At 7digital, we are running a weekly discussion group on design patterns, and I have been creating some small projects to exlore the issues raised in these discussions. The other week my colleague Emily presented the State Pattern, and it soon became clear that this is one of the more controversial patterns, as it seems to ride roughshod over the SOLID principles.

In this post I will give a bit of background, and then discuss the various approaches illustrated in my project.

The State Pattern

The classic case for the State Pattern is when an object’s behaviour changes according to the state it is in. A vending machine can only sell when it contains products, a shopping basket can only be checked out when it has products in it, a person can only get divorced if they are already married.

This pattern is implemented as an extension of the Strategy Pattern: the context object (vending machine, shopping basket, person) has a dependency on a state object, which in turn implements the behaviour (sell, check out, divorce) in an appropriate manner. Unlike cardinal examples of the Strategy Pattern, in which the strategy is set once and left alone, the State Pattern lets the strategy change during execution — often as the result of a method call on the state itself (if the vending machine sells the last product, it then becomes empty; if the only item is removed from the basket, it cannot be checked out; if the person gets divorced, they are no longer married).

Smelly

An immediate criticism of the State Pattern is that it leads to a system that is fairly tightly coupled. Not only is the context coupled to the state in order to make method calls, but a mechanism is then needed for updating the state, and a classical implementation of this is for the state to be coupled back to the context — a two-way coupling that is already a code smell.

Other implementations are possible: the context can update its own state after each state method call, or the state method can give the successor state as a return value; however, putting this question aside, it is the State Pattern’s lack of SOLIDity that makes it most problematic.

VAPID

Consider a context object that exposes several methods and has several states. Now consider that only some of these methods will be appropriate for any given state of the object (just as in the examples given above). Under the classic State Pattern, each concrete state must implement a method for every delegated method of the context object, even if it is not appropriate for it to do so (often implemented by throwing an exception).

First, this is a violation of the Single Responsibility Principle, as each concrete state now has responsibilities that it is designed to shirk. If we consider the ‘reason to change’ criterion, each concrete state can change because a) it changes the implementation of the stuff it does do, and b) it changes the implementation of failing to do the stuff it does not do.

More graphically, this is a violation of the Liskov Substitution Principle. This principle requires interchangeable objects (eg, those with the same interface) to behave in the same way, ie, given the same input conditions, they produce the same output conditions. The problem with the State Pattern is that it requires each concrete state to have different behaviour from its peers: violation of the LSP is seemingly baked into this pattern.

Finally, this pattern can lead to a violation of the Interface Segregation principle, insofar as a multiplicity of methods on the context class (which may be justified as being appropriate to that object) can then delegate to a a multiplicity of methods on the state interface, which, given their state-dependence, will no longer form a coherent collection.

Sharlotka

Let’s take a break from theory here, and look at my implementation.

I considered the stereotypical State Pattern examples, but let’s be honest: there are quite enough vending machines and shopping baskets in the world, and discussion of design patterns is not the context to be locking horns with Cardinal O’Brien.

So I thought I would use a culinary example instead.

Sharlotka (шарлотка) is a Russian apple cake. It is made by filling a cake tin with chopped apples, then pouring an eggy batter over the apples and baking in a warm oven till the batter has set. It is served dusted with sugar and cinnamon, and is really rather lovely.

It also makes a good, simple example of the State Pattern: each stage of cooking must be carried out in sequence, but there’s a nice loop in the middle, where you have to keep checking the cake until it’s cooked through before turning it out.

Classic Implementation

My first implementation follows the classic pattern.

You can see that there is one ISharlotkaState interface, which contains a method for each of calls that the Sharlotka context object delegates to its state. Each of these methods takes a reference to the Sharlotka as a parameter, so it can alter its state if need be. (The Sharlotka is passed with the interface IHasState<ISharlotkaState> to avoid making the State part of the public interface of the Sharlotka class itself.)

If you look at any of the concrete states (eg, ReadyToAddApplesState), you will see that most of the methods are left throwing a WrongStateException, as well as the mechanism for updating the context’s state. The _successor state is injected in the constructor, as this makes it possible to leave the mechanics of wiring everything together to the IoC container, rather than having to new up states within other states. In a small way this is reminiscent of the Chain of Resposibility Pattern, and does something to alleviate the tight coupling smell.

If you want to unit test one of the concrete states (eg, ReadyToAddApplesStateTests) then you are again left with a lot of boilerplate code.

This implementation highlights some of the deficiencies of the classic State Pattern implementation; however, it does still work, and may be appropriate for cases where it is not so much the behaviour of the context that is dependent on state, but rather its internal implementation.

Segregating the Interfaces

A more parsimonious approach has been suggested by Jan van Ryswyck, and I have implemented a version of this as the Small Interface project.

The key difference here is that rather than having a single, monolithic interface for all the states, the individual behaviours are broken into their own interfaces. When the context comes to call each state method, it first checks whether it can cast the state to the appropriate interface, and only then makes the call.

This implementation makes no further efforts to remedy the tight coupling, but does makes some improvements to the SOLID violations:

The Single Responsibility Principle is better supported, as each state is now only responsible for implementing the interfaces that are actually appropriate; we no longer have states also taking responsibility for throwing WrongStateExceptions, as this is done by the context object.

The Liskov Substitution Principle is better supported, as we no longer have sets of defective implementations of interface methods; instead we have used the Interface Segregation Principle to split out a set of atomic interfaces, each of which can more easily support the LSP. There are still opportunities for violation of the LSP, as it is entirely possible to implement the same interface in several states, and for each implementation to be different, but this problem is no longer inherent in the implementation.

This implementation makes good steps towards SOLIDity, but still has a major flaw, which is intrinsic to the state pattern: you can call methods on the stateful object that are simply not appropriate to its state; to continue the sharlotka example, you can call Serve before you have called Bake. Whilst doing this will throw an exception, it should arguably not be possible to do so in the first place.

A Fluent Implementation

A third possibility came up in our discussion — I think it was Greg who raised it —: you can implement your stateful class with a fluent interface.  You can see my version of this in the Fluent project.

Like the Small Interface implementation, this uses several atomic interfaces instead of one monolithic one; unlike that implementation however the interfaces are directly implemented by the Sharlotka class, rather than being delegated to a state object, and at any point in its lifecycle the Sharlotka is only considered as an implementation of the interface that is appropriate to its current state, ie, you never have access to methods that are not currently supported.

The trick in implementing this is in the fluent idiom: rather than calling

sharlotka.AddApples();
sharlotka.AddBatter();
sharlotka.Bake();

we want to chain the methods like this:

sharlotka.AddApples()
	.AddBatter()
	.Bake();

We can do this by making each method return this, but with a return type of the appropriate next interface.

For example, TurnOut is implemented like this:

ICanDustWithSugar ICanTurnOut.TurnOut() {
	return this;
}

which means that the next call must be a method of ICanDustWithSugar.

This implementation does away with the smelly tight coupling of the State-Pattern examples, as state is represented by the return type interface of each method, rather than as a property of the Sharlotka object. The application of the Single Responsibility Principle is rather less clear in this example, as the methods are stubbed out, rather than being implemented in any meaningful way. It is quite likely that in a real system the implementation of each method would be delegated to another object in order to honour this principle; this would look rather similar to the Small Interface implementation, with the crucial distinction that the implementation would not be responsible for updating the Sharlotka’s state.  The Liskov Substitution Principle ceases to be a question here, as we have moved to small interfaces with single implementations, and this fact also supports the Interface Segregation Principle.

Where this implementation is not so suitable is in cases where the state of the object after a method call is not clearly determined. For instance, withdrawing money from a bank account can leave the account in credit or overdrawn; in such a case the trick of returning a particular interface is not sufficient. A small example of this in my implementation is the Bake loop, and I have overcome the problem in this particular case by checking the return type until it is not null. However, this technique is already a significant departure from the fluent idiom, as it relies on runtime checking to make sure that the code actually works.

There is another danger with this implementation, in that it relies on the consumer knowing to use it fluently. There is no protection against storing the return value of any method in a variable, and calling the same method more than once (indeed, this is what is done during the Bake loop), and any person coding with this interface needs to know that methods must be chained. However, the fluent idiom is fairly commonplace now, and neither of these considerations is one of code quality.

& Messe it Forth

These three implementations illustrate different approaches to the same problem. The Classic implementation is unlikely to be the best in most circumstances because of its tendency to produce Liskov-violating defective methods; the Small Interface implementation overcomes these problems, and is probably most suitable for situations where the state of the object does no change in easily predictable ways; the Fluent implementation is handy when a particular sequence of methods should be called, but less idiomatic when the sequence can branch at runtime.

There are also tantalising prospects for implementing this type of system with monads, but I’m going to leave that for another day.