v2.4.0

Highlights

• Backtest specifications: new endpoints for defining backtests and running them across multiple models, with parameter defaults and metadata exposed, and target column selection for stored datasets.
• Marketplace model management via chap-admin: chap install moved into chap-admin, which now installs, updates (update-all to stable versions) and uninstalls marketplace models and registers them from the running service.
• Alerts: alert policies and alerts are persisted, behind a release gate.
• Prediction setups: prediction-setup API routes are promoted out of experimental, and /v1/analytics/make-prediction is deprecated in their favour.

Breaking changes

• chap install has moved to chap-admin (#637).
• A changed model revision under an existing version label is now refused (#616).

Features

• Hyperparameter optimisation metadata on evaluations (#618)
• Metric unit and target exposed in the metrics API (#603)
• Configured model health status exposed, worked out per name and version (#656, #646)
• Job failure reason reported in GET /v1/jobs (#654)
• Download of original job requests (#625)
• dryRun for make-dataset to check data before import (#644)
• Datasets created manually are flagged (#639)
• Prediction setup id included in prediction responses (#607)
• Model output streamed to the debug log as it arrives (#587)
• valid_hpo_objective replaced with a proper_scoring_rule flag (#641)

Bug Fixes

• Honor the requested stride for weekly backtest imports (#692)
• Keep /docs, /redoc and /openapi.json public when CHAP_API_TOKEN is set (#688)
• Keep flat HPO metadata when loading an evaluation from file (#672)
• Remove the naive predictor's shared CSV round-trip (#678)
• Send chapkit user options flat and fail when the service drops them; tell converted values apart from dropped ones (#640, #658)
• Count a chapkit service's stored covariates in the unused-covariate warning (#657)
• Handle concurrent chapkit registrations racing on model template insert (#648)
• Make startup alembic upgrades persist and replay safely (#638)
• Record the train cutoff and honour n_retrain in dry runs (#590)
• Treat a redis connect timeout as redis being unavailable (#653)
• Disable the gunicorn control socket in the read-only container (#649)
• Create the worker data directories in Dockerfile.inla (#651)
• Ship the sanity-check-model example dataset in the package (#642)
• Make export-metrics --output-file a named flag (#643)
• Remove the ewars_plus n_periods override in run_backtest (#636)

Other

• EWARS dev overlay pinned to 1.0.4
• Dependency updates (pydantic, cyclopts 5, optuna 5, uvicorn, scikit-learn, altair, matplotlib, docker, and others)