In Spark MLlib, the decision tree algorithm is scaled by leveraging the distributed nature of Spark. Each node in the cluster builds a separate decision tree, and the final model is an ensemble of these individual trees, such as a random forest or gradient boosted trees. This approach allows Spark to handle large datasets efficiently, as the computation is distributed across multiple nodes. Additionally, ensemble methods like random forests and gradient boosted trees can improve the accuracy and robustness of the decision tree model by combining the predictions of multiple trees.