Github user jkbradley commented on the issue: https://github.com/apache/spark/pull/19904 Strong +1 for unpersisting the data at the end. In the long-term, I don't think we'll even cache the training and validation datasets. Our caching of the training & validation datasets is a temporary hack to get around the issue that we don't have a DataFrame k-fold splitting method. Our current workaround is to go down to RDDs: DataFrame -> RDD -> k-fold split -> DataFrame, and as I recall, we cache to lower the SerDe costs in these conversions. Once we have a k-fold split method for DataFrames, we can just cache the original (full) dataset and not the k splits.
--- --------------------------------------------------------------------- To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org For additional commands, e-mail: reviews-h...@spark.apache.org