In a Spark streaming application that processes events from Azure Event Hubs, which method is best for ensuring stateful computation across batches for windowed aggregations?