Output Build Target

26 April 2007

In the past few days I've been reviewing an in-progress article by a Julian Simpson, a colleague of mine, on refactoring ant files. Julian is one our "deployment dudes" who've been responsible for applying our agile-oriented work habits to the deployment of systems. In doing this Julian has run into more than his fair share of gnarly ant build scripts. His article is a nice description of some his favorite ways to clean up the mess.

One of his observations particularly intrigued me, pointing out that ant files tend to decompose into targets named after tasks (e.g. "compile", "unit-test"). These tasks often get combined into imperative target invocations

<target name="cruise" depends="clean, compile, copy-static-files, 
unit-test, publish, javadoc, tag"/>

This kind of style smells wrong because it's a mismatch with the dependency-based computational model that build files are based on. In using this you look at each target, ask what it depends on, and put those dependencies into the target. So here your unit-test task should depend on the compile and copy-static-files targets directly and shouldn't be mentioned by a later target.

When faced with a peice of apparant idiocy the easy thing to do is make some flippant comment like "maybe they have such a simple system that they can run their tests without compiling". But often it's worth delving a little deeper.

Both Julian and I have a Unix background, so we are familiar with the make build language. Like Ant, make uses dependency-based programming. Both systems break down their work into targets linked through dependency relationships. But there's a subtle difference. In ant you give a target a name and indicate which other targets you depend on. In make, however, you state an output file, the input files it depends on, and how to get from one to the other.

So compiling a two file hello world (in C) in make looks something like this

hello : hello.o greet.o
  gcc -o hello hello.o greet.o

hello.o : hello.c 
  gcc -c hello.c

greet.o: greet.c
  gcc -c greet.c

The first line says that the target file hello depends on the files hello.o and greet.o. The second line tells you how to make the target from the source files (and in a real make file much of this is general rules so you don't have to duplicate them). We then have similar rules to show how to make the inputs to the top rule.

In ant (with Java) looks like this

<project name = "hello" default = "test">
  <target name = "compile">
    <javac 
      srcdir = "src"
      destdir = "build"
    />
  </target>
  <target name = "test" depends = "compile">
    <junit printsummary="on">
      <classpath location = "build"/>
      <test name = "Tester"/>
    </junit>
  </target>   
</project>

Here we say we have a test task that depends on the compile task, meaning that the compile task needs to be run at some point before we run the test task.

Both of these fragments demonstrate dependencies but this difference in naming affects how we think about them. I wonder how that naming convention pulls your brain away from thinking about dependencies and into thinking about imperative rules.

Another aspect of this is how we deal with unneccessary work. Builds always seem to take longer than we'd like, so we do a great deal to avoid doing things we don't need to do. Indeed this is the whole point of using dependency based programming in the first place. Even if a target is mentioned multiple times, it will only be run once.

But this desire to avoid unnecessary work also goes into running the targets themselves. The fundamental mechanism in make for avoiding work is to compare the modification times of the output and input files. If the output is younger than the youngest input, then it's clearly up to date and doesn't need to be re-made.

Ant doesn't do this, instead it relies on ant tasks doing their own analysis to see what needs to be run. This makes a lot of sense when a compiler needs to figure out which output classes need to be rebuilt after a change. Often, however, people write targets themselves that don't do some kind of checking. With any target you should be sure that it won't do any significant work unless something could have changed; focusing on the output is a good way to do this.

Sometimes the output isn't obvious. What is the output of a unit-test task? Putting my make hat on, I'd say "some sort of file". You could have the test report be the result and check in the task that test report file was older than any of the inputs - although that may not be as easy as you would like. The report needn't even contain anything, it could be just a TouchFile.

Are my unix-make roots going overboard here? Possibly. But I would suggest that you look at your build files and to consider:

  • Name each target after its product (not the activity it does).
  • Make sure a target does no significant work if it doesn't need to.
  • State each target's dependencies directly, let the build system work out the imperative sequence to follow.