System map¶
ExaMLOps is easiest to understand as a transit map. Five kinds of work travel through the platform, and each has its own line. The lines share stations: Ray Serve is a stop on the data line and on the control line, and the control plane is where the control line and the decision line meet.
Open any station for what it does, the commands that operate it and the guide that explains it. Pick a line in the legend to follow it on its own.
The five lines¶
Data line¶
A prediction: from a client, across the bus, through batching and routing to a model on Ray Serve, and back. A chat request takes the same line through the model gateway to a vLLM server.
Control line¶
A retrain: from a trigger to the control plane, through Prefect and a scheduler, into the registry and back into serving.
Decision line¶
A person decides: approvals, promotions, traffic splits, the autopilot switch and cluster admission.
Compute line¶
A cluster job: discover a cluster, approve it, place the job, run it and account for its GPU-hours.
Signal line¶
Telemetry and evidence: metrics, logs, traces, alerts and the hash-chained audit trail.
Under construction¶
Dotted track marks work that is designed but not built. Every dotted segment is listed on the roadmap.
Where to go next¶
| You want to… | Read |
|---|---|
| Know what each service does and which port it listens on | Services and their jobs |
| Serve an LLM and put keys, budgets and guardrails in front of it | Follow an LLM request |
| Find the command for a task | Every capability — all exa commands, searchable |
| Know what is built and what is next | Roadmap |
| Run it yourself | Quick Start |