What are the actual advantages of the visitor pattern? What are the alternatives?

Question

I read quite a lot about the visitor pattern and its supposed advantages. To me however it seems they are not that much advantages when applied in practice:

• "Convenient" and "elegant" seems to mean lots and lots of boilerplate code
• Therefore, the code is hard to follow. Also 'accept'/'visit' is not very descriptive
• Even uglier boilerplate code if your programming language has no method overloading (i.e. Vala)
• You cannot in general add new operations to an existing type hierarchy without modification of all classes, since you need new 'accept'/'visit' methods everywhere as soon as you need an operation with different parameters and/or return value (changes to classes all over the place is one thing this design pattern was supposed to avoid!?)
• Adding a new type to the type hierarchy requires changes to all visitors. Also, your visitors cannot simply ignore a type - you need to create an empty visit method (boilerplate again)

It all just seems to be an awful lot of work when all you want to do is actually:

// Pseudocode
int SomeOperation(ISomeAbstractThing obj) {
    switch (type of obj) {
        case Foo: // do Foo-specific stuff here
        case Bar: // do Bar-specific stuff here
        case Baz: // do Baz-specific stuff here
        default: return 0; // do some sensible default if type unknown or if we don't care
    }
}

The only real advantage I see (which btw i haven't seen mentioned anywhere): The visitor pattern is probably the fastest method to implement the above code snippet in terms of cpu time (if you don't have a language with double dispatch or efficient type comparison in the fashion of the pseudocode above).

Questions:

• So, what advantages of the visitor pattern have I missed?
• What alternative concepts/data structures could be used to make the above fictional code sample run equally fast?

Answer 1

For as far as I have seen so far there are two uses / benefits for the visitor design pattern:

1. Double dispatch
2. Separate data structures from the operations on them

Double dispatch

Let's say you have a Vehicle class and a VehicleWasher class. The VehicleWasher has a Wash(Vehicle) method:

VehicleWasher
    Wash(Vehicle)

Vehicle

Additionally we also have specific vehicles like a car and in the future we'll also have other specific vehicles. For this we have a Car class but also a specific CarWasher class that has an operation specific to washing cars (pseudo code):

CarWasher : VehicleWasher
    Wash(Car)

Car : Vehicle

Then consider the following client code to wash a specific vehicle (notice that x and washer are declared using their base type because the instances might be dynamically created based on user input or external configuration values; in this example they are simply created with a new operator though):

Vehicle x = new Car();
VehicleWasher washer = new CarWasher();

washer.Wash(x);

Many languages use single dispatch to call the appropriate function. Single dispatch means that during runtime only a single value is taken into account when determining which method to call. Therefore only the actual type of washer we'll be considered. The actual type of x isn't taken into account. The last line of code will therefore invoke CarWasher.Wash(Vehicle) and NOT CarWasher.Wash(Car).

If you use a language that does not support multiple dispatch and you do need it (I can honoustly say I have never encountered such a situation though) then you can use the visitor design pattern to enable this. For this two things need to be done. First of all add an Accept method to the Vehicle class (the visitee) that accepts a VehicleWasher as a visitor and then call its operation Wash:

Accept(VehicleWasher washer)
    washer.Wash(this);

The second thing is to modify the calling code and replace the washer.Wash(x); line with the following:

x.Accept(washer);

Now for the call to the Accept method the actual type of x is considered (and only that of x since we are assuming to be using a single dispatch language). In the implementation of the Accept method the Wash method is called on the washer object (the visitor). For this the actual type of the washer is considered and this will invoke CarWasher.Wash(Car). By combining two single dispatches a double dispatch is implemented.

Now to eleborate on your remark of the terms like Accept and Visit and Visitor being very unspecific. That is absolutely true. But it is for a reason.

Consider the requirement in this example to implement a new class that is able to repair vehicles: a VehicleRepairer. This class can only be used as a visitor in this example if it would inherit from VehicleWasher and have its repair logic inside a Wash method. But that ofcourse doesn't make any sense and would be confusing. So I totally agree that design patterns tend to have very vague and unspecific naming but it does make them applicable to many situations. The more specific your naming is, the more restrictive it can be.

Your switch statement only considers one type which is actually a manual way of single dispatch. Applying the visitor design pattern in the above way will provide double dispatch. This way you do not necessarily need additional Visit methods when adding additional types to your hierarchy. Ofcourse it does add some complexity as it makes the code less readable. But ofcourse all patterns come at a price.

Ofcourse this pattern cannot always be used. If you expect lots of complex operations with multiple parameters then this will not be a good option.

An alternative is to use a language that does support multiple dispatch. For instance .NET did not support it until version 4.0 which introduced the dynamic keyword. Then in C# you can do the following:

washer.Wash((dynamic)x);

Because x is then converted to a dynamic type its actual type will be considered for the dispatch and so both x and washer will be used to select the correct method so that CarWasher.Wash(Car) will be called (making the code work correctly and staying intuitive).

Separate data structures and operations

The other benefit and requirement is that it can separate the data structures from the operations. This can be an advantage because it allows new visitors to be added that have there own operations while it also allows data structures to be added that 'inherit' these operations. It can however be only applied if this seperation can be done / makes sense. The classes that perform the operations (the visitors) do not know the structure of the data structures nor do they have to know that which makes code more maintainable and reusable. When applied for this reason the visitors have operations for the different elements in the data structures.

Say you have different data structures and they all consist of elements of class Item. The structures can be lists, stacks, trees, queues etc.

You can then implement visitors that in this case will have the following method:

Visit(Item)

The data structures need to accept visitors and then call the Visit method for each Item.

This way you can implement all kinds of visitors and you can still add new data structures as long as they consist of elements of type Item.

For more specific data structures with additional elements (e.g. a Node) you might consider a specific visitor (NodeVisitor) that inherits from your conventional Visitor and have your new data structures accept that visitor (Accept(NodeVisitor)). The new visitors can be used for the new data structures but also for the old data structures due to inheritence and so you do not need to modify your existing 'interface' (the super class in this case).