Details
-
Bug
-
Status: Resolved
-
P3
-
Resolution: Not A Problem
-
None
-
None
Description
We have being experiencing OOM errors on GroupByKey in batch mode in both Spark runners even if behind the woods spark spills data to disk in such cases: taking a look at the translation in the two runners, it might be due to using ReduceFnRunner for merging windows in GBK translation. ReduceFnRunner.processElements expects to have all elements to merge the windows between each other.:
RDD spark runner:
structured streaming spark: runner: https://github.com/apache/beam/blob/752798e3b0e6911fa84f8d138dacccdb6fcc85ef/runners/spark/2/src/main/java/org/apache/beam/runners/spark/structuredstreaming/translation/batch/functions/GroupAlsoByWindowViaOutputBufferFn.java#L74
Even replacing the Iterable with an Iterator in ReduceFnRunner to avoid materialization does not work because deep in ReduceFnRunner.processElements(), the collection is iterated twice.
It could be better to do what flink runner does and translate GBK as CombinePerKey with a Concatenate combine fn and thus avoid elements materialization.