Flink is pretty neat and it didn't feel like an "all or nothing" monolithic chunk when I played with it.

The page based inner loop makes it very predictable when it does data operations - they were doing this early last year and this mechanism made it very cpu cache friendly & didn't trigger the massive GC pauses.

I was playing with Flink a bit earlier, because Flink can be integrated into Tez, so that Tez could do DAG scheduling while Flink ran with its inner loops on turbo.

That inner loop can edge out even hand-written java code I wrote for page-rank (delta iterations are nice).

Great post detailing how Flink is managing data within the JVM and implementing internal operators (hashing, sorting, ..) working with that serialized data. I also like the performance analysis in the post.