Flink Streaming: From one window, lookup state in another window - apache-flink

Question

I have two streams:
Measurements
WhoMeasured (metadata about who took the measurement)
These are the case classes for them:
case class Measurement(var value: Int, var who_measured_id: Int)
case class WhoMeasured(var who_measured_id: Int, var name: String)
The Measurement stream has a lot of data. The WhoMeasured stream has little. In fact, for each who_measured_id in the WhoMeasured stream, only 1 name is relevant, so old elements can be discarded if one with the same who_measured_id arrives. This is essentially a HashTable that gets filled by the WhoMeasured stream.
In my custom window function
class WFunc extends WindowFunction[Measurement, Long, Int, TimeWindow] {
override def apply(key: Int, window: TimeWindow, input: Iterable[Measurement], out: Collector[Long]): Unit = {
// Here I need access to the WhoMeasured stream to get the name of the person who took a measurement
// The following two are equivalent since I keyed by who_measured_id
val name_who_measured = magic(key)
val name_who_measured = magic(input.head.who_measured_id)
}
}
This is my job. Now as you might see, there is something missing: The combination of the two streams.
val who_measured_stream = who_measured_source
.keyBy(w => w.who_measured_id)
.countWindow(1)
val measurement_stream = measurements_source
.keyBy(m => m.who_measured_id)
.timeWindow(Time.seconds(60), Time.seconds(5))
.apply(new WFunc)
So in essence this is sort of a lookup table that gets updated when new elements in the WhoMeasured stream arrive.
So the question is: How to achieve such a lookup from one WindowedStream into another?
Follow Up:
After implementing in the way Fabian suggested, the job always fails with some sort of serialization issue:
[info] Loading project definition from /home/jgroeger/Code/MeasurementJob/project
[info] Set current project to MeasurementJob (in build file:/home/jgroeger/Code/MeasurementJob/)
[info] Compiling 8 Scala sources to /home/jgroeger/Code/MeasurementJob/target/scala-2.11/classes...
[info] Running de.company.project.Main dev MeasurementJob
[error] Exception in thread "main" org.apache.flink.api.common.InvalidProgramException: The implementation of the RichCoFlatMapFunction is not serializable. The object probably contains or references non serializable fields.
[error] at org.apache.flink.api.java.ClosureCleaner.clean(ClosureCleaner.java:100)
[error] at org.apache.flink.streaming.api.environment.StreamExecutionEnvironment.clean(StreamExecutionEnvironment.java:1478)
[error] at org.apache.flink.streaming.api.datastream.DataStream.clean(DataStream.java:161)
[error] at org.apache.flink.streaming.api.datastream.ConnectedStreams.flatMap(ConnectedStreams.java:230)
[error] at org.apache.flink.streaming.api.scala.ConnectedStreams.flatMap(ConnectedStreams.scala:127)
[error] at de.company.project.jobs.MeasurementJob.run(MeasurementJob.scala:139)
[error] at de.company.project.Main$.main(Main.scala:55)
[error] at de.company.project.Main.main(Main.scala)
[error] Caused by: java.io.NotSerializableException: de.company.project.jobs.MeasurementJob
[error] at java.io.ObjectOutputStream.writeObject0(ObjectOutputStream.java:1184)
[error] at java.io.ObjectOutputStream.defaultWriteFields(ObjectOutputStream.java:1548)
[error] at java.io.ObjectOutputStream.writeSerialData(ObjectOutputStream.java:1509)
[error] at java.io.ObjectOutputStream.writeOrdinaryObject(ObjectOutputStream.java:1432)
[error] at java.io.ObjectOutputStream.writeObject0(ObjectOutputStream.java:1178)
[error] at java.io.ObjectOutputStream.writeObject(ObjectOutputStream.java:348)
[error] at org.apache.flink.util.InstantiationUtil.serializeObject(InstantiationUtil.java:301)
[error] at org.apache.flink.api.java.ClosureCleaner.clean(ClosureCleaner.java:81)
[error] ... 7 more
java.lang.RuntimeException: Nonzero exit code returned from runner: 1
at scala.sys.package$.error(package.scala:27)
[trace] Stack trace suppressed: run last MeasurementJob/compile:run for the full output.
[error] (MeasurementJob/compile:run) Nonzero exit code returned from runner: 1
[error] Total time: 9 s, completed Nov 15, 2016 2:28:46 PM
Process finished with exit code 1
The error message:
The implementation of the RichCoFlatMapFunction is not serializable. The object probably contains or references non serializable fields.
However, the only field my JoiningCoFlatMap has is the suggested ValueState.
The signature looks like this:
class JoiningCoFlatMap extends RichCoFlatMapFunction[Measurement, WhoMeasured, (Measurement, String)] {

Answer 1

I think what you want to do is a window operation followed by a join.
You can implement the a join of a high-volume stream and a low-value update-by-key stream using a stateful CoFlatMapFunction as in the example below:
val measures: DataStream[Measurement] = ???
val who: DataStream[WhoMeasured] = ???
val agg: DataStream[(Int, Long)] = measures
.keyBy(_._2) // measured_by_id
.timeWindow(Time.seconds(60), Time.seconds(5))
.apply( (id: Int, w: TimeWindow, v: Iterable[(Int, Int, String)], out: Collector[(Int, Long)]) => {
// do your aggregation
})
val joined: DataStream[(Int, Long, String)] = agg
.keyBy(_._1) // measured_by_id
.connect(who.keyBy(_.who_measured_id))
.flatMap(new JoiningCoFlatMap)
// CoFlatMapFunction
class JoiningCoFlatMap extends RichCoFlatMapFunction[(Int, Long), WhoMeasured, (Int, Long, String)] {
var names: ValueState[String] = null
override def open(conf: Configuration): Unit = {
val stateDescrptr = new ValueStateDescriptor[String](
"whoMeasuredName",
classOf[String],
"" // default value
)
names = getRuntimeContext.getState(stateDescrptr)
}
override def flatMap1(a: (Int, Long), out: Collector[(Int, Long, String)]): Unit = {
// join with state
out.collect( (a._1, a._2, names.value()) )
}
override def flatMap2(w: WhoMeasured, out: Collector[(Int, Long, String)]): Unit = {
// update state
names.update(w.name)
}
}
A note on the implementation: A CoFlatMapFunction cannot decide which input to process, i.e., the flatmap1 and flatmap2 functions are called depending on what data arrives at the operator. It cannot be controlled by the function. This is a problem when initializing the state. In the beginning, the state might not have the correct name for an arriving Measurement object but return the default value. You can avoid that by buffering the measurements and joining them once, the first update for the key from the who stream arrives. You'll need another state for that.