luhenry edited a comment on pull request #32415: URL: https://github.com/apache/spark/pull/32415#issuecomment-832285023
Some runs with various values of `max_iter`: __100:__ ![image](https://user-images.githubusercontent.com/660779/117075564-4fe1c480-ad35-11eb-889a-6df72eecacbc.png) __300:__ ![image](https://user-images.githubusercontent.com/660779/117075724-96cfba00-ad35-11eb-9831-6c0640c735ea.png) __1,000:__ ![image](https://user-images.githubusercontent.com/660779/117075621-6be56600-ad35-11eb-81e2-a537936593cd.png) However, when playing with the parallelism provided by Spark, that is where I can easily reproduce the huge variation for `summary.logLikelihood`. __`pyspark.SparkContext('local[1]')`:__ ![image](https://user-images.githubusercontent.com/660779/117076248-5f154200-ad36-11eb-8c47-38676dabede2.png) __`pyspark.SparkContext('local[2]')`:__ ![image](https://user-images.githubusercontent.com/660779/117076284-705e4e80-ad36-11eb-8bd1-291b7bd2a21d.png) __`pyspark.SparkContext('local[3]')`:__ ![image](https://user-images.githubusercontent.com/660779/117076424-a6033780-ad36-11eb-88c3-be61700a45fd.png) __`pyspark.SparkContext('local[4]')`:__ ![image](https://user-images.githubusercontent.com/660779/117076684-15792700-ad37-11eb-88c2-5b6c70f4e9b4.png) That is all with Spark 3.1.1 and so unrelated to my changes. I would say that the flakiness observed in this PR is explained by this change as if it's a potential race in the implementation, changing the performance profile of the underlying operations can make the race more likely to happen. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org --------------------------------------------------------------------- To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org For additional commands, e-mail: reviews-h...@spark.apache.org