Introducing HashSet<T> [Kim Hamilton]

Introducing HashSet<T> [Kim Hamilton]

  • Comments 31

HashSet<T> is in our latest CTP, and you can find it in the System.Collections.Generic namespace. The naming discussion over the last month has motivated me to recap some naming highlights for HashSet, so hang in til the end if you’re interested.

HashSet is an unordered collection containing unique elements. It has the standard collection operations Add, Remove, Contains, but since it uses a hash-based implementation, these operation are O(1). (As opposed to List<T> for example, which is O(n) for Contains and Remove.) HashSet also provides standard set operations such as union, intersection, and symmetric difference.

    HashSet<int> theSet1 = new HashSet<int>();
    theSet1.Add(1);
    theSet1.Add(2);
    theSet1.Add(2);
    // theSet1 contains 1,2

    HashSet<int> theSet2 = new HashSet<int>();
    theSet2.Add(1);
    theSet2.Add(3);
    theSet2.Add(4);
    // theSet2 contains 1,3,4

    theSet1.UnionWith(theSet2);
    // theSet1 contains 1,2,3,4

HashSet’s default Add operation returns a bool letting you know whether the item was added, so in the code sample above, you could check the return type to check whether the item was already in the set.

    bool added = theSet1.Add(2); // added is true
    added = theSet1.Add(2); // added is false

If you’re familiar with our ICollection<T> interface, notice that this means ICollection<T>.Add (returning void) has an explicit implementation, allowing HashSet<T> to introduce its own Add.

A note on uniqueness: HashSet determines equality according to the EqualityComparer you specify, or the default EqualityComparer for the type (if you didn’t specify). In the above example we didn’t specify an EqualityComparer so it will use the default for Int32. In the next example, we’ll use an OddEvenComparer, which considers items equal if they are both even or both odd.

    class OddEvenComparer : IEqualityComparer<int> {
        public OddEvenComparer() {}
        public bool Equals(int x, int y) {
            return (x & 1) == (y & 1);
        }

        public int GetHashCode(int x) {
            return (x & 1);
        }
    }

    ...

    // Now use the comparer
    HashSet<int> oddEvenSet = new HashSet<int>(new OddEvenComparer());
    oddEvenSet.Add(1);
    oddEvenSet.Add(3);
    oddEvenSet.Add(4);
    // oddEventSet contains 1,4; it considered 1 and 3 equal.

Notice the name UnionWith in the first example. UnionWith, as with the other set operations, modifies the set it’s called on and doesn’t create a new set. This distinction is important because the Linq operations Union, Intersect, etc on IEnumerable create a new set. So HashSet’s methods aren’t duplicating Linq; they’re provided in case you want to avoid creating a new set, and they’re distinguished by the With suffix.

Now for some naming fun, which will demonstrate some other framework guidelines. We would have liked to name this feature Set. This is because it’s preferred to use a common name rather than one that reveals details about the implementation. To borrow an example from Krzysztof Cwalina and Brad Abram’s book Framework Design Guidelines, a type used to submit print jobs to a print queue should be named Printer, and not PrintQueue. Applying this guideline to this class – HashSet, while more technically precise, isn’t as recognizable at Set. You can see this guideline in other class names in the System.Collections.Generic namespace: List<T> instead of ArrayList<T>, Dictionary<T> instead of Hashtable<T>.

This brings up the question of whether naming it Set would have been a bad idea if we add other sets in the future, such as an OrderedSet. However, a hash-based unordered set can reasonably be considered the “go-to” set because of its good performance, so distinguishing it with the name Set would still be acceptable.

Any guesses as to why we didn’t go with the name Set?

  • Because "set" is a reserved contextual keyword

    http://msdn2.microsoft.com/en-gb/library/x53a06bb.aspx

  • FXCop rule prohibits naming a type after a language keyword. "set" is also used in C#, but also in VB6, though it is either gone or optional in VB.NET.

    It would be nice to support operator overloading for sets.

  • I would strongly prefer the name Set instead of HashSet.  As you state, HashSet refers too strongly to the implementation.  Users of the class don't really care about the implementation, we just care about the functionality and performance characteristics.

  • Even though you no longer have to use Set when assigning references in VB, it is still a keyword.  It is used in declaring property accessors:

    Property Text() As String

    Get

     Return _text

    End

    Set(ByVal value As String)

     _text = value

    End Set

    End Property

    I'm not sure if Set is a true keyword or a contextual keyword, but either way there's potential for confusion.

  • PowerCollections uses Set<T>... I can't say I've ever had a problem with keyword collision with it.

  • You've been kicked (a good thing) - Trackback from DotNetKicks.com

  • HashSet is a new generic collection that has been added to the System.Collections.Generic namespace.

  • > Any guesses as to why we didn’t

    > go with the name Set?

    I assume you mean Set<T>.

    That makes the VB argument mostly irrelevant.

    I am _really_ glad that you did not call it Set<T>. I eagerly await the day that ISet<T> is introduced along with other set types, TreeSet<T>, OrderSet<T> etc.

    However, I think it is a mistake to introduce HashSet<T> now without an ISet<T>. A lot of code will need to be rewritten later (such as updating signatures) because of this choice.

    I hope the type is unsealed. Lets say I want to design a TreeSet<T>. It really does not make sense for me to extend HashSet<T> to do this. Yet without an ISet<T>, there really aren't any other options. I could ignore HashSet<T> entirely and write everything from scratch. However, that does not fit well into the _framework_.

    Please reconsider this issue.

  • A Set type is great to have. I've gotten by with Dictionary<KeyType, object> so far (leaving the value as null), but this is even better.

    Another type that I use often is Dictionary<KeyType, List<ValueType>> - a key maps to a set of zero or more values. I would love to see something like this in the BCL.

  • Hello Kim,

    When and where this collection will be available?

  • The BCL Team work in a new type collection called (temporary) HashSet that is a collection containing

  • I think this is a great idea, as it is really a missing feature currently. Thus why don't you go a step beyond by creating an HashList which would be o(1) (or did I miss it)

  • Hi Isreal,

    It's in our latest Orcas CTP, which you can download here:

    http://www.microsoft.com/downloads/details.aspx?FamilyId=C09B5A2D-EB6A-44B6-8BBD-3764A2FDA9CE&displaylang=en

    It's in System.Core.dll, in the System.Collections.Generic namespace.

    Thanks,

    Kim

  • Hi Sebastien,

    I'm interpreting HashList as a collection in which you can perform both index-based lookups and hash-based lookups. In other words, a generalization of NameValueCollection (which accepts only Strings).

    We haven't gotten a lot of requests for this type of collection, and as with NVC, it would have some additional overhead. So we'd still recommend picking one or the other (List or HashSet) if you don't actually require both types of lookups. See the related blog on NVC overhead here:

    http://blogs.msdn.com/bclteam/archive/2006/09/05/741660.aspx

    Thanks,

    Kim

  • Hi Dono,

    >> A lot of code will need to be rewritten later (such as updating signatures) because of this choice.

    We had a lot of discussion on this point during the design of HashSet (which Justin summarized); we'll most likely follow up soon with some more details.

    Thanks,

    Kim

Page 1 of 3 (31 items) 123