The second alternative, with the addition of methods that take functions
with Scala types, seems the most sensible. I wonder if there is a need
then to maintain the *J Java parameter methods, or whether users could just
access the functionality by converting the Scala DataStreams to Java via
.javaStream and whatever the equivalent is for DataSets.
On Thu, Oct 4, 2018 at 8:10 AM Aljoscha Krettek <aljoscha@xxxxxxxxxx>
I'm currently working on
with the goal of adding support for Scala 2.12. There is a bit of a
and I have to explain some context first.
With Scala 2.12, lambdas are implemented using the lambda mechanism of
Java 8, i.e. Scala lambdas are now SAMs (Single Abstract Method). This
means that the following two method definitions can both take a lambda:
def map[R](mapper: MapFunction[T, R]): DataSet[R]
def map[R](fun: T => R): DataSet[R]
The Scala compiler gives precedence to the lambda version when you call
map() with a lambda in simple cases, so it works here. You could still
map() with a lambda if the lambda version of the method weren't here
because they are now considered the same. For Scala 2.11 we need both
signatures, though, to allow calling with a lambda and with a
The problem is with more complicated method signatures, like:
def reduceGroup[R](fun: (scala.Iterator[T], Collector[R]) => Unit):
def reduceGroup[R](reducer: GroupReduceFunction[T, R]): DataSet[R]
(for reference, GroupReduceFunction is a SAM with void
reduce(java.lang.Iterable<T> values, Collector<O> out))
These two signatures are not the same but similar enough for the Scala
2.12 compiler to "get confused". In Scala 2.11, I could call
with a lambda that doesn't have parameter type definitions and things
be fine. With Scala 2.12 I can't do that because the compiler can't
out which method to call and requires explicit type definitions on the
I see some solutions for this:
1. Keep the methods as is, this would force people to always explicitly
specify parameter types on their lambdas.
2. Rename the second method to reduceGroupJ() to signal that it takes a
user function that takes Java-style interfaces (the first parameter is
java.lang.Iterable while the Scala lambda takes a scala.Iterator). This
disambiguates the code, users can use lambdas without specifying explicit
parameter types but breaks the API.
One effect of 2. would be that we can add a reduceGroup() method that
takes a api.scala.GroupReduceFunction that takes proper Scala types, thus
it would allow people to implement user functions without having to cast
the various Iterator/Iterable parameters.
Either way, people would have to adapt their code when moving to Scala
2.12 in some way, depending on what style of methods they use.
There is also solution 2.5:
2.5 Rename the methods only in the Scala 2.12 build of Flink and keep the
old method names for Scala 2.11. This would require some infrastructure
I don't yet know how it can be done in a sane way.
What do you think? I personally would be in favour of 2. but it breaks