When “Agile”, “Dynamic” and “Typeless” Become a Hindrance.

In recent years, I’ve seen the apparent rising popularity of “Agile” programming, powered by “dynamic” languages such as “Ruby”. While these things seem warm and fluffy at first, in the long term with large projects, they really can become a difficult beast to control.

Let’s take “Ruby on Rails” as an example. I was recently tasked with upgrading our version of Rails from 1.2.3 all the way up to 2.3.2, no easy task to be sure. We decided that a gradual, version-by-version upgrade would work the best, since it would allow us to catch deprecations and handle the framework changes a bite at a time.

What I discovered in the end was that Rails itself uses so much “magic”, and directly modifies so many core Ruby API’s that it was impossible to predict what would break between versions. As an example, a minor version change from 1.2.3 to 1.2.4 caused unexpected changes in how dates were handled. The change-logs made no reference to such modifications. The change from 2.1 to 2.2 resulted in a host of modules no longer being identified. Oh sure, if you printed it out, Ruby “thought” it was available, but the minute you tried to actually use anything within the module, you would receive “uninitialized constant” errors on the module name itself.

I know everyone has their own opinions on the matter, but from where I stand using Ruby on Rails on a system even vaguely large and complex simply raises far too many question marks on predictability and a clean upgrade path.

As a comparison, let’s refer to Java 2 Enterprise. Every piece of code I have ever written in J2EE version 1.0 runs without any issue under a J2EE version 3 container. Not even a re-compile was required. I do realise that the comparison is somewhat unfair – but the point still remains. I have very rarely needed to make major changes to any Java-based application to cater for a new framework – especially between minor version numbers.

I hate to make a blanket case for all agile/dynamic frameworks, so I’ll admit that was probably unfair (Grails, Django are on my to-do list). However, my case is even held fast on a more superficial level – my own benchmarks on a variety of languages clearly show that although time and effort may be saved on initial development time, dynamic/typeless languages will always suffer in raw speed, and agile frameworks will always suffer from a degree of unpredictability when upgrading (this becomes immensely exaggerated when “auto-magic” comes into play, as there is no clear path when anything does break).

To be clear, I won’t abandon agile frameworks – if I just want a simple website up, and expect maybe a dozen hits a month, I’m certainly not going to do it in a full blown enterprise container – but in the same token, I would certainly not use any such framework in a banking environment!

Partial Mocking/Stubbing and why it’s not evil.

Over the last few years, I’ve started expanding the places where I use automated testing (both test driven development and behaviour driven development, even though one would have to admit that in reality there’s little difference between them). I had only every written very simplistic JUnit tests for Java modules, but never really for the greater overall project. This changed when I starting working on a Ruby on Rails, where we were required to make extensive use of RSpec.

One of the things I learned to appreciate was the heavy use of stubbing (and expectations) that RSpec provided. Obviously, this was easier to do for a dynamic language such as Ruby (ignoring all the other shortcomings of the language itself).

When I started writing Java applications again, I obviously wanted to be able to perform similar tests with similar stubbings. To my dismay, the more pedantic amongst us programmers somehow decided that partial mocking/stubbing was somehow a truly terrible thing to do. I began reading all sort of justifications for why it was bad, and how one should refactor the class being tested in order to test it.

jMock and Mockito both adhered to this belief, EasyMock seemed less stringent, but stubbing was overly string-based (making refactorying later on a little annoying since tests would fail for no other reason than renaming a method).

Following is an example of the EasyMock mechanism (as shown in their documentation)

ToMock mock = createMock(ToMock.class,
   ToMock.class.getMethod("mockedMethod", null));

Unfortunately, all these folks fail to take into account the real world, and situation under which you code might not want to allow users to screw around with dependent objects. Here is a real example:

I created a class which did a load of work (let’s call it FlowClass), and part of this work was to use another class (let’s call it LoaderClass) which started up operating system processes. I needed to test the flow of FlowClass, but I didn’t want to start firing off all sorts of new processes. Easily enough, I was able to move the instantiation of LoaderClass into another method, but using something like jMock, how do I prevent that one single method from causing processes to load? I emphatically do not want the user to have to set instance of LoaderClass (like a JavaBean property), because I needed to keep strict control over the flow. The idea of having to introduce potential erratic program behaviour simply in order to test is completely unacceptable by any stretch of the imagination.

One site I came across suggested subclassing FlowClass directly in the test (even though, once more, the author makes all sort of claims of this being evil but delicious). This did the job, but it seems like so many lines of code to do nothing more than force a value to be returned. On top of that, what if I wanted to check to make sure the method was even ever called?

My needs eventually led me to create JavaStubs. It doesn’t do all the work that the other mocking frameworks do, but what it does do is allow me to quickly and easily stub method, including capturing exceptions for testing, and counting how many times they were called. I tried to be as string-less as possible (thank goodness for Java generics), and make it easily readable (much thanks to Mockito for suggestions on that one).