Watch this video on YouTube
Why should SparkTrials not be used on autoscaling clusters, and what issue can arise in such cases?