How to Check That a Registered Model Will Load Before Deploying¶

This guide shows you how to find out, before a deployment, whether an environment can load a registered forecaster. Use it when you upgrade yohou, scikit-learn or polars in the environment that runs scheduled forecasts.

Prerequisites¶

A forecaster registered with yohou_mlflow.log_model
The new environment installed, with access to the tracking server and registry

Steps¶

1. Run the check in the new environment¶

Run this with the new environment's packages, for example in a CI job that installs the upgraded lockfile:

import yohou_mlflow

report = yohou_mlflow.check_compatibility("models:/daily-demand/latest")
print(report)

When the model would load, the output is:

Loadable: a strict load_model would succeed.

When it would not, each reason is listed:

Not loadable:
  - A strict load would refuse the model, version mismatch: yohou: saved with 0.1.0a12, installed 0.1.0a13 (must match exactly)

The check reads only the model's metadata and the type list inside the saved file. It constructs nothing from the file, so it is safe on a model you have not reviewed.

If the forecaster needs extra trusted types, pass them as the second argument:

report = yohou_mlflow.check_compatibility("models:/daily-demand/latest", ["my_project.models.DampedRidge"])

2. Fail the deployment when the model will not load¶

if not report.loadable:
    raise SystemExit(str(report))

To check several versions, for example every version a rollback could use, run the check for each URI (models:/daily-demand/3, models:/daily-demand/4, ...).

3. Act on a version mismatch¶

If the check reports a mismatch, choose one:

Refit the forecaster in the new environment and register it, then deploy.
Keep the old versions of the packages it names in the deployed environment.
Load it anyway with load_model(..., strict=False), which only warns. Compare its predictions with the previous environment's before relying on them.

Troubleshooting¶

Problem: the report lists format version as the only problem: The model was saved by a newer yohou-mlflow. Upgrade yohou-mlflow in the new environment.
Problem: the report lists types outside the trust policy: Pass the types you trust, as in step 1, or see How to save a forecaster that contains your own estimator classes.