MLOps Pipeline

deployml provisions a complete end-to-end MLOps pipeline on GCP. Here's how the components fit together.

Architecture

                        ┌─────────────────────────────────────────────┐
                        │                   GCP (Cloud Run)            │
                        │                                              │
  Your Code ──────────▶ │  MLflow          FastAPI         Grafana     │
  (training,            │  (tracking +     (model          (monitoring │
   serving,             │   registry +     serving)        dashboard)  │
   monitoring)          │   artifacts)                                  │
                        │       │               │               │       │
                        │       ▼               ▼               ▼       │
                        │  Cloud SQL       BigQuery          BigQuery   │
                        │  (Postgres)      mlops dataset     metrics DB │
                        └─────────────────────────────────────────────┘

Components

MLflow

Handles experiment tracking, artifact storage, and model registry. All three share a single Cloud Run service backed by Cloud SQL for metadata and a GCS bucket for model files.

• Log experiments and compare runs
• Store model artifacts in GCS
• Register and version models
• Promote models through stages (Staging → Production)

FastAPI

Serves predictions from the latest registered model in MLflow. Pulls the model on startup — no redeployment needed when you register a new model version.

• GET /health — health check
• POST /predict — run inference

Grafana

Monitoring dashboard connected to the metrics Postgres database. Use it to visualize model performance, drift metrics, and prediction volume over time.

BigQuery (mlops dataset)

Four tables are created automatically on deploy:

Table	What goes in it
offline_features	Precomputed features used for training and serving
predictions	Model outputs logged when FastAPI serves a request
ground_truth	Actual outcomes matched back to predictions by ID
drift_metrics	Aggregated statistics for tracking model drift

Data Flow

Raw data
   │
   ▼
offline_features (BigQuery)
   │
   ├──▶ Training script ──▶ MLflow (log experiment) ──▶ Model Registry
   │
   └──▶ FastAPI (serve) ──▶ predictions (BigQuery)
                                    │
              ground_truth (BigQuery)┘
                                    │
                                    ▼
                          drift_metrics (BigQuery)
                                    │
                                    ▼
                              Grafana dashboard