feat(etl,ml): orchestre l'entrainement et le scoring LightGBM via deux DAGs Airflow

docker-compose.yml

@@ -5,6 +5,32 @@
 
 name: enervision
 
+# Piege : LocalExecutor fait tourner les taches comme sous-processus du scheduler, jamais du
+# webserver. `airflow_ml_state` (modele entraine, magasin MLflow) n'a donc besoin d'etre monte
+# que sur `airflow-scheduler` en pratique, mais reste partage avec le webserver pour que ce
+# dernier puisse au besoin l'inspecter sans en devenir dependant.
+x-airflow-common: &airflow-common
+  build:
+    context: .
+    dockerfile: etl/airflow/Dockerfile
+  environment: &airflow-common-env
+    AIRFLOW__CORE__EXECUTOR: LocalExecutor
+    AIRFLOW__CORE__LOAD_EXAMPLES: "false"
+    AIRFLOW__CORE__FERNET_KEY: ${AIRFLOW_FERNET_KEY:?}
+    AIRFLOW__WEBSERVER__SECRET_KEY: ${AIRFLOW_WEBSERVER_SECRET_KEY:?}
+    AIRFLOW__DATABASE__SQL_ALCHEMY_CONN: postgresql+psycopg2://${POSTGRES_USER}:${POSTGRES_PASSWORD}@db:5432/airflow
+    # Role `enervision_ml` dedie pas encore provisionne (dette assumee, cf. ADR 0003/CLAUDE.md) :
+    # memes identifiants que le backend en attendant.
+    ML_DATABASE_URL: postgresql+psycopg://${POSTGRES_USER}:${POSTGRES_PASSWORD}@db:5432/${POSTGRES_DB}
+    MLFLOW_TRACKING_URI: sqlite:////opt/ml/state/mlflow.db
+  volumes:
+    - ./etl/airflow/dags:/opt/airflow/dags
+    - ./etl/airflow/plugins:/opt/airflow/plugins
+    - ./etl/airflow/include:/opt/airflow/include
+    - airflow_logs:/opt/airflow/logs
+    - airflow_ml_state:/opt/ml/state
+  restart: unless-stopped
+
 services:
   db:
     image: timescale/timescaledb-ha:pg17
@@ -19,6 +45,7 @@ services:
       - pgdata:/home/postgres/pgdata/data
       - ./db/init/100-extensions.sql:/docker-entrypoint-initdb.d/100-extensions.sql:ro
       - ./db/init/110-test-database.sql:/docker-entrypoint-initdb.d/110-test-database.sql:ro
+      - ./db/init/120-airflow-database.sql:/docker-entrypoint-initdb.d/120-airflow-database.sql:ro
     healthcheck:
       test: ["CMD-SHELL", "pg_isready -U $${POSTGRES_USER} -d $${POSTGRES_DB}"]
       interval: 10s
@@ -64,7 +91,54 @@ services:
     ports:
       - "${FRONTEND_PORT:-3000}:80"
     restart: unless-stopped
-        
+
+  # Conteneur unique, jamais redemarre : migre la base de metadonnees puis cree le premier compte
+  # (idempotent, `|| true` sur la creation qui echoue si le compte existe deja). `webserver` et
+  # `scheduler` attendent qu'il se termine avec succes avant de demarrer.
+  airflow-init:
+    <<: *airflow-common
+    restart: "no"
+    command:
+      - bash
+      - -c
+      - |
+        airflow db migrate
+        airflow users create \
+          --username "${AIRFLOW_ADMIN_USERNAME:-admin}" \
+          --password "${AIRFLOW_ADMIN_PASSWORD:?}" \
+          --firstname Admin \
+          --lastname EnerVision \
+          --role Admin \
+          --email "${AIRFLOW_ADMIN_EMAIL:-admin@enervision.fr}" \
+          || true
+
+  airflow-webserver:
+    <<: *airflow-common
+    command: webserver
+    ports:
+      - "${AIRFLOW_PORT:-8080}:8080"
+    depends_on:
+      db:
+        condition: service_healthy
+      airflow-init:
+        condition: service_completed_successfully
+    healthcheck:
+      test: ["CMD", "curl", "--fail", "http://localhost:8080/health"]
+      interval: 30s
+      timeout: 10s
+      retries: 5
+      start_period: 60s
+
+  airflow-scheduler:
+    <<: *airflow-common
+    command: scheduler
+    depends_on:
+      db:
+        condition: service_healthy
+      airflow-init:
+        condition: service_completed_successfully
 
 volumes:
   pgdata:
+  airflow_logs:
+  airflow_ml_state: