Overview

FoehnCast is a local-first ML system that helps you pick the best kiteboarding spot. This page shows how it's organized and where local vs. cloud differ.

The Core Design

The same Feature → Training → Inference pattern runs everywhere — locally or in the cloud.

flowchart TD classDef data fill:#eef2f7,stroke:#475569 classDef pipeline fill:#e6f4f1,stroke:#0f766e classDef app fill:#0f2530,stroke:#0f766e,color:#fff DATA["Weather forecasts"]:::data subgraph Build ["Build path (offline)"] direction TB FEAT["Feature pipeline"]:::pipeline CURATED["Curated features"]:::data TRAIN["Training pipeline"]:::pipeline REG["MLflow registry"]:::data FEAT --> CURATED --> TRAIN --> REG end subgraph Serve ["Serve path (online)"] direction TB API["FastAPI app"]:::app RANK["Ranked spots"]:::data API --> RANK end DATA --> FEAT CURATED --> API REG --> API

Local vs. Cloud

Both environments run the same Python code. The difference is where data lives and how things get scheduled.

flowchart TD classDef shared fill:#f1f5f9,stroke:#475569,stroke-dasharray: 5 5 classDef local fill:#e6f4f1,stroke:#0f766e classDef cloud fill:#fff4e6,stroke:#c2410c CORE["Same Python code"]:::shared subgraph Local ["Local (Docker Compose)"] direction LR LOPS["Airflow + MLflow + Prometheus"]:::local LDATA["MinIO + Feast emulator"]:::local end subgraph Cloud ["Cloud (GCP)"] direction LR CDATA["BigQuery + GCS"]:::cloud CAPI["Cloud Run (public)"]:::cloud end CORE --> Local CORE --> Cloud

Concern	Local	Cloud
Storage	MinIO (S3-compatible)	BigQuery + GCS
Orchestration	Airflow (Docker)	Cloud Workflows
Serving	FastAPI on localhost	Cloud Run
Monitoring	Prometheus + StatsD	Managed Prometheus
Cost	Free	~$8/mo (Cloud SQL)

Reading Order

1. Use Case and Data — what problem we're solving
2. Getting Started — run it locally
3. Architecture — the FTI split and how pipelines connect
4. Feature, Training, Inference — each pipeline in detail
5. Monitoring — metrics, drift detection, alerts
6. Cloud Architecture — the GCP deployment