Which Spark feature allows for recomputing data efficiently across distributed data sets?

The feature that allows for recomputing data efficiently across distributed datasets in Spark is caching. When you cache a dataset in Spark, it stores the data in memory after the first computation. This means any subsequent actions that need to use that data can access it from memory instead of recomputing it from scratch or fetching it from disk, which enhances performance significantly.

Caching is particularly useful in iterative algorithms and machine learning processes, where the same dataset is repeatedly accessed. By keeping the data in memory, Spark minimizes the overhead associated with data retrieval and ensures that the calculations can be performed more quickly.

Other options like resilience, parallelism, and aggregation do not specifically address the efficiency of recomputing data directly. Resilience refers to Spark's ability to handle node failures and maintain computations through lineage, parallelism pertains to how tasks are distributed across the cluster for concurrent processing, and aggregation deals with combining data from multiple sources. While all these features are crucial to Spark's operation, caching is the standout feature for enhancing the efficiency of recomputing data.