How negativexq/ml-platform-infrastructure works
ML Platform is a self-hosted Kubernetes platform for training and serving models, LLMs, and user functions. Its main workflow runs from project-scoped management in the web UI and control API, through persisted desired state and reconciliation, to training or deployment on external systems; callers use a separate gateway for inference. The graph preserves documented access, model, rollout, monitoring, and UI capabilities. Some subsystem links are based on README architecture descriptions because their implementation files were not sampled.
Main entry points: api/app.py Control API · reconciliation Reconcilers · gateway/app.py Inference gateway · router.tsx Web UI
Components 24
User experience
- Web UI
controlplane/ui/web/src/router.tsx
Control plane
- Control API
controlplane/api/app.py - Identity and access
controlplane/application/identity.py - Projects and jobs
controlplane/application/projects.py - Pipeline definitions
controlplane/application/pipelines.py - Pipeline runs
controlplane/application/pipeline_runs.py - Models and versions
controlplane/application/models.py - Deployments and rollouts
controlplane/application/deployments.py - Gateway access
controlplane/application/api_access.py - Domain state
controlplane/domain/entities.py
Lifecycle execution
- SQL persistence
controlplane/persistence/sql.py - Reconcilers
controlplane/reconciliation
Platform integrations
- Workflow adapter
controlplane/adapters/workflow/argo.py - Kubernetes provisioner
controlplane/adapters/kubernetes/provisioner.py - MLflow tracking
controlplane/adapters/mlflow/tracking.py - KServe adapter
controlplane/adapters/serving/kserve.py
Serving and observability
- Inference gateway
controlplane/gateway/app.py - Model inference
app/inference.py - Monitoring
controlplane/observability/metrics.py
Other components
- People
- PostgreSQL
- API callers
- Serving endpoints
- Telemetry and dashboards
Connections 25
- People → Web UI: use
- Web UI → Control API: calls
- Control API → Identity and access: authenticates
controlplane/api/app.py - Control API → Projects and jobs: routes requests
controlplane/api/app.py - Control API → Pipeline definitions: routes requests
controlplane/api/app.py - Control API → Pipeline runs: routes requests
controlplane/api/app.py - Control API → Models and versions: manages
controlplane/api/app.py - Control API → Deployments and rollouts: manages
controlplane/api/app.py - Control API → Gateway access: manages
controlplane/api/app.py - Projects and jobs → Domain state: uses entities
controlplane/application/projects.py - Pipeline definitions → Domain state: creates definitions
controlplane/application/pipelines.py - Control API → SQL persistence: writes intent
- SQL persistence → PostgreSQL: stores state
- Reconcilers → PostgreSQL: read desired state
- Reconcilers → Kubernetes provisioner: provisions projects
- Reconcilers → Workflow adapter: executes pipelines
- Reconcilers → MLflow tracking: syncs model registry
- Reconcilers → KServe adapter: reconciles serving
- API callers → Inference gateway: send inference requests
- Inference gateway → Gateway access: checks keys
- Inference gateway → Serving endpoints: routes inference
- KServe adapter → Serving endpoints: manages endpoints
- Monitoring → Telemetry and dashboards: exports telemetry
- Pipeline runs → Workflow adapter: tracks execution
- Model inference → Serving endpoints: serves predictions