The Danger of Subtype Polymorphism Applied to Tuples

Lukas EderOctober 23rd, 2015Last Updated: October 23rd, 2015

1 48 4 minutes read

Java 8 has lambdas and streams, but no tuples, which is a shame. This is why we have implemented tuples in jOOλ – Java 8’s missing parts. Tuples are really boring value type containers. Essentially, they’re just an enumeration of types like these:

public class Tuple2<T1, T2> {
    public final T1 v1;
    public final T2 v2;

    public Tuple2(T1 v1, T2 v2) {
        this.v1 = v1;
        this.v2 = v2;
    }

    // [...]
}


public class Tuple3<T1, T2, T3> {
    public final T1 v1;
    public final T2 v2;
    public final T3 v3;

    public Tuple3(T1 v1, T2 v2, T3 v3) {
        this.v1 = v1;
        this.v2 = v2;
        this.v3 = v3;
    }

    // [...]
}

Writing tuple classes is a very boring task, and it’s best done using a source code generator.

Tuples in other languages and APIs

jOOλ‘s current version features tuples of degrees 0 – 16. C# and other .NET languages have tuple types between 1 – 8. There’s a special library just for tuples called Javatuples with tuples between degrees 1 and 10, and the authors went the extra mile and gave the tuples individual English names:

Unit<A> // (1 element)
Pair<A,B> // (2 elements)
Triplet<A,B,C> // (3 elements)
Quartet<A,B,C,D> // (4 elements)
Quintet<A,B,C,D,E> // (5 elements)
Sextet<A,B,C,D,E,F> // (6 elements)
Septet<A,B,C,D,E,F,G> // (7 elements)
Octet<A,B,C,D,E,F,G,H> // (8 elements)
Ennead<A,B,C,D,E,F,G,H,I> // (9 elements)
Decade<A,B,C,D,E,F,G,H,I,J> // (10 elements)

Why?

because Ennead really rings that sweet bell when I see it

Last, but not least, jOOQ also has a built-in tuple-like type, the org.jooq.Record, which serves as a base type for nice subtypes like Record7<T1, T2, T3, T4, T5, T6, T7>. jOOQ follows Scala and defines records up to a degree of 22.

Watch out when defining tuple type hierarchies

As we have seen in the previous example, Tuple3 has much code in common with Tuple2.

As we’re all massively brain-damaged by decades of object orientation and polymorphic design anti-patters, we might think that it would be a good idea to let Tuple3<T1, T2, T3> extend Tuple2<T1, T2>, as Tuple3 just adds one more attribute to the right of Tuple2, right? So…

public class Tuple3<T1, T2, T3> extends Tuple2<T1, T2> {
    public final T3 v3;

    public Tuple3(T1 v1, T2 v2, T3 v3) {
        super(v1, v2);
        this.v3 = v3;
    }

    // [...]
}

The truth is: That’s about the worst thing you could do, for several reasons. First off, yes. Both Tuple2 and Tuple3 are tuples, so they do have some common features. It’s not a bad idea to group those features in a common super type, such as:

public class Tuple2<T1, T2> implements Tuple {
    // [...]
}

But the degree is not one of those things. Here’s why:

Permutations

Think about all the possible tuples that you can form. If you let tuples extend each other, then a Tuple5 would also be assignment-compatible with a Tuple2, for instance. The following would compile perfectly:

Tuple2<String, Integer> t2 = tuple("A", 1, 2, 3, "B");

When letting Tuple3 extend Tuple2, it may have seemed like a good default choice to just drop the right-most attribute from the tuple in the extension chain.

But in the above example, why don’t I want to re-assign (v2, v4) such that the result is (1, 3), or maybe (v1, v3), such that the result is ("A", 2)?

There are a tremendous amount of permutations of possible attributes that could be of interest when “reducing” a higher degree tuple to a lower degree one. No way a default of dropping the right-most attribute will be sufficiently general for all use-cases

Type systems

Much worse than the above, there would be drastic implications for the type system, if Tuple3 extended Tuple2. Check out the jOOQ API, for instance. In jOOQ, you can safely assume the following:

// Compiles:
TABLE1.COL1.in(select(TABLE2.COL1).from(TABLE2))

// Must not compile:
TABLE1.COL1.in(select(TABLE2.COL1, TABLE2.COL2).from(TABLE2))

The first IN predicate is correct. The left hand side of the predicate has a single column (as opposed to being a row value expression). This means that the right hand side of the predicate must also operate on single-column expressions, e.g. a SELECT subquery that selects a single column (of the same type).

The second example selects too many columns, and the jOOQ API will tell the Java compiler that this is wrong.

This is guaranteed by jOOQ via the Field.in(Select) method, whose signature reads:

public interface Field<T> {
    ...
    Condition in(Select<? extends Record1<T>> select);
    ...
}

So, you can provide a SELECT statement that produces any subtype of the Record1<T> type.

Luckily, Record2 does not extend Record1

If now Record2 extended Record1, which might have seemed like a good idea at first, the second query would suddenly compile:

// This would now compile
TABLE1.COL1.in(select(TABLE2.COL1, TABLE2.COL2).from(TABLE2))

… even if it forms an invalid SQL statement. It would compile because it would generate a Select<Record2<Type1, Type2>> type, which would be a subtype of the expected Select<Record1<Type1>> from the Field.in(Select) method.

Conclusion

Tuple2 and Tuple5 types are fundamentally incompatible types. In strong type systems, you mustn’t be lured into thinking that similar types, or related types should also be compatible types.

Type hierarchies are something very object-oriented, and by object-oriented, I mean the flawed and over-engineered notion of object orientation that we’re still suffering from since the 90s. Even in “the Enterprise”, most people have learned to favour Composition over Inheritance. Composition in the case of tuples means that you can well transform a Tuple5 to a Tuple2. But you cannot assign it.

In jOOλ, such a transformation can be done very easily as follows:

// Produces (1, 3)
Tuple2<String, Integer> t2_4 = 
    tuple("A", 1, 2, 3, "B")
    .map((v1, v2, v3, v4, v5) -> tuple(v2, v4));

// Produces ("A", 2)
Tuple2<String, Integer> t1_3 = 
    tuple("A", 1, 2, 3, "B")
    .map((v1, v2, v3, v4, v5) -> tuple(v1, v3));

The idea is that you operate on immutable values, and that you can easily extract parts of those values and map / recombine them to new values.

Reference:

The Danger of Subtype Polymorphism Applied to Tuples from our JCG partner Lukas Eder at the JAVA, SQL, AND JOOQ blog.