Understanding Matyan

Matyan is built around a few core ideas:

• Run isolation — Each training run is a logical unit; data is associated with a run and sent to a central backend (no local repo).
• Scalability — The backend is stateless and reads from FoundationDB; ingestion goes through the frontier and Kafka so you can run many parallel experiments and many workers.
• Flexibility — The UI and MatyanQL let you filter, group, and compare runs and metrics.

Matyan is made of the backend, frontier, workers, client, and UI. See Architecture for their roles and data flow. You run the backend, frontier, and workers as separate services (e.g. Docker Compose or Kubernetes). See Getting started to run them locally.