public void perform(String operatorName) {
      BigDecimal op1 = stack.pop();

      if ("+".equals(operatorName)) {
         BigDecimal op2 = stack.pop();
         stack.push(op1.add(op2));
         currentMode = Mode.inserting;
      } else if ("-".equals(operatorName)) {
         BigDecimal op2 = stack.pop();
         stack.push(op2.subtract(op1));
         currentMode = Mode.inserting;
      } else if ("!".equals(operatorName)) {
         op1 = op1.round(MathContext.UNLIMITED);
         BigDecimal result = BigDecimal.ONE;
         while (op1.compareTo(BigDecimal.ONE) > 0) {
            result = result.multiply(op1);
            op1 = op1.subtract(BigDecimal.ONE);
         }
         stack.push(result);
      } else {
         throw new MathOperatorNotFoundException();
      }
   }

Unlike the last example, I’ll provide the entire class. Feel free to make changes to this class as well. However, for now focus on the perform(...) method.

One note, Philip Schwarz recommended a change to what I proposed to avoid the command/query separation violation. I applied his recommendation before posting this updated version.

The Whole Class

package com.scrippsnetworks.calculator;

import java.math.BigDecimal;
import java.math.MathContext;

public class RpnCalculator {
   private OperandStack stack = new OperandStack();
   private Mode currentMode = Mode.accumulating;

   enum Mode {
      accumulating, replacing, inserting
   };

   public RpnCalculator() {
   }

   public void take(BigDecimal value) {
      if (currentMode == Mode.accumulating)
         value = determineNewTop(stack.pop(), value);

      if (currentMode == Mode.replacing)
         stack.pop();

      stack.push(value);
      currentMode = Mode.accumulating;
   }

   private BigDecimal determineNewTop(BigDecimal currentTop, BigDecimal value) {
      BigDecimal newTopValue = currentTop;
      String digits = value.toString();
      while (digits.length() > 0) {
         newTopValue = newTopValue.multiply(BigDecimal.TEN);
         newTopValue = newTopValue.add(new BigDecimal(Integer.parseInt(digits
               .substring(0, 1))));
         digits = digits.substring(1);
      }

      return newTopValue;
   }

   public void enter() {
      stack.dup();
      currentMode = Mode.replacing;
   }

   public void perform(String operatorName) {
      BigDecimal op1 = stack.pop();

      if ("+".equals(operatorName)) {
         BigDecimal op2 = stack.pop();
         stack.push(op1.add(op2));
         currentMode = Mode.inserting;
      } else if ("-".equals(operatorName)) {
         BigDecimal op2 = stack.pop();
         stack.push(op2.subtract(op1));
         currentMode = Mode.inserting;
      } else if ("!".equals(operatorName)) {
         op1 = op1.round(MathContext.UNLIMITED);

         BigDecimal result = BigDecimal.ONE;
         while (op1.compareTo(BigDecimal.ONE) > 0) {
            result = result.multiply(op1);
            op1 = op1.subtract(BigDecimal.ONE);
         }
         stack.push(result);
      } else {
         throw new MathOperatorNotFoundException();
      }
   }

   public BigDecimal getX() {
      return stack.x();
   }

   public BigDecimal getY() {
      return stack.y();
   }

   public BigDecimal getZ() {
      return stack.z();
   }

   public BigDecimal getT() {
      return stack.t();
   }
}

Today we picked one method in particular and attempted to work our way through it. This method was a classic legacy method (no unit tests). It also had a switch on type and then it also did one or more things based on a set of options. All of this was in one method.

If you read Fowler’s Refactoring Book, it mentions a combination of encapsulating the type code followed by replacing switch with polymorphism for the first problem in this method (the switch). We were able to skip encapsulating the type code since we wanted to keep the external API unchanged (legacy code).

So we first created a base strategy for the switch and then several empty derived classes, one for each of the enums. This is a safe legacy refactoring because it only involved adding new code.

Next, we created a factory to create the correct strategy based on the type code and added that to the existing method (we also added a few virtual methods). Again, a safe refactoring since it only involved adding effectively unused code (we did create the factory using nearly strict TDD). Finally, we delegated from the original method to the strategy returned from the factory. Safe again, since we had tested the factory.

So far, so good. But next, we wanted to push the method down to each of the subclasses and remove the parts of the logic that did not apply to each given type. We did a spike refactoring to see what that’d be like and it was at least a little dicey. We finally decided to get the original method under test so that as we refactored, we had the safety net necessary to refactor with confidence.

We started by simply calling the method with null’s and 0 values. We worked our way through the method, adding hand-rolled test doubles until we came across our first static class.

• Added an instance of the class as a static member (essentially creating a singleton).
• Added a property setter and getter (making it an overridable singleton).
• We then copied the body of the static method into an instance method, which we made virtual.
• We then delegated the static method to the virtual method.
• Then, in the unit test, we set the singleton to a hand-coded test double in the setup and reset the singleton in the tear down method.

We had to do this several times and on the third time (I think it was the third time), the hand-rolled test double would have had to implement several (17ish) methods and it became clear that we were ready to use a mocking framework. They are using Moq so we started using Moq to accomplish the remainder of the mocking.

After some time, we managed to get a test that essentially sent a tracer bullet through one path of the method we wanted to get under test. When the test turned green there was much rejoicing.

• This method currently makes calls to the service layers and those calls depend on both an enumeration (replaced with a shallow and wide hierarchy of strategies) and options (to be replaced with a composition of strategies).
• It also changes some values in an underling domain object.

So that’s what we needed to characterize.

• Begin as we already had, get through the method with a tracer.
• Determine the paths we want to get under test.
• For each path:
  • Create a test using strict mocks (which fail as soon as an unexpected method is called)
  • Use a Setup to document this call as expected – this is essentially one of the assertions for the characterization test.
  • Continue until we have all the Setups required to get through the test.
  • Add any final assertions based on state-based checks and call VerifyAll on the Moq-based mock object.

This would be a way we could work through the method and characterize it before we start refactoring it in earnest.

This might sound like a lot of work and it certainly is no cake walk, but all of this work was done by one of the attendees and as a group they certainly have the expertise to do this work. And in reality, it did not take too long. As they practice and get some of the preliminary work finished, this will be much easier.

• Spent time on one project practicing TDD and refactoring to patterns (they implemented 5 of the GoF patterns).
• Spent time practicing some of Fowler’s refactorings and Feather’s legacy refactorings.
• Spent a day practicing TDD using mocks for everything but the unit under test. At the end they had a test class, one production class and several interfaces.

In retrospect, the work they did in the first three days was nearly exactly what they needed to practice for the final day of effort. When we started tackling their legacy code, they had already practiced everything they used in getting the method under test.

So overall great week with a fun group of guys in Canada.

I say this because of recent trends in our industry and my hands-on experience with many enterprise and Internet applications, mostly at client sites. You might be thinking that I’m referring to the mainstream breakout of Functional Programming (FP), which is happening right now. The killer app for FP is concurrency. We’ve all heard that more and more applications must be concurrent these days (which doesn’t necessarily mean multithreaded). When we remove side effects from functions and disallow mutable variables, our concurrency issues largely go away. The success of the Actor model of concurrency, as used to great effect in Erlang, is one example of a functional-style approach. The rise of map-reduce computations is another example of a functional technique going mainstream. A related phenomenon is the emergence of key-value store databases, like BigTable and CouchDB, is a reaction to the overhead of SQL databases, when the performance cost of the Relational Model isn’t justified. These databases are typically managed with functional techniques, like map-reduce.

But actually, I’m thinking of something else. Hybrid languages like Scala, F#, and OCaml have demonstrated that OOP and FP can complement each other. In a given context, they let you use the idioms that make the most sense for your particular needs. For example, immutable “objects” and functional-style pattern matching is a killer combination.

What’s really got me thinking that objects are losing their supremacy is a very mundane problem. It’s a problem that isn’t new, but like concurrency, it just seems to grow worse and worse.

The problem is that there is never a stable, clear object model in applications any more. What constitutes a BankAccount or Customer or whatever is fluid. It changes with each iteration. It’s different from one subsystem to another even within the same iteration! I see a lot of misfit object models that try to be all things to all people, so they are bloated and the teams that own them can’t be agile. The other extreme is “balkanization”, where each subsystem has its own model. We tend to think the latter case is bad. However, is lean and mean, but non-standard, worse than bloated, yet standardized?

The fact is, for a lot of these applications, it’s just data. The ceremony of object wrappers doesn’t carry its weight. Just put the data in a hash map (or a list if you don’t need the bits “labeled”) and then process the collection with your iterate, map, and reduce functions. This may sound heretical, but how much Java code could you delete today if you replaced it with a stored procedure?

These alternatives won’t work for all situations, of course. Sometimes polymorphism carries its weight. Unfortunately, it’s too tempting to use objects as if more is always better, like cow bell.

So what would replace objects for supremacy? Well, my point is really that there is no one true way. We’ve led ourselves down the wrong path. Or, to be more precise, we followed a single, very good path, but we didn’t know when to take a different path.

Increasingly, the best, most nimble designs I see use objects with a light touch; shallow hierarchies, small objects that try to obey the Single Responsibility Principle, composition rather than inheritance, etc. Coupled with a liberal use of functional idioms (like iterate, map, and reduce), these designs strike the right balance between the protection of data hiding vs. openness for easy processing. By the way, you can build these designs in almost any of our popular languages. Some languages make this easier than others, of course.

Despite the hype, I think Domain-Specific Languages (DSLs) are also very important and worth mentioning in this context. (Language-Oriented Programming – LOP – generalizes these ideas). It’s true that people drink the DSL Kool-Aid and create a mess. However, when used appropriately, DSLs reduce a program to its essential complexity, while hiding and modularizing the accidental complexity of the implementation. When it becomes easy to write a user story in code, we won’t obsess as much over the details of a BankAccount as they change from one story to another. We will embrace more flexible data persistence models, too.

Back to OOP and FP, I see the potential for their combination to lead to a rebirth of the old vision of software components, but that’s a topic for another blog post.

For example, here is a Scala example that uses a trait to trace calls to a Worker.work method.

    
// run with "scala example.scala" 

class Worker {
    def work() = "work" 
}

trait WorkerTracer extends Worker {
    override def work() = "Before, " + super.work() + ", After" 
}

val worker = new Worker with WorkerTracer

println(worker.work())        // => Before, work, After
    

Note that WorkerTracer extends Worker so it can override the work method. Since Scala is statically typed, you can’t just define an override method and call super unless the compiler knows there really is a “super” method!

Here’s a Ruby equivalent.

    
# run with "ruby example.rb" 

module WorkerTracer
    def work; "Before, #{super}, After"; end
end

class Worker 
    def work; "work"; end
end

class TracedWorker < Worker 
  include WorkerTracer
end

worker = TracedWorker.new

puts worker.work          # => Before, work, After
    

Note that we have to create a subclass, which isn’t required for the Scala case (but can be done when desired).

If you know that you will always want to trace calls to work in the Ruby case, you might be tempted to dispense with the subclass and just add include WorkerTracer in Worker. Unfortunately, this won’t work. Due to the way that Ruby resolves methods, the version of work in the module will not be found before the version defined in Worker itself. Hence the subclass seems to be the only option.

However, we can work around this using metaprogramming. We can use WorkerTracer#append_features(...). What goes in the argument list? If we pass Worker, then all instances of Worker will be effected, but actually we’ll still have the problem with the method resolution rules.

If we just want to affect one object and work around the method resolution roles, then we need to pass the singleton class (or eigenclass or metaclass ...) for the object, which you can get with the following expression.

    
metaclass = class << worker; self; end
    

So, to encapsulate all this and to get back to the original goal of implementing with-style semantics, here is an implementation that adds a with method to Object, wrapped in an rspec example.

    
# run with "spec ruby_with_spec.rb" 

require 'rubygems'
require 'spec'

# Warning, monkeypatching Object, especially with a name
# that might be commonly used is fraught with peril!!

class Object
  def with *modules
    metaclass = class << self; self; end
    modules.flatten.each do |m|
      m.send :append_features, metaclass
    end
    self
  end
end

module WorkerTracer
    def work; "Before, #{super}, After"; end
end

module WorkerTracer1
    def work; "Before1, #{super}, After1"; end
end

class Worker 
    def work; "work"; end
end

describe "Object#with" do
  it "should make no changes to an object if no modules are specified" do
    worker = Worker.new.with
    worker.work.should == "work" 
  end

  it "should override any methods with a module's methods of the same name" do
    worker = Worker.new.with WorkerTracer
    worker.work.should == "Before, work, After" 
  end

  it "should stack overrides for multiple modules" do
    worker = Worker.new.with(WorkerTracer).with(WorkerTracer1)
    worker.work.should == "Before1, Before, work, After, After1" 
  end

  it "should stack overrides for a list of modules" do
    worker = Worker.new.with WorkerTracer, WorkerTracer1
    worker.work.should == "Before1, Before, work, After, After1" 
  end

  it "should stack overrides for an array of modules" do
    worker = Worker.new.with [WorkerTracer, WorkerTracer1]
    worker.work.should == "Before1, Before, work, After, After1" 
  end
end
    

You should carefully consider the warning about monkeypatching Object! Also, note that Module.append_features is actually private, so I had to use m.send :append_features, ... instead.

The syntax is reasonably intuitive and it eliminates the need for an explicit subclass. You can pass a single module, or a list or array of them. Because with returns the object, you can also chain with calls.

A final note; many developers steer clear of metaprogramming and reflection features in their languages, out of fear. While prudence is definitely wise, the power of these tools can dramatically accelerate your productivity. Metaprogramming is just programming. Every developer should master it.

The Liskov Substitution Principle was coined by Barbara Liskov in Data Abstraction and Hierarchy (1987).

If for each object o1 of type S there is an object o2 of type T such that for all programs P defined in terms of T, the behavior of P is unchanged when o1 is substituted for o2, then S is a subtype of T.

I’ve always liked the elegant simplicity, yet power, of LSP. In less formal terms, it says that if a client (program) expects objects of one type to behave in a certain way, then it’s only okay to substitute objects of another type if the same expectations are satisfied.

This is our best definition of inheritance. The well-known is-a relationship between types is not precise enough. Rather, the relationship has to be behaves-as-a, which unfortunately is more of a mouthful. Note that is-a focuses on the structural relationship, while behaves-as-a focuses on the behavioral relationship. A very useful, pre-TDD design technique called Design by Contract emerges out of LSP, but that’s another topic.

Note that there is a slight assumption that I made in the previous paragraph. I said that LSP defines inheritance. Why inheritance specifically and not substitutability, in general? Well, inheritance has been the main vehicle for substitutability for most OO languages, especially the statically-typed ones.

For example, a Java application might use a simple tracing abstraction like this.

    
public interface Tracing {
    void trace(String message);
}
    

Clients might use this to trace methods calls to a log. Only classes that implement the Tracer interface can be given to these clients. For example,

    
public class TracerClient {
    private Tracer tracer;

    public TracerClient(Tracer tracer) {
        this.tracer = tracer;
    }

    public void doWork() {
        tracer.trace("in doWork():");
        // ...
    }
}
    

However, Duck Typing is another form of substitutability that is commonly seen in dynamically-typed languages, like Ruby and Python.

If it walks like a duck and quacks like a duck, it must be a duck.

Informally, duck typing says that a client can use any object you give it as long as the object implements the methods the client wants to invoke on it. Put another way, the object must respond to the messages the client wants to send to it.

The object appears to be a “duck” as far as the client is concerned.

In or example, clients only care about the trace(message) method being supported. So, we might do the following in Ruby.

    
class TracerClient 
  def initialize tracer 
    @tracer = tracer
  end

  def do_work
    @tracer.trace "in do_work:" 
    # ... 
  end
end

class MyTracer
  def trace message
    p message
  end
end

client = TracerClient.new(MyTracer.new)
    

No “interface” is necessary. I just need to pass an object to TracerClient.initialize that responds to the trace message. Here, I defined a class for the purpose. You could also add the trace method to another type or object.

So, LSP is still essential, in the generic sense of valid substitutability, but it doesn’t have to be inheritance based.

Is Duck Typing good or bad? It largely comes down to your view about dynamically-typed vs. statically-typed languages. I don’t want to get into that debate here! However, I’ll make a few remarks.

On the negative side, without a Tracer abstraction, you have to rely on appropriate naming of objects to convey what they do (but you should be doing that anyway). Also, it’s harder to find all the “tracing-behaving” objects in the system.

On the other hand, the client really doesn’t care about a “Tracer” type, only a single method. So, we’ve decoupled “client” and “server” just a bit more. This decoupling is more evident when using closures to express behavior, e.g., for Enumerable methods. In our case, we could write the following.

    
class TracerClient2 
  def initialize &tracer 
    @tracer = tracer
  end

  def do_work 
    @tracer.call "in do_work:" 
    # ... 
  end
end

client = TracerClient2.new {|message| p "block tracer: #{message}"}
    

For comparison, consider how we might approach substitutability in Scala. As a statically-typed language, Scala doesn’t support duck typing per se, but it does support a very similar mechanism called structural types.

Essentially, structural types let us declare that a method parameter must support one or more methods, without having to say it supports a full interface. Loosely speaking, it’s like using an anonymous interface.

In our Java example, when we declare a tracer object in our client, we would be able to declare that is supports trace, without having to specify that it implements a full interface.

To be explicit, recall our Java constructor for TestClient.

    
public class TracerClient {
    public TracerClient(Tracer tracer) { ... }
    // ...
    }
}
    

    
class ScalaTracerClient(val tracer: { def trace(message:String) }) {
    def doWork() = { tracer.trace("doWork") }
}

class ScalaTracer() {
    def trace(message: String) = { println("Scala: "+message) }
}

object TestScalaTracerClient {
    def main() {
        val client = new ScalaTracerClient(new ScalaTracer())
        client.doWork();
    }
}
TestScalaTracerClient.main()
    

Recall from my previous blogs on Scala, the argument list to the class name is the constructor arguments. The constructor takes a tracer argument whose “type” (after the ’:’) is { def trace(message:String) }. That is, all we require of tracer is that it support the trace method.

So, we get duck type-like behavior, but statically type checked. We’ll get a compile error, rather than a run-time error, if someone passes an object to the client that doesn’t respond to tracer.

To conclude, LSP can be reworded very slightly.

If for each object o1 of type S there is an object o2 of type T such that for all programs P defined in terms of T, the behavior of P is unchanged when o1 is substituted for o2, then S is substitutable for T.

I replaced a subtype of with substitutable for.

An important point is that the idea of a “contract” between the types and their clients is still important, even in a language with duck-typing or structural typing. However, languages with these features give us more ways to extend our system, while still supporting LSP.

Several colleagues and I really groked Fusion. It started with system sequence diagrams showing interactions much like the specification I mentioned above. It also described a difference between analysis and design (and if Uncle Bob reads this, he’ll probably have some strong words about so-called object-oriented analysis, well this was 15 years ago… though I still there there is some value to be found there). Anyway, this is mostly about system sequence diagrams so I won’t say much more about that excellent process.

Each system sequence diagram represented a scenario. To represent many scenarios, Fusion offered a BNF-based syntax to express those various scenarios. (I understand that this was also because for political reasons within HP they were not allowed to include a state model, but I don’t know if that is true or not.) For several years I practiced Fusion and really I often revert back to that if I’m not trying to do anything in particular.

Spending a little time up front thinking a little about the logical interaction between the system and its various actors helps me get a big picture view of the system boundary and its general flow. I have also found it helps others as well, but your mileage may vary.

So when I viewed their protocol specification, it really brought back some good memories. And in fact, that’s how we decided to look at the problem.

(What follows is highly-idealized)

We reviewed their specification and decided we’d try to work through the initialization sequence and then work through one sequence that involved “completing a simple job.” I need to keep this high level to keep the identity of the company a secret.

There was prior work and we kept that in mind but really started from scratch. In our very first attempt, there had been some work done along the lines of using the Command pattern, so we started there. Of course, once we did our first command, we backed off and when with a more basic design that seemed to fit the complexity a bit better (starting with the command pattern at the beginning is an example of solution-problemming to use a Weinberg term – and one of the reasons I’m sometimes skeptical when people start talking in patterns).

• Receive a message from the outside as some array of bytes
• Determine the “command” represented by the bytes
• Process the parameters within the command
• Issue a message to the simulator
• Create a logical response
• Format the logical response into the underlying protocol
• Send the response back

At the time, they were considering using JNI, so we spent just over a day validating that we could communicate bi-directionally, maintaining a single process space.

• It is easy to end up with a bunch of tested objects but no fully-connected system
• Sharing setup between various mocks is difficult and often not done so there’s a lot of violation of DRY
• You have to learn a new syntax

We accepted learning a new syntax because it was deemed less painful than maintaining existing hand-rolled test doubles (though there are several reasonable solution for that, ask if you want to know what it is). There is the issue of sharing setup on mocks, but we did not have enough work yet to really notice that as a problem. However, they were at least aware of that and we briefly discussed how to share common expectation-setting (it’s well supported).

• Could be tested up to but excluding the JNI stuff
• Could be configured to stub out JNI or use real JNI
• Was easily configurable
• Was automatically configured by C++ (since it was a C++ process that was started to get the whole system in place)

We designed that (15 minute white-board session), coded it and ended up with a few integration tests. Along the way, we built a simple factory for creating the system fully connected. That factory was used both in tests as well as by the JNI-based classes to make sure that we had a fully-connected systems when it was finally started by C++.

Near the end, we decided we wanted to demonstrate asynchronous computation, which we did using tests. I stumbled a bit but we got it done in a few hours. We demonstrated that the system receiving messages from the outside world basically queued up requests rather than making the sender wait synchronously (we demonstrated this indirectly – that might be a later blog post – let me know if you’re interested).

By the way, that was the first week. These guys were good and I had a great time.

There was still a little work to be done on the C++ side and I only had a week, so I asked them to keep me posted. The following Tuesday they had the first end-to-end interaction, system initialization.

By Wednesday (so 3 business days later), they had a complete demonstration of end-to-end interaction with a single, simple job finishing. Not long after that they demonstrated several simple jobs finishing. The next thing on their list? Completing more complex jobs, system configuration, etc.

• Select a system interaction
• List all of the steps it needs to accomplish (some requests required a response, some did not)
• Write unit tests for each “arm” of the interaction

Select a set of end-to-end interactions that add value to the user of the system

For each system-level interaction, enumerate all of its steps and then add implementing those steps as individual back-log items

Now some of those individual steps will end up being small (less than an hour) but some will be quite large when they start working with variable parameters and commands that need to operate at a higher priority.

But they are well on their way and I was reminded of just how much I really enjoyed using Fusion.

End-to-end automated integration tests are a well-established way of catching these problems early on. The team I’m mentoring has set up its own continuous-integration (CI) server and the team is getting pretty good at writing acceptance tests using FitNesse. However, these tests only cover the components provided by the team, not the true end-to-end user stories. So, they are imperfect as both acceptance tests and integration tests. Our longer-term goal is to automate true end-to-end acceptance and integration tests, across all components and services.

In this particular case, the other team is following a waterfall-style of development, with big design up front. Therefore, my team needed to give them an interface to design against, before we were ready to actually implement the service.

There are a couple of problems with this approach. First, the two teams should really “pair” to work out the interface and behavior across their components. As I said, we’re just starting to go Agile, but my goal is to have virtual feature teams, where members of the required component teams come together as needed to implement features. This would help prevent the miscommunication of one team defining an interface and sharing it with another team through documentation, etc. Getting people to communicate face-to-face and to write code together would minimize miscommunication.

Second, defining a service interface without the implementation is risky, because it’s very likely you will miss important details. The best way to work out the details of the interface is to test drive it in some way.

This suggests another technique I want to introduce to the team. When defining an interface for external consumption, don’t just deliver the “static” interface (source files, documentation, etc.), also deliver working Mock Objects that the other team can test against. You should develop these mocks as you test drive the interface, even if you aren’t yet working on the full implementation (for schedule or other reasons).

The mocks encapsulate and enforce the behavioral contract of the interface. Design by Contract is a very effective way of thinking about interface design and implementing automated enforcement of it. Test-driven development mostly serves the same practical function, but thinking in “contractual” terms brings clarity to tests that is often missing in many of the tests I see.

Many developers already use mocks for components that don’t exist yet and find that the mocks help them design the interfaces to those components, even while the mocks are being used to test clients of the components.

Of course, there is no guarantee that the mocks faithfully represent the actual behavior, but they will minimize surprises. Whether you have mocks or not, there is no substitute for running automated integration tests on real components as soon as possible.

The first project is using Adobe Flex to create a rich interface. The team decided to adopt FunFX for acceptance testing. You write your tests in Ruby, typically using Test::Unit or RSpec.

FunFX places some constraints on your Flex application. You have to define the GUI objects in MXML, the XML-based file format for Flex applications, rather than ActionScript, and you need to add ids to all elements you want to reference.[1]

These are reasonable constraints and the first constraint promotes better quality, in fact. The MXML format is more succinct (despite the XML “noise”) and declarative than ActionScript code. This is almost always true of UI code in most languages (with notable exceptions…). Declarative vs. imperative code tends to improve quality because less code means fewer bugs, less to maintain, and it frees the implementor of the declarative “language” to pick the best implementation strategies, optimizations, etc. This characteristic is typical of Functional Languages and well-designed Domain Specific Languages, as well.

I don’t think you can underestimate the benefit of writing less code. I see too many teams whose problems would diminish considerably if they just got rid of duplication and learned to be concise.

The second project is a wiki-based application written in Java. To make deployment as simple as possible, the implementors avoided the Servlet API (no need to install Tomcat, etc.) and rolled their own web server and page rendering components. (I’m not sure I would have made these decisions myself, but I don’t think they are bad, either…)

The rendering components are object-oriented and use a number of design patterns, such as page factories with builder objects that reflect the “widgets” in the UI, HTML tags, etc. This approach makes the UI very testable with JUnit and FitNesse. In fact, the development process was a model of test-driven development.

However, the final result is flawed! It is much too difficult to change the look and feel of the application, which is essential for most UI’s, especially web UI’s. The project made the wrong tradeoffs; the design choices met the requirements of TDD very well, but they made maintenance and enhancement expensive and tedious. The application is now several years old and it has become dated, because of the expense of “refreshing” the look and feel.

What should have been done? These days, most dynamic web UI’s are built with templating engines, of which there are many in the most common programming languages. Pages defined in a templating engine are very declarative, except for the special tags where behavior is inserted. The pages are easy to change. It is mostly obvious where a particular visual element is generated, since most of the “tags” in the template look exactly like the tags in the rendered page. “Declarative” templates, like good DSL’s, can be read, understood, and even edited by the stakeholders, in this case the graphical designers.

But how do you test these page templates? When test-driving UI’s it is important to decide what to test and what not to test. The general rule for TDD is to test anything that can break. The corollary, especially relevant for UI’s, is don’t test anything when you don’t care if it changes.

It is usually the dynamic behavior of the UI that can break and should be tested. Templating engines provide special tags for inserting dynamic behavior in the underlying language (Java, Ruby, etc.). This is what you should test. It is usually best to keep the scripts in these tags as small as possible; the scripts just delegate to code, which can be test-driven in the usual way.

I see too many UI tests that compare long strings of HTML. These tests break whenever someone makes a minor look and feel or other inconsequential change. Part of the art of UI TDD is knowing how to test just what can break and nothing more. In the second project, incidental changes to the UI break tests that should be agnostic to such changes.

To conclude, keep your UI’s as declarative as you can. Only test the “declarations” (e.g., templates) in areas where they might break, meaning if it changes, it’s a bug. You’ll get the full benefits of TDD and the freedom to change the UI easily and frequently, as needed.

¹ Disclaimer: my information on FunFX is second hand, so I might not have the details exactly correct; see the FunFX documentation for details.