecture
Chamber operates as a SaaS platform with lightweight agents deployed in your Kubernetes clusters:How It Works
- Agents report capacity: Lightweight agents in each cluster report available GPU resources
- Chamber schedules workloads: When teams submit workloads, Chamber decides where to run them
- Agents execute locally: Agents create and manage workloads on your clusters
- Dashboard provides visibility: View utilization, manage teams, and track workloads
The agent only makes outbound connections to Chamber. No inbound firewall rules are required.
Data Model
Chamber’s data model reflects enterprise organizational structures:Key Relationships
| Relationship | Description |
|---|---|
| Organization → Teams | An org contains a hierarchy of teams |
| Organization → Pools | An org can have multiple capacity pools |
| Team ↔ Pool | Many-to-many via reservations |
| Team → Workloads | Workloads belong to exactly one team |
Workload Classes
| Class | Capacity Source | Preemptible | Use Case |
|---|---|---|---|
| Reserved | Team’s reservation | No | Production training, SLA workloads |
| Elastic | Reserved capacity first, then idle pool | Yes | Experiments, batch processing, dev work |
Elastic workloads maximize cluster utilization by using capacity that would otherwise sit idle, while still guaranteeing reserved workloads always have resources.
Next Steps
Teams
Understand the organizational hierarchy
Capacity Pools
Learn about GPU resource pools
Reservations
How capacity is allocated to teams
Scheduling
Deep dive into the scheduler