Hi Bill, Flink uses pipelined data shipping by default. If you have a program like: source -> map -> reduce -> sink, the mapper will immediately start to shuffle data over the network to the reducer. The reducer collects that data and starts to sort it in batches. When the mapper is done and the reducer is done with the sorting, the sorted batches are merged and a sorted stream is fed into the reduce function. So, the shuffling and sorting happens while the mapper is running, but the reduce function can only be applied after the last mapper has finished.
I am not aware of benchmarks that compare Flink's and Spark's in-memory operations, but this blog post compares the performance of sorting binary data in managed memory (like Flink) to naive sorting on the heap (Spark might do it differently!). --> http://flink.apache.org/news/2015/05/11/Juggling-with-Bits-and-Bytes.html Fabian 2015-07-17 16:07 GMT+02:00 Bill Sparks <jspa...@cray.com>: > Does flink require all the map tasks to finish before the reducers can > proceed like Spark, or can the reducer operations start before all the > mappers have finished like the older Hadoop mapreduce. > > Also my understanding is that flink manages it's own heap, do you/we > have a sense of the performance impact of this as compared to say …. Spark > where it's all in the JVM. > > Regards, > Bill. > > -- > Jonathan (Bill) Sparks > Software Architecture > Cray Inc. >
