Java Lambda Stream Distinct() on arbitrary key?

I frequently ran into a problem with Java lambda expressions where when I wanted to distinct() a stream on an arbitrary property or method of an object, but wanted to keep the object rather than map it to that property or method. I started to create containers as discussed here but I started to do it enough to where it became annoying and made a lot of boilerplate classes.

I threw together this Pairing class, which holds two objects of two types and allows you to specify keying off the left, right, or both objects. My question is... is there really no built-in lambda stream function to distinct() on a key supplier of some sorts? That would really surprise me. If not, will this class fulfill that function reliably?

Here is how it would be called

BigDecimal totalShare = orders.stream().map(c -> Pairing.keyLeft(c.getCompany().getId(), c.getShare())).distinct().map(Pairing::getRightItem).reduce(BigDecimal.ZERO, (x,y) -> x.add(y));

Here is the Pairing class

    public final class Pairing<X,Y>  {
           private final X item1;
           private final Y item2;
           private final KeySetup keySetup;
           
           private static enum KeySetup {LEFT,RIGHT,BOTH};
           
           private Pairing(X item1, Y item2, KeySetup keySetup) {
                  this.item1 = item1;
                  this.item2 = item2;
                  this.keySetup = keySetup;
           }
           public X getLeftItem() { 
                  return item1;
           }
           public Y getRightItem() { 
                  return item2;
           }
    
           public static <X,Y> Pairing<X,Y> keyLeft(X item1, Y item2) { 
                  return new Pairing<X,Y>(item1, item2, KeySetup.LEFT);
           }
           
           public static <X,Y> Pairing<X,Y> keyRight(X item1, Y item2) { 
                  return new Pairing<X,Y>(item1, item2, KeySetup.RIGHT);
           }
           public static <X,Y> Pairing<X,Y> keyBoth(X item1, Y item2) { 
                  return new Pairing<X,Y>(item1, item2, KeySetup.BOTH);
           }
           public static <X,Y> Pairing<X,Y> forItems(X item1, Y item2) { 
                  return keyBoth(item1, item2);
           }
           
           @Override
           public int hashCode() {
                  final int prime = 31;
                  int result = 1;
                  if (keySetup.equals(KeySetup.LEFT) || keySetup.equals(KeySetup.BOTH)) {
                  result = prime * result + ((item1 == null) ? 0 : item1.hashCode());
                  }
                  if (keySetup.equals(KeySetup.RIGHT) || keySetup.equals(KeySetup.BOTH)) {
                  result = prime * result + ((item2 == null) ? 0 : item2.hashCode());
                  }
                  return result;
           }
    
           @Override
           public boolean equals(Object obj) {
                  if (this == obj)
                         return true;
                  if (obj == null)
                         return false;
                  if (getClass() != obj.getClass())
                         return false;
                  Pairing<?,?> other = (Pairing<?,?>) obj;
                  if (keySetup.equals(KeySetup.LEFT) || keySetup.equals(KeySetup.BOTH)) {
                         if (item1 == null) {
                               if (other.item1 != null)
                                      return false;
                         } else if (!item1.equals(other.item1))
                               return false;
                  }
                  if (keySetup.equals(KeySetup.RIGHT) || keySetup.equals(KeySetup.BOTH)) {
                         if (item2 == null) {
                               if (other.item2 != null)
                                      return false;
                         } else if (!item2.equals(other.item2))
                               return false;
                  }
                  return true;
           }
           
    }

UPDATE:

Tested Stuart's function below and it seems to work great. The operation below distincts on the first letter of each string. The only part I'm trying to figure out is how the ConcurrentHashMap maintains only one instance for the entire stream

public class DistinctByKey {

	public static <T> Predicate<T> distinctByKey(Function<? super T,Object> keyExtractor) {
	    Map<Object,Boolean> seen = new ConcurrentHashMap<>();
	    return t -> seen.putIfAbsent(keyExtractor.apply(t), Boolean.TRUE) == null;
	}
	
	public static void main(String[] args) { 
		
		final ImmutableList<String> arpts = ImmutableList.of("ABQ","ALB","CHI","CUN","PHX","PUJ","BWI");
		
		arpts.stream().filter(distinctByKey(f -> f.substring(0,1))).forEach(s -> System.out.println(s));
	}

Output is...

ABQ
CHI
PHX
BWI

The distinct operation is a stateful pipeline operation; in this case it's a stateful filter. It's a bit inconvenient to create these yourself, as there's nothing built-in, but a small helper class should do the trick:

/**
 * Stateful filter. T is type of stream element, K is type of extracted key.
 */
static class DistinctByKey<T,K> {
    Map<K,Boolean> seen = new ConcurrentHashMap<>();
    Function<T,K> keyExtractor;
    public DistinctByKey(Function<T,K> ke) {
        this.keyExtractor = ke;
    }
    public boolean filter(T t) {
        return seen.putIfAbsent(keyExtractor.apply(t), Boolean.TRUE) == null;
    }
}

I don't know your domain classes, but I think that, with this helper class, you could do what you want like this:

BigDecimal totalShare = orders.stream()
    .filter(new DistinctByKey<Order,CompanyId>(o -> o.getCompany().getId())::filter)
    .map(Order::getShare)
    .reduce(BigDecimal.ZERO, BigDecimal::add);

Unfortunately the type inference couldn't get far enough inside the expression, so I had to specify explicitly the type arguments for the DistinctByKey class.

This involves more setup than the collectors approach described by Louis Wasserman, but this has the advantage that distinct items pass through immediately instead of being buffered up until the collection completes. Space should be the same, as (unavoidably) both approaches end up accumulating all distinct keys extracted from the stream elements.

UPDATE

It's possible to get rid of the K type parameter since it's not actually used for anything other than being stored in a map. So Object is sufficient.

/**
 * Stateful filter. T is type of stream element.
 */
static class DistinctByKey<T> {
    Map<Object,Boolean> seen = new ConcurrentHashMap<>();
    Function<T,Object> keyExtractor;
    public DistinctByKey(Function<T,Object> ke) {
        this.keyExtractor = ke;
    }
    public boolean filter(T t) {
        return seen.putIfAbsent(keyExtractor.apply(t), Boolean.TRUE) == null;
    }
}

BigDecimal totalShare = orders.stream()
    .filter(new DistinctByKey<Order>(o -> o.getCompany().getId())::filter)
    .map(Order::getShare)
    .reduce(BigDecimal.ZERO, BigDecimal::add);

This simplifies things a bit, but I still had to specify the type argument to the constructor. Trying to use diamond or a static factory method doesn't seem to improve things. I think the difficulty is that the compiler can't infer generic type parameters -- for a constructor or a static method call -- when either is in the instance expression of a method reference. Oh well.

(Another variation on this that would probably simplify it is to make DistinctByKey<T> implements Predicate<T> and rename the method to eval. This would remove the need to use a method reference and would probably improve type inference. However, it's unlikely to be as nice as the solution below.)

UPDATE 2

Can't stop thinking about this. Instead of a helper class, use a higher-order function. We can use captured locals to maintain state, so we don't even need a separate class! Bonus, things are simplified so type inference works!

public static <T> Predicate<T> distinctByKey(Function<? super T,Object> keyExtractor) {
    Map<Object,Boolean> seen = new ConcurrentHashMap<>();
    return t -> seen.putIfAbsent(keyExtractor.apply(t), Boolean.TRUE) == null;
}

BigDecimal totalShare = orders.stream()
    .filter(distinctByKey(o -> o.getCompany().getId()))
    .map(Order::getShare)
    .reduce(BigDecimal.ZERO, BigDecimal::add);

You more or less have to do something like

 elements.stream()
    .collect(Collectors.toMap(
        obj -> extractKey(obj), 
        obj -> obj, 
       (first, second) -> first
           // pick the first if multiple values have the same key
       )).values().stream();

Another way of finding distinct elements

List<String> uniqueObjects = ImmutableList.of("ABQ","ALB","CHI","CUN","PHX","PUJ","BWI")
            .stream()
            .collect(Collectors.groupingBy((p)->p.substring(0,1))) //expression 
            .values()
            .stream()
            .flatMap(e->e.stream().limit(1))
            .collect(Collectors.toList());

A variation on Stuart Marks second update. Using a Set.

public static <T> Predicate<T> distinctByKey(Function<? super T, Object> keyExtractor) {
    Set<Object> seen = Collections.newSetFromMap(new ConcurrentHashMap<>());
    return t -> seen.add(keyExtractor.apply(t));
}

We can also use RxJava (very powerful reactive extension library)

Observable.from(persons).distinct(Person::getName)

or

Observable.from(persons).distinct(p -> p.getName())

To answer your question in your second update:

> The only part I'm trying to figure out is how the ConcurrentHashMap maintains only one instance for the entire stream:

public static <T> Predicate<T> distinctByKey(Function<? super T,Object> keyExtractor) {
        Map<Object,Boolean> seen = new ConcurrentHashMap<>();
        return t -> seen.putIfAbsent(keyExtractor.apply(t), Boolean.TRUE) == null;
    }

In your code sample, distinctByKey is only invoked one time, so the ConcurrentHashMap created just once. Here's an explanation:

The distinctByKey function is just a plain-old function that returns an object, and that object happens to be a Predicate. Keep in mind that a predicate is basically a piece of code that can be evaluated later. To manually evaluate a predicate, you must call a method in the Predicate interface such as test. So, the predicate

t -> seen.putIfAbsent(keyExtractor.apply(t), Boolean.TRUE) == null

is merely a declaration that is not actually evaluated inside distinctByKey.

The predicate is passed around just like any other object. It is returned and passed into the filter operation, which basically evaluates the predicate repeatedly against each element of the stream by calling test.

I'm sure filter is more complicated than I made it out to be, but the point is, the predicate is evaluated many times outside of distinctByKey. There's nothing special* about distinctByKey; it's just a function that you've called one time, so the ConcurrentHashMap is only created one time.

*Apart from being well made, @stuart-marks :)

You can use the distinct(HashingStrategy) method in Eclipse Collections.

List<String> list = Lists.mutable.with("ABQ", "ALB", "CHI", "CUN", "PHX", "PUJ", "BWI");
ListIterate.distinct(list, HashingStrategies.fromFunction(s -> s.substring(0, 1)))
    .each(System.out::println);

If you can refactor list to implement an Eclipse Collections interface, you can call the method directly on the list.

MutableList<String> list = Lists.mutable.with("ABQ", "ALB", "CHI", "CUN", "PHX", "PUJ", "BWI");
list.distinct(HashingStrategies.fromFunction(s -> s.substring(0, 1)))
    .each(System.out::println);

HashingStrategy is simply a strategy interface that allows you to define custom implementations of equals and hashcode.

public interface HashingStrategy<E>
{
    int computeHashCode(E object);
    boolean equals(E object1, E object2);
}

Note: I am a committer for Eclipse Collections.

It can be done something like

Set<String> distinctCompany = orders.stream()
        .map(Order::getCompany)
        .collect(Collectors.toSet());

Set.add(element) returns true if the set did not already contain element, otherwise false. So you can do like this.

Set<String> set = new HashSet<>();
BigDecimal totalShare = orders.stream()
    .filter(c -> set.add(c.getCompany().getId()))
    .map(c -> c.getShare())
    .reduce(BigDecimal.ZERO, BigDecimal::add);

If you want to do this parallel, you must use concurrent map.