Scaling decision trees in Spark involves partitioning the data across multiple nodes and using parallel processing to train the trees efficiently. Spark's MLlib provides optimized algorithms for feature selection and node splitting, which contribute to the overall performance and scalability of decision tree models.