Android & Jacoco

January 01, 2015

Empirical Development

For a while now, I've been wanting to get code coverage working with MinimalBible, and it's finally at a point that I'm mostly satisfied with. I certainly wouldn't argue that it's in a "good" state, but it's enough for now. So, I wanted to write a post outlining how this was all set up since I haven't been able to find many people using this style. Plus, it took a very long time to set up, so I hope I can save someone else some pain in the future!

Before I get too much farther though, let me explain the concept of code coverage. Code coverage is intended to answer the question of "What code have I written tests for?" This allows you to quickly spot code that is untested by your existing suite, and let you know where is best to focus your time. Coverage statistics are generated line-by-line and branch-by-branch - if your test doesn't execute a specific branch of code during the test, your coverage system will let you know that not everything is "covered" by a test. All said, you get an easy way to see what potential problems exist and proactively solve them.

Test Setup

So, onward to how I set up testing in MinimalBible. First things first, I have to say that I really changed how I did testing in Android. Instead of using the existing Android build tools, I'm using pure-Java testing. Nothing Android. You can find the setup process here (massive shoutouts to Blundell Apps, this has been incredibly useful) but I'll give a quick overview of how this is set up.

The basic principle is that the testing support in Android is so broken as to be useless. I won't go into all my gripes, but suffice to say I have had issues with the JUnit API (Android uses something below 3.8), code coverage, and UI tests with Espresso. So instead of trying to use the native Android tools for testing, we create a new project based on the source code of our existing project. This allows us to run everything in a native Java environment, without worrying about any of the platform considerations. If you're interested in what that setup would look like, you can find my version here.

There are two projects: app and app-test. App is where the actual source code resides, and App-test is where the test code resides. However, the test code includes the original code as a dependency so that we can still write tests against it. I've included the important bits of app-test/build.gradle below:

apply plugin: 'java'

def androidModule = project(':app')
dependencies {
    compile androidModule
    testCompile androidModule.android.applicationVariants.toList().first().javaCompile.classpath // 1
    testCompile androidModule.android.applicationVariants.toList().first().javaCompile.outputs.files // 2
    testCompile files(androidModule.plugins.findPlugin("com.android.application").getBootClasspath()) // 3
    testCompile 'junit:junit:4.+' // 4
    testCompile 'org.robolectric:robolectric:2.2'
}

And a quick breakdown of what's going on:

  1. Make sure to include all dependencies of the actual project in the test project
  2. Include all code for the actual project in the test project.
  3. Include all dependencies of the android Gradle plugin in the test project. Not sure why this is here.
  4. Finally, include all the other libraries and things we actually need for testing.

So at this point we have a test-ready project setup. We're not quite ready for code coverage yet, but we're about to change that.

Next steps: Enabling Code Coverage

So, we have a test project set up to run the tests we put inside it. The next step is setting up Jacoco to report on what code is being tested during the test suite. I'm going to present the solution first to make it easy on anyone reading this - if you want a full explanation of what's going on check it out here.

In order to enable Jacoco testing we need to change app-test/build.gradle to look like this:

apply plugin: 'java'
apply plugin: 'jacoco'

def androidModule = project(':app')
def firstVariant = androidModule.android.applicationVariants.toList().first()

def testIncludes = [
    '**/*Test.class'
]
def jacocoExcludes = [
    'android/**',
    'org/bspeice/minimalbible/R*',
    '**/*$$*'
]

First steps first, this is the easy part. We're defining what files we want to include in testing, alongside the files we want to exclude from the final report. For example, we exclude the "R" file, since it's all generated code. In addition, anything containing a "$$" is generated by Dagger/Butterknife, so we ignore those too.

If you want to adapt the solution I'm outlining here to your own project, these should be the only sections you need to edit.

The next section is a whole lot more complicated:

dependencies {
    compile androidModule
    testCompile 'junit:junit:4.+'
    testCompile 'org.robolectric:robolectric:+'
    testCompile 'org.mockito:mockito-core:+'
    testCompile 'com.jayway.awaitility:awaitility:+'
    testCompile 'org.jetbrains.spek:spek:+'
    testCompile firstVariant.javaCompile.classpath
    testCompile firstVariant.javaCompile.outputs.files
    testCompile files(androidModule.plugins.findPlugin("com.android.application").getBootClasspath())
}
def buildExcludeTree(path, excludes) {
    fileTree(path).exclude(excludes)
} // 1
jacocoTestReport {
    doFirst {
        // First we build a list of our base directories
        def fileList = new ArrayList<String>()
        def outputsList = firstVariant.javaCompile.outputs.files
        outputsList.each { fileList.add(it.absolutePath.toString()) }
        
        // And build a fileTree from those
        def outputTree = fileList.inject { tree1, tree2 ->
            buildExcludeTree(tree1, jacocoExcludes) +
            buildExcludeTree(tree2, jacocoExcludes)
        }
        
        // And finally tell Jacoco to only include said files in the report
        classDirectories = outputTree
    } // 3
}
tasks.withType(Test) {
    scanForTestClasses = false
    includes = testIncludes
} // 2
  1. Define a quick function that will exclude a list of file paths from a given path.
  2. Set up Gradle to run the tests we defined earlier
  3. Set up Jacoco to exclude the paths we specified earlier from the report. This step is so complicated because we have to get the outputs paths from the Android project, and exclude our paths from each of those.

Wrapping Up

So given the above build.gradle file, we now have a project capable of testing your actual application code and producing coverage statistics on it. While I haven't outlined it above, because the testing code is separate from the Android project, you're free to write your tests in JUnit, Spock, or Spek. I'm going to be using Spek moving forward.

We can include tests using the testIncludes list, and make sure that classes don't get reported using the jacocoExcludes list. All said, that's what we were out for in the first place, so I'll call it a success.

If you want to take this solution further, the next step would be to add Robolectric tests into the suite, but I've been having issues with that too.

Appendix: Jacoco Full Explanation

In order to fully understand what's going on with how Jacoco excludes things from reporting, we have to step back and take a visit to Gradle first to understand your build lifecycle.

Gradle: Configure, Run

Gradle is an incredibly powerful tool, but it is massively confusing if you don't already know what you're doing. In my opinion, the documentation is still missing many examples that would be super-helpful, and is generally dense to try and get through.

That aside, to understand what's going on, you must understand that the Gradle build process happens in two phases: Configuration, and then Build.

For our purposes, you don't need to understand what each one does, but understanding the semantics is crucial. Because there's a two phase build, we can't write a build.gradle that tries to exclude files from Jacoco like this:

jacocoTestReport {
    // First we build a list of our base directories
    def fileList = new ArrayList<String>()
    def outputsList = firstVariant.javaCompile.outputs.files
    outputsList.each { fileList.add(it.absolutePath.toString()) }
    
    // And build a fileTree from those
    def outputTree = fileList.inject { tree1, tree2 ->
        buildExcludeTree(tree1, jacocoExcludes) +
        buildExcludeTree(tree2, jacocoExcludes)
    }
    
    // And finally tell Jacoco to only include said files in the report
    classDirectories = outputTree
}

Did you notice the difference? In the second example, we're missing the doFirst closure. Keep this in mind during the next sections.

Under the hood, Jacoco reports on all classes specified in the classDirectories variable. So, all we need to do is make sure that we include all the classes to report on in classDirectories, and exclude the ones we don't want to see.

However, if you skip the doFirst closure, you'll be in deep trouble. Without that closure, Groovy will run the code in the jacocoTestReport closure before testing is actually run, since it will be in the configuration build phase.

What the code actually does is exclude everything in jacocoExcludes from the global class path. This isn't a great solution, but I'm not sure how else to do it.

The problem comes when you exclude files like the android package that we don't want to report on, but are needed for testing. When things in android aren't loaded during the tests, you'll get lots of nasty NoClassDefFoundException exceptions, because Java can't find the code it needs for testing.

The solution? We need to modify the class path only right before Jacoco runs. This way, the tests are allowed to run successfully, and Jacoco never knows about those classes.

To do this, we need to move the class path configuration into the build phase instead of configure. The way to do that? You guessed it, surround the code in a doFirst closure.

So the end result is that we can exclude specific classes from reporting without interfering in the test setup process. It took me forever to figure out how exactly to implement this, but I hope this can help someone avoid the same issues in the future.

Side note: Much of the above solution was adapting the procedures outlined here to the world of Android. Thanks to everyone for putting in the effort to make it easier for me!

Robolectric + Android Lollipop

Getting actual Android code running on your computer----------------------------------------------------So far in the life of [MinimalBib...… Continue reading

Boundary Value Testing

Published on November 07, 2014

The State of Affairs

Published on November 06, 2014