Scikit‑learn Estimator Caching ↔ Joblib Memory: When Is Cached Result Invalidated by Code Changes?

Answer

The cache is invalidated only when the value that joblib.Memory uses to build its cache key changes. Editing the body of an estimator’s fit (or any other) method, or changing an imported utility function, does not alter the key, so the previously cached result is returned.

Confirmed facts

• joblib.Memory constructs the key from the estimator’s class name, its __module__ (import path), and the exact hyper‑parameter values passed to the constructor.
• The source code of methods is not hashed; therefore a change in method logic leaves the key unchanged.
• Changing an external utility that the estimator imports does not affect the key unless the import path of the estimator itself (or the utility’s module) changes.

Likely explanation

If you modify only the fit method body while keeping the class name, module, and hyper‑parameters identical, the cache key stays the same and joblib returns the stale result. The same holds for changes to helper functions unless the estimator’s import path changes. When GridSearchCV runs with n_jobs>1 using the default loky backend, each worker shares the same Memory directory; workers may read outdated entries or overwrite them inconsistently, producing non‑deterministic scores.

Steps to force a refresh after code changes

1. Clear the Memory base directory (e.g., rm -rf cachedir), or
2. Alter something that participates in the key – rename the estimator class, move it to a different module, or add a dummy hyper‑parameter (e.g., dummy_tag=1) – so the key differs, or
3. Disable caching for the decorator by setting memory=None or using backend='threading' with location=None.

If you are running GridSearchCV with n_jobs>1 and the default loky backend, you must ensure all worker processes see the same code version; otherwise some may hit stale cache while others recompute. Clarify: Are you using the loky backend with parallel jobs? Knowing this determines whether you need to restart the worker pool or simply clear the cache directory.