Databricks-ML-Pro Model Development Practice Question

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✗

  SparkTrials distributes trials across workers, and the NaN is caused by a serialization error when returning the loss from the worker to the driver.

  Why it's wrong here

  Serialization errors would raise exceptions and cause the trial to fail, not produce a NaN loss. SparkTrials uses Spark to parallelize trials, but the loss value is a simple float that serializes fine. If there were a serialization issue, the entire trial would be marked as failed and Hyperopt would not include it in the results. NaN losses typically originate from the objective function itself, not from the distribution mechanism.

• ✗

  The parallelism parameter is set too high, causing race conditions in the shared search space that corrupt the loss values.

  Why it's wrong here

  Hyperopt's SparkTrials handles parallelism safely by using a centralized search algorithm (TPE or random) on the driver and distributing trials to workers. There are no race conditions that would corrupt loss values; each trial is independent. Setting parallelism too high might slow down the search or cause resource contention, but it would not produce NaN losses. The NaN issue is almost always due to the objective function's return value.

• ✓

  The objective function returns a NaN when the model fails to converge, and Hyperopt treats NaN as a valid loss, potentially selecting a failed trial as best.

  Why this is correct

  Hyperopt does not automatically filter out NaN losses; it compares them numerically, and NaN comparisons can lead to unpredictable selection. If the objective function returns NaN due to convergence failure or invalid parameters, those trials can be incorrectly considered as having a low loss (since NaN comparisons are false, it may not update the best, but in some cases it can cause issues). The best practice is to return a large finite value or use a try-except to handle failures. This directly explains the poor best model.

• ✗

  The model's hyperparameters include a regularization parameter that, when set too high, causes the coefficients to become NaN, and Hyperopt does not handle this.

  Why it's wrong here

  While extreme regularization can cause numerical issues, scikit-learn typically raises a warning or error rather than producing NaN coefficients. Moreover, Hyperopt would still receive a finite loss if the model trains successfully. The scenario describes NaN losses, which are more commonly caused by the objective function returning NaN, for example due to a division by zero in a custom metric or an invalid parameter combination that causes the model to fail silently.