(Coursenotes for CSC 305 Individual Software Design and Development)

Strategy and command design patterns

Reading:

Strategy design pattern

Example: A routing application

Suppose you’re working on a maps application (kind of like Google Maps or Apple Maps). Your application has a graph that represents geographic locations, and it allows the user to request a driving route from point A to point B. So put simply your app can do something like this:

List<Point> path = this.pathfinder.computePath(graph, source, dest);

The computePath function above finds a driving path from source to dest and life is good. You now get a request that the app should also allow people to request walking paths or biking paths. You could implement these as separate methods, e.g.,

this.pathfinder.computeWalkingPath(...)

this.pathfinder.computeBikingPath(...)

But this can get unwieldy—your pathfinder object is now supporting three behaviours that are, by definition, distinct from each other. This may indicate a lack of cohesion. Worse, any clients that rely on this pathfinder would have to now have code added to them that can optionally call the correct path finding algorithm. Effectively, each new pathfinding strategy you support will also result in a bunch of additions to other classes that need to consume the pathfinding feature.

What SOLID principle are we violating? How?

The Strategy design pattern tries to solve this problem. It is useful when you have a class that accomplishes a specific task in a lot of different ways, and extract these different algorithms into their own routines are called strategies.

It requires a few pieces:1

  • The Strategy interface: This is the common interface to which all your individual strategies will adhere. In this case you might have a PathingStrategy interface with a computePath abstract method.
  • Concrete strategy implementations: You could have several classes implement the PathingStrategy interface. So for example, DrivingStrategy, WalkingStrategy, PublicTransportStrategy would all have their own computePath implementations.
  • The Client: This is the consumer of the various strategies. The client initialises a new concrete strategy and uses their computePath method to achieve this task. Exactly which strategy object is initialised is based on decisions made at runtime (e.g., the user selects a specific strategy). What’s important is that the client’s code doesn’t change when the chosen strategy changes.

So the strategy pattern helps us go from having a bunch of different methods (computeDrivingPath, computeWalkingPath, computeBikingPath) in the same class to having one method implemented (computePath) in a few different ways. You might be wondering why the latter way is deemed preferable.

With dependency injection, the Client is able to be totally decoupled from the specific strategy it is using. Without the common StrategyInterface binding it all together, each new strategy added in the future would require code added in various points in the Client (everywhere the strategy was used). This problem becomes compounded when the “strategies” have multiple behaviours (e.g., route finding, but also things like checking tolls or estimating time).

When to use the strategy design pattern:

  • When you have multiple ways of accomplishing a task and you want to be able to swap between those ways at runtime.
  • You want to isolate details of how an algorithm works from the client that uses it.

Cons of using this pattern:

  • It can be overkill for just a couple of strategies. That’s why in project 2 I’ve suggested that you can simply use a Function object whose value is determined programmitically, instead of creating a Parser interface with two separate implementations for prefix and postfix notation.
  • Refactoring guru says that another con is that clients must be aware of the differences between strategies in order to select the right one. I don’t think this is a con. This is just the cost of doing business—we wouldn’t have different strategies if there weren’t some difference between them that is noticeable to someone.

Command design pattern

The Command design pattern is used to turn a “request” into a standalone object. I’m not going to talk too much about details here because this pattern mostly tries to bring “functions as values” to languages that don’t have that ability.

The classical Command pattern requires two pieces:

  • A Command interface that has an execute abstract method.
  • Concrete implementations that implement the execute method.

Voila! You have a function that can be stored in a variable, given to other functions, etc. This is what lambdas let you do in Java. Just use lambdas.

In fact, this is how lambdas are implemented in Java. Each lambda is really an instance of a functional interface; a Java interface that has only one abstract method (e.g., Function, Predicate, Consumer, Comparator). The only difference is instead of creating a whole new class that implements the interface, just to implement that one function, you use the lambda syntax to concisely implement just the function (which is usually all you care about anyway).

Undo-able operations

One benefit of using the Command pattern (other than all the benefits of functions as values) is that you can support things like undo-able operations. Imagine a function that could be applied as well as reversed. You can do this in two ways.

First, each command simply saves a copy of its input. That way its “undo” operation is to return give back that old version (before the execute function was applied). This is not always feasible.

Another way is to define the Command interface with an execute method and also an undo method. That way each “command” that you create must also define its reverse operation.

In either case, you wouldn’t just use a lambda, since each “command” requires two functions, not just one. This is a common operation for database migrations. A database migration is when some change is made to the schema of a database. For example, suppose we’re maintaining Cal Poly’s student management system, and we want to add a field to the Student table, indicating whether they first enrolled in Cal Poly on the quarter system or the semester system.

In production databases that are in active use and already have live data, a migration is not as simple as tacking on the new column. You need to define what should be done for existing rows in the table. This is what would be defined in the execute function. You also need to write a “rollback” operation, because you want this change to be atomic. Your table has millions of rows; if the execute function fails for some row in the middle, you don’t want to quit now with your table in two inconsistent states. So your undo function tells the table how to roll back the change.

Undo-able operations are not always possible to support. For example, the old states might be too large to feasibly save for each command. Or the undo operation might be impossible to implement (e.g., encryptions).


  1. Note that the Refactoring Guru page on this pattern includes a fourth piece: the Context. This sits in between the Client and the Strategy interface. I…don’t know why they recommend this. There is such a thing as too much abstraction. In this case, the Context class adds (what I think is) a needless layer between the Client and the Strategy interface, so I recommend ignoring that part.