問題文
A data scientist wants to run an Optuna study across the executors of a multi-node Databricks cluster and have every trial recorded in one MLflow experiment, with as little custom plumbing as possible. Which arrangement achieves that?
選択肢
- Use the MLflow tracking server as the study's storage backend and launch the study with the Spark-backed study class so that trials are distributed to executors.
- Create one MLflow run per trial in advance, then let each trial pick an unused run by polling the experiment for runs without metrics, claiming the first empty one it finds.
- Wrap the objective function in a scalar user-defined function and call it from a SQL query so that Spark parallelizes the trials row by row.
- Run the study in a single Python process on the driver and start several background threads inside the objective function.