Hi, the problem is that the elements emitted from the count window operation all have a timestamp of Long.MAX_VALUE. The reason for this is that "countWindow(int, int)" de-sugars to this
input .keyBy(...) .window(GlobalWindows.create()) .trigger(CountTrigger.of(1)) .evictor(CountEvictor.of(2)) elements emitted for windows have the end timestamp of that window assigned to them and GlobalWindow has the end timestamp Long.MAX_VALUE. Using a stateful map should indeed work and I would encourage to use that instead of a count window. Cheers, Aljoscha On Thu, 8 Dec 2016 at 01:05 Nico <[email protected]> wrote: > Hi, > > I am a little bit confused regarding the windows in Flink. Is it possible > to use multiple window operators in a single flink job? In my example I > receive events every 5s, which need to be edited before further > investigation. For this I use a keyBy(ID) followed by a sliding > Count-Window (2,1)... so I always use an old and new event in order to > calculate the missing attribute (for this attribute I need 2 events with > the same ID). > > For further investigation I would like to use a Tumbling TimeWindow to > analyze events within the last 10s... however It doesn't seem to work and I > don't know why. > > How it looks like: > > stream > .keyBy("id") > .countWindow(2,1) > .reduce(new Reduce()) > .keyBy("area") > .timeWindow(Time.seconds(10)) > .fold(new Fold()) > .print() > > > When I use both windows separately, it works: > > stream.keyBy("id").countWindow(2,1).reduce(new Reduce()).print() > stream.keyBy("area").timeWindow(Time.seconds(10)).fold(new Fold()).print() > > Would it be better to use a stateful Mapoperation instead of the > countWindow? > > > Best regards, > Nico >
