Hello all,
Currently, I examine the effects of stream partitioning on performance for
simple state-full scenarios.
My toy application for the rest of my question will be the following: A stream
of non-negative integers, each one annotated with a timestamp, and the goal is
to get the top-10 most frequent non-negative integers on tumbling windows of 10
seconds. In other words, my input is a stream of tuples with two fields,
Tuple2<Long, Integer>(timestamp, key), where key is the non-negative integer
value, and timestamp is used to assign each event to a window. The execution
plan I am considering is to have a first phase (Phase 1), where the stream is
partitioned and the partial aggregations are processed in parallel (set
parallelism to N > 1). Afterwards, the second phase (Phase 2) involves
gathering all partial aggregations on a single node (set parallelism to 1), and
calculate the full aggregation for each key, order the keys based on windowed
frequency and outputs the top-10 keys for each window.
As I mentioned earlier, my goal is to compare the performance of different
partitioning policies on this toy application. Initially, I want to compare
shuffle-grouping (round-robin) and hash-grouping and then move on to different
partitioning policies by using Flink's CustomPartitioner API. After reading
Flink's documentation, I managed to develop the toy application using
hash-partitioning. Below, I present the different parts of my code:
// Phase 0: input setup
DataStream<Tuple3<Long, Integer, Integer>> stream = env.fromCollection(...)
.assignTimestampsAndWatermarks(new
AscendingTimestampExtractor<Tuple2<Long, Integer>>() {
@Override
public long extractAscendingTimestamp(Tuple2<Long, Integer>
event) { return event.f0; }
}).map( (Tuple2<Long, Integer> e) -> new Tuple3<Long, Integer,
Integer>(e.f0, e.f1, 1));
On Phase 0, I collect the input stream, from an in-memory list, define the
event timestamp which will be used for windowing, and extend each event with a
value of 1 for calculating the appearance of each number on every window.
Afterwards, for the parallel Phase 1, I use hash partitioning by first using
.keyBy() operation on the key of each tuple (i.e., field 1), followed by a
.window() operation, to assign each tuple on a different window, and end with a
.sum(). My code for (parallel) Phase 1 is the following:
// Phase 1: parallel partial sum, with a parallelism of N (N > 1)
DataStream<Tuple3<Long, Integer, Integer> phaseOne =
stream.keyBy(1).window(TumblingEventTimeWindows.of(Time.seconds(10)).sum(2).setParallelism(N);
Moving on to Phase 2, to aggregate all partial results of a single window in
one operator for producing the full aggregation, ordering based on frequency,
and return the top-10 keys, I have the following:
// Phase 2: serial full aggregation and ordering, with a parallelism of 1
DataStream<String> phaseTwo = phaseOne
.windowAll(TumblingEventTimeWindows.of(Time.seconds(10))
.apply(new AllWindowsFunction<Tuple3<Long, Integer, Integer>,
String, TimeWindow>() {
@Override
public void apply(TimeWindow window, Iterable<Tuple3<Long,
Integer, Integer>> values, Collector<String> out) throws Exception {
...
List<Integer> topTenValues = ...;
StringBuilder strBuilder = new StringBuilder();
for (Integer t : topTenValues)
strBuilder.append(Integer.toString(t) + ",");
out.collect(strBuilder.toString());
});
The previous code makes use of hash-partitioning for its parallel phase. From
what I understand, Flink allows the .window() operation only on a KeyedStream.
Furthermore, the .customPartition() method transforms a DataStream to a
DataStream (and the same is true for .shuffle() which round-robins events).
Therefore, I am confused on how I can use a shuffle policy with windows. One
Idea that came to me is to provide an irrelevant field on the .keyBy() method,
or define my own KeySelector<IN, KEY> that will simulate shuffle grouping
through key generation. Unfortunately, I have two concerns regarding the
previous alternatives: For the keyBy() approach, I need to control the internal
hashing mechanisms, which entails cherry-picking fields on different workloads
and performing an exhaustive search on the behavior of different random fields
(not practical). For the KeySelector<IN, KEY>approach, I need to maintain state
among different calls of getKey(), which (as far as I know) is not offered by
the KeySelector<IN, KEY> interface and I do not want to rely on external state
that will lead to additional overhead. Therefore, my first question is how will
I be able to effectively use round-robin grouping with windows on my toy
application?
The bigger point I am trying to address revolves around custom partitioning
policies and windows in general. My understanding is that the benefit of a
custom partitioning policy is to have the ability to control the partitioning
process based on a pre-defined set of resources (e.g., partitions, task slots
etc.). Hence, I am confused on how I would be able to use partitionCustom()
followed by .window() on the (parallel) phase one, to test the performance of
different execution plans (i.e., partitioning policies).
I apologize for the long question, but I believe that I had to provide enough
details for the points/questions I currently have (highlighted with bold).
Thank you very much for your time.
Kind Regards,
Nikos R. Katsipoulakis,
Department of Computer Science
University of Pittsburgh