api.md

fitMultiple.md

I wanted to circulate an alternative API for addressing the issue of model specific optimizations in MLlib, https://issues.apache.org/jira/browse/SPARK-22126. This builds on top of the work by Weichen and the API he's proposed. The idea here is break up the description of how multiple model instances can be fit in parallel and the actual execution of this fitting. I propose that we add a new method called fitMultiple (working title) which returns Array[Callable[Model[_]]]. fitMultiple(dataset, paramMaps).map(callable => callable.call()) should be the same as paramMaps.map(pm => fit(dataset, pm)), but could include some model specific performance optimizations. Callables returned by fitMultiple should be designed to be thread safe so that they can be run in parallel (eg by CrossValidator).

We would provide a default implementation of fitMultiple(...) so that even developers who write their own Estimators/Transformers would not need to be implement this method unless they wanted to include performance optimizations.

def fitMultiple(dataset: Dataset[_], paramMaps: Array[ParamMap]): Array[Callable[M]] = {
  paramMaps.map { paramMap =>
    new Callable[M] {
      override def call(): M = fit(dataset, paramMap)
    }
  }
}

We would also update fit(dataset, paramMaps) to use fitMultiple so that it would be a wrapper around fitMultiple which would benefit from any performance optimizations but would run the fitting synchronously in the current thread.

def fit(dataset: Dataset[_], paramMaps: Array[ParamMap]): Seq[M] = {
   fitMultiple(dataset, paramMaps).map{ _.call() }
}

I do not understand why we would need to make batches of callables together, this is something that can be dealt with when creating the callables. For example, you can still produce 4 callables, but only 2 do some more complex work, and the last 2 are just piggybacking on the hard work of the first 2. Of course, in terms of scheduling, one may make some mistakes, but for large numbers of configurations or large thread pools this is not going to be relevant.

What Weichen is proposing makes sense to me. I don't see how would you efficiently synchronize the Callables if they were piggy-backing on each other's results. And I think this would matter even with large thread pools since the number of dependent tasks is not bound and can be O(N) (if I understood Tim's comment correctly).

Let's move the discussion to the JIRA, https://issues.apache.org/jira/browse/SPARK-22126. I'm also updating the gist to name the method fitMultiple. Sorry if that makes this thread harder to read, but I want to present the API as fitMultiple in OSS.

Making Callables piggy-backing on each other will bring more complexity, and it seems will bring risks of dead-lock (if scheduled in wrong order)? for example, the default implementation, callables are scheduled serially, if a certain callable piggy-backing on others are scheduled first, then dead-lock occurs ?