Question
A production Databricks job uses `MlflowSparkStudy` to run Optuna trials on executors. Trials fail intermittently because some workers cannot import the same training library version used by the driver. What should be fixed before increasing `n_jobs`?