Tuples( or arrays ) as Dictionary keys in C#

If you are on .NET 4.0 use a Tuple:

lookup = new Dictionary<Tuple<TypeA, TypeB, TypeC>, string>();

If not you can define a Tuple and use that as the key. The Tuple needs to override GetHashCode, Equals and IEquatable:

struct Tuple<T, U, W> : IEquatable<Tuple<T,U,W>>
{
    readonly T first;
    readonly U second;
    readonly W third;

    public Tuple(T first, U second, W third)
    {
        this.first = first;
        this.second = second;
        this.third = third;
    }

    public T First { get { return first; } }
    public U Second { get { return second; } }
    public W Third { get { return third; } }

    public override int GetHashCode()
    {
        return first.GetHashCode() ^ second.GetHashCode() ^ third.GetHashCode();
    }

    public override bool Equals(object obj)
    {
        if (obj == null || GetType() != obj.GetType())
        {
            return false;
        }
        return Equals((Tuple<T, U, W>)obj);
    }

    public bool Equals(Tuple<T, U, W> other)
    {
        return other.first.Equals(first) && other.second.Equals(second) && other.third.Equals(third);
    }
}

If you're on C# 7, you should consider using value tuples as your composite key. Value tuples typically offer better performance than the traditional reference tuples (Tuple<T1, …>) since value tuples are value types (structs), not reference types, so they avoid the memory allocation and garbage collection costs. Also, they offer conciser and more intuitive syntax, allowing for their fields to be named if you so wish. They also implement the IEquatable<T> interface needed for the dictionary.

var dict = new Dictionary<(int PersonId, int LocationId, int SubjectId), string>();
dict.Add((3, 6, 9), "ABC");
dict.Add((PersonId: 4, LocationId: 9, SubjectId: 10), "XYZ");
var personIds = dict.Keys.Select(k => k.PersonId).Distinct().ToList();

Between tuple and nested dictionaries based approaches, it's almost always better to go for tuple based.

From maintainability point of view,

  • its much easier to implement a functionality that looks like:

    var myDict = new Dictionary<Tuple<TypeA, TypeB, TypeC>, string>();
    

    than

    var myDict = new Dictionary<TypeA, Dictionary<TypeB, Dictionary<TypeC, string>>>();
    

    from the callee side. In the second case each addition, lookup, removal etc require action on more than one dictionary.

  • Furthermore, if your composite key require one more (or less) field in future, you will need to change code a significant lot in the second case (nested dictionary) since you have to add further nested dictionaries and subsequent checks.

From performance perspective, the best conclusion you can reach is by measuring it yourself. But there are a few theoretical limitations which you can consider beforehand:

  • In the nested dictionary case, having an additional dictionary for every keys (outer and inner) will have some memory overhead (more than what creating a tuple would have).

  • In the nested dictionary case, every basic action like addition, updation, lookup, removal etc need to be carried out in two dictionaries. Now there is a case where nested dictionary approach can be faster, i.e., when the data being looked up is absent, since the intermediate dictionaries can bypass the full hash code computation & comparison, but then again it should be timed to be sure. In presence of data, it should be slower since lookups should be performed twice (or thrice depending on nesting).

  • Regarding tuple approach, .NET tuples are not the most performant when they're meant to be used as keys in sets since its Equals and GetHashCode implementation causes boxing for value types.

I would go with tuple based dictionary, but if I want more performance, I would use my own tuple with better implementation.


On a side note, few cosmetics can make the dictionary cool:

  1. Indexer style calls can be a lot cleaner and intuitive. For eg,

    string foo = dict[a, b, c]; //lookup
    dict[a, b, c] = ""; //update/insertion
    

    So expose necessary indexers in your dictionary class which internally handles the insertions and lookups.

  2. Also, implement a suitable IEnumerable interface and provide an Add(TypeA, TypeB, TypeC, string) method which would give you collection initializer syntax, like:

    new MultiKeyDictionary<TypeA, TypeB, TypeC, string> 
    { 
        { a, b, c, null }, 
        ...
    };