feat(backend): surveille la derive du modele de prevision
EC06 attendait une reponse a « comment savez-vous que le modele se degrade ? ». Elle n'existait nulle part : `docs/architecture/00-vue-ensemble.md` et `docs/ML-START.md` le disaient tous les deux. Le calcul vit dans le backend, et `ml/` ne gagne pas une ligne. Trois raisons : `prediction` n'est pas dans le perimetre de lecture que `ML_DATABASE_URL` vise (ADR 0003 et ML-START le bornent a `reading` et `site`) ; l'alignement prevu contre realise existe deja une fois ici, dans `AlertService._detect_anomaly`, et le dupliquer en SQL brut creerait une seconde source de verite, ce que l'ADR 0006 refuse ; et FastAPI continue de ne jamais faire tourner LightGBM. Ce qui est mesure : la jointure `prediction` x `reading` sur `(site_id, target_at)`, avec un `DISTINCT ON` des deux cotes. Les runs de scoring s'empilent volontairement, et `uq_reading_source` autorise deux lectures au meme instant quand la source differe : sans ce dedoublonnage, la meme heure pesait plusieurs fois dans la moyenne. La fenetre est fermee a droite par un delai de grace, sinon la derniere heure, dont le realise n'est pas encore ingere, ferait chuter la couverture a chaque execution. Le verdict a trois valeurs, pas deux : avec trois points on ne declare pas une derive, on dit qu'on ne sait pas. La comparaison se fait entre deux fenetres vives de meme duree, jamais contre la metrique loguee a l'entrainement : celle-ci mesure un backtest a meteo connue, le scoring prevoit une heure dont la meteo ne l'est pas. `drift_report` porte une ligne par site plus une ligne globale, que `site_id` a NULL designe. L'idempotence passe par un index a `coalesce` et non par une contrainte d'unicite, sans quoi deux lignes globales ne seraient jamais egales.
This commit is contained in:
Johan LEROY
@@ -0,0 +1,187 @@
|
||||
from datetime import UTC, datetime, timedelta
|
||||
|
||||
import pytest
|
||||
from sqlalchemy.dialects import postgresql
|
||||
from sqlalchemy.ext.asyncio import AsyncSession
|
||||
from sqlalchemy.sql import ClauseElement
|
||||
|
||||
from app.repositories.drift import (
|
||||
DriftRepository,
|
||||
NouveauRapportDerive,
|
||||
_lectures_retenues,
|
||||
_predictions_retenues,
|
||||
)
|
||||
from tests.repositories.test_prediction import creer_prediction
|
||||
from tests.repositories.test_reading import creer_lecture
|
||||
from tests.repositories.test_site import creer as creer_site
|
||||
|
||||
DEBUT = datetime(2026, 9, 15, tzinfo=UTC)
|
||||
FIN = datetime(2026, 9, 22, tzinfo=UTC)
|
||||
CIBLE = datetime(2026, 9, 16, 12, tzinfo=UTC)
|
||||
|
||||
|
||||
def sql(requete: ClauseElement) -> str:
|
||||
return str(requete.compile(dialect=postgresql.dialect())) # type: ignore[no-untyped-call]
|
||||
|
||||
|
||||
def rapport(**remplacements: object) -> NouveauRapportDerive:
|
||||
defauts: dict[str, object] = {
|
||||
"site_id": None,
|
||||
"window_start": DEBUT,
|
||||
"window_end": FIN,
|
||||
"reference_start": None,
|
||||
"reference_end": None,
|
||||
"n_observations": 10,
|
||||
"mae": 1.0,
|
||||
"mape": 5.0,
|
||||
"bias": 0.1,
|
||||
"reference_mae": None,
|
||||
"coverage_ratio": 1.0,
|
||||
"insufficient_data_ratio": 0.0,
|
||||
"model_references": ["lightgbm-aaa"],
|
||||
"status": "stable",
|
||||
"reason": None,
|
||||
}
|
||||
return NouveauRapportDerive(**{**defauts, **remplacements}) # type: ignore[arg-type]
|
||||
|
||||
|
||||
def test_predictions_keep_one_row_per_site_and_target_in_sql() -> None:
|
||||
requete = sql(_predictions_retenues(debut=DEBUT, fin=FIN, site_id=None).element)
|
||||
|
||||
assert "DISTINCT ON (prediction.site_id, prediction.target_at)" in requete
|
||||
assert "prediction.prediction_id DESC" in requete
|
||||
|
||||
|
||||
def test_readings_keep_one_row_per_site_and_instant_in_sql() -> None:
|
||||
requete = sql(_lectures_retenues(debut=DEBUT, fin=FIN, site_id=None).element)
|
||||
|
||||
assert "DISTINCT ON (reading.site_id, reading.timestamp)" in requete
|
||||
assert "reading.reading_id DESC" in requete
|
||||
|
||||
|
||||
def test_predictions_restrict_themselves_to_the_requested_site_in_sql() -> None:
|
||||
requete = sql(_predictions_retenues(debut=DEBUT, fin=FIN, site_id="SITE001").element)
|
||||
|
||||
assert requete.count("prediction.site_id = ") == 1
|
||||
|
||||
|
||||
def test_readings_ignore_a_missing_consumption_in_sql() -> None:
|
||||
requete = sql(_lectures_retenues(debut=DEBUT, fin=FIN, site_id=None).element)
|
||||
|
||||
assert "reading.consumption_kwh IS NOT NULL" in requete
|
||||
|
||||
|
||||
@pytest.mark.integration
|
||||
async def test_repository_pairs_a_prediction_with_the_reading_of_the_same_instant(
|
||||
session: AsyncSession,
|
||||
) -> None:
|
||||
site = await creer_site(session)
|
||||
await creer_prediction(session, site_id=site.site_id, target_at=CIBLE, predicted_value=12.0)
|
||||
await creer_lecture(session, site_id=site.site_id, timestamp=CIBLE, consumption_kwh=10.0)
|
||||
|
||||
paires = await DriftRepository(session).paires(debut=DEBUT, fin=FIN, site_id=site.site_id)
|
||||
await session.rollback()
|
||||
|
||||
assert [(p.predicted_value, p.actual_value) for p in paires] == [(12.0, 10.0)]
|
||||
|
||||
|
||||
@pytest.mark.integration
|
||||
async def test_repository_keeps_the_latest_run_when_several_predictions_share_a_target(
|
||||
session: AsyncSession,
|
||||
) -> None:
|
||||
site = await creer_site(session)
|
||||
await creer_prediction(session, site_id=site.site_id, target_at=CIBLE, predicted_value=12.0)
|
||||
await creer_prediction(session, site_id=site.site_id, target_at=CIBLE, predicted_value=99.0)
|
||||
await creer_lecture(session, site_id=site.site_id, timestamp=CIBLE, consumption_kwh=10.0)
|
||||
|
||||
paires = await DriftRepository(session).paires(debut=DEBUT, fin=FIN, site_id=site.site_id)
|
||||
await session.rollback()
|
||||
|
||||
assert [p.predicted_value for p in paires] == [99.0]
|
||||
|
||||
|
||||
@pytest.mark.integration
|
||||
async def test_repository_keeps_one_reading_per_instant_when_two_sources_wrote_the_same_hour(
|
||||
session: AsyncSession,
|
||||
) -> None:
|
||||
site = await creer_site(session)
|
||||
await creer_prediction(session, site_id=site.site_id, target_at=CIBLE, predicted_value=12.0)
|
||||
await creer_lecture(
|
||||
session, site_id=site.site_id, timestamp=CIBLE, source="api_current", consumption_kwh=10.0
|
||||
)
|
||||
await creer_lecture(
|
||||
session, site_id=site.site_id, timestamp=CIBLE, source="api_history", consumption_kwh=20.0
|
||||
)
|
||||
|
||||
paires = await DriftRepository(session).paires(debut=DEBUT, fin=FIN, site_id=site.site_id)
|
||||
await session.rollback()
|
||||
|
||||
assert [p.actual_value for p in paires] == [20.0]
|
||||
|
||||
|
||||
@pytest.mark.integration
|
||||
async def test_repository_excludes_an_insufficient_data_prediction_from_the_pairs(
|
||||
session: AsyncSession,
|
||||
) -> None:
|
||||
site = await creer_site(session)
|
||||
await creer_prediction(
|
||||
session,
|
||||
site_id=site.site_id,
|
||||
target_at=CIBLE,
|
||||
predicted_value=None,
|
||||
status="insufficient_data",
|
||||
failure_reason="historique trop court",
|
||||
)
|
||||
await creer_lecture(session, site_id=site.site_id, timestamp=CIBLE, consumption_kwh=10.0)
|
||||
|
||||
depot = DriftRepository(session)
|
||||
paires = await depot.paires(debut=DEBUT, fin=FIN, site_id=site.site_id)
|
||||
comptages = await depot.comptages(debut=DEBUT, fin=FIN, site_id=site.site_id)
|
||||
await session.rollback()
|
||||
|
||||
assert paires == []
|
||||
assert [(c.status, c.nombre) for c in comptages] == [("insufficient_data", 1)]
|
||||
|
||||
|
||||
@pytest.mark.integration
|
||||
async def test_repository_excludes_a_target_outside_the_window(session: AsyncSession) -> None:
|
||||
site = await creer_site(session)
|
||||
hors_fenetre = FIN + timedelta(hours=1)
|
||||
await creer_prediction(
|
||||
session, site_id=site.site_id, target_at=hors_fenetre, predicted_value=12.0
|
||||
)
|
||||
await creer_lecture(session, site_id=site.site_id, timestamp=hors_fenetre, consumption_kwh=10.0)
|
||||
|
||||
paires = await DriftRepository(session).paires(debut=DEBUT, fin=FIN, site_id=site.site_id)
|
||||
await session.rollback()
|
||||
|
||||
assert paires == []
|
||||
|
||||
|
||||
@pytest.mark.integration
|
||||
async def test_repository_reads_back_the_global_report_it_wrote(session: AsyncSession) -> None:
|
||||
depot = DriftRepository(session)
|
||||
fenetre = datetime(2035, 3, 1, tzinfo=UTC)
|
||||
|
||||
ecrites = await depot.enregistre([rapport(window_end=fenetre)])
|
||||
derniers = await depot.derniers()
|
||||
globaux = [r for r in derniers if r.site_id is None and r.window_end == fenetre]
|
||||
await session.rollback()
|
||||
|
||||
assert ecrites == 1
|
||||
assert len(globaux) == 1
|
||||
|
||||
|
||||
@pytest.mark.integration
|
||||
async def test_repository_ignores_a_second_report_for_the_same_window_and_site(
|
||||
session: AsyncSession,
|
||||
) -> None:
|
||||
depot = DriftRepository(session)
|
||||
fenetre = datetime(2035, 4, 1, tzinfo=UTC)
|
||||
|
||||
premiere = await depot.enregistre([rapport(window_end=fenetre)])
|
||||
seconde = await depot.enregistre([rapport(window_end=fenetre, status="derive", reason="x")])
|
||||
await session.rollback()
|
||||
|
||||
assert premiere == 1
|
||||
assert seconde == 0
|
||||
Reference in New Issue
Block a user