The recommended approach for loading small datasets (~10MB or less) in Hyperopt with SparkTrials is to load the dataset on the driver and call it directly from the objective function. This method is preferred because:

• Overhead Reduction: Broadcasting and loading data from DBFS involve additional overhead that can be significant for small datasets. These methods are more suitable for larger datasets.
• Driver Memory Efficiency: Small datasets can typically be loaded efficiently into the driver's memory without causing memory issues.
• Faster Access: Accessing data directly from the driver within the objective function is often faster than retrieving it from broadcasted variables or DBFS.
• Simplified Code: This approach leads to simpler code, as you don't need to deal with explicit broadcasting or DBFS interactions.

For larger datasets, broadcasting or loading from DBFS might be necessary for efficient distribution across workers. Always consider the dataset size and the trade-offs between different loading approaches to optimize performance and resource usage in Hyperopt with SparkTrials.