Hi Matt, it is not possible to share an object across different task of the same operator or even different operators. This would be globally mutable state which is in general hard to get efficient in distributed systems.
Something that might work is to use a CoFlatMapOperator with one input being the training data and the other the actual input. Then you can train the model and query the model in the same operator. You would have multiple models, one in each parallel task. If you can partition the training (and input) data in a meaningful way, you would have a partition or key specific model. You can also use random partitioning and have models which are based on random samples of the training data. Or if you want each model to be based on the same input data, you can broadcast the training data. This would look as follows: val input = ??? val training = ??? val predictions = input.keyBy(xxx).connect(training.keyBy(xxx)).flatMap(YourFlatMap)) // partitioned variant val predictions = input.connect(training.shuffle).flatMap(YourFlatMap)) // random variant val predictions = input.connect(training.broadcast).flatMap(YourFlatMap)) // broadcasted variant An example of a CoFlatMap which trains and queries a prediction model can be found in the Flink Training [1] (code [2]). Hope this helps, Fabian [1] http://dataartisans.github.io/flink-training/exercises/timePrediction.html [2] https://github.com/dataArtisans/flink-training-exercises/blob/master/src/main/java/com/dataartisans/flinktraining/exercises/datastream_java/state/TravelTimePrediction.java 2017-01-13 15:38 GMT+01:00 Matt <dromitl...@gmail.com>: > Errata: How can an *object (such as the classifier, line 1)* be accessed > by any Flink node [...] > > Just in case, the classifier itself can't be serialized I believe, it's > part of a framework which I can't modify. In any case, even if it's > serialized, I guess the cost of moving it to one node and then another > makes the whole data flow unpractical. It's better to move all created > instances to one single node where only one instance of the classifier > is maintained. > > I'm not sure if this is possible or how to do this. > > On Thu, Jan 12, 2017 at 11:11 PM, Matt <dromitl...@gmail.com> wrote: > >> Hello, >> >> I have a stream of objects which I use to update the model of a >> classification algorithm and another stream with the objects I need to >> classify in real time. >> >> The problem is that the instances for training and evaluation are >> processed on potentially different Flink nodes, but the classifier should >> be applied to all instances no matter in what node it was generated (ie, >> the classifier should be accessible from any Flink node). >> >> Just to make it clearer, here is what would NOT work since these sink >> functions are not serializable: https://gist.github.com/b979bf >> 742b0d2f3da8cc8e5e91207151 >> >> Two questions here: >> >> *1. How can an instance be accessed by any Flink node like this (line 11 >> and 19)? Maybe there's a better approach to this problem.* >> >> *2. In the example the second stream (line 15) is started right away but >> at startup the classifier is not ready to use until it has been trained >> with enough instances. Is it possible to do this? If I'm not wrong >> env.execute (line 24) can be used only once.* >> >> Regards, >> Matt >> > >
