I believe that the work on keeping quality high should start from the very beginning of the project. When it comes to actual implementation setting up build configuration is the very first thing one makes. The choice of tools has a huge impact on the process and results.
Additionally, the build itself is a program as well (and an important one!), so there is no excuse in avoiding good practices like readability, DRY, SOLID, etc.
That is why in this post I want to write down some good ideas about SBT usage that I’ve learned in both commercial and my own small projects, that help me write better code, keep build maintainable and improve projects in general.
In case of simpler projects we should find out that our project follows Maven-like layout similar to:
/src should be obvious for everyone who ever worked on projects with Maven-ish directory structure. We have 2 directories here,
/src/test, which in turn group source code by languages (so Java files would be under
/java subdirectory, Scala files within
/scala, etc) and resources that are in
resources directory (there are exceptions like Android build configuration, but we’ll leave that for another day).
/project are more interesting to us. The former is the most important file to lookup by SBT when we run
sbt command within
/project is kind of a second class citizen here: we can use it to empower
build.sbt file and make sure that version of SBT used to build project will be consistent in all environments.
Simple build definition could look like this one:
What we see here is DSL created using SBT magic. As a matter of fact it is somewhat restricted Scala subset with several implicit imports already made for us. Those properties may look as something mutable, but underneath they are actually immutable values! If we import project into our favorite IDE with SBT support we can check that all of those are actually
sbt.SettingKey instances, and operators like
++= are used to create modified copies of those keys. Those keys are then used underneath as arguments for something similar to
project.settings(settings1, settings2), which returns modified instance of immutable project. So despite the mutable looking DSL everything stays immutable at the core.
Those pieces of information are quite useful when we consider multiproject. What are the reasons to do that? For me personally it’s about keeping things simple: it is easier to work on project when some responsibilities are clearly separated. Because order of compilation and direction of dependency is clearly defined we can use modules to enforce concepts like layered architecture, hexagonal architecture and (to a degree) boundary contexts.
Of course it comes with a price: maintenance of such build could be more complex and (as for now) SBT has trouble with caching dependency resolutions, meaning that checking libraries might take a while. However, I have seen more than once that keeping things tiny and clean is definitely worth it.
Basic setup of a multiproject would look like that:
Of course, we have to make sure that there are modules
modules directory which are following same Maven conventions like singular build described before.
That setup will load aggregating project on
sbt named after our directory:
If we wanted to have more control over it, we can create it explicitly:
root project on start as expected.
Let us stop here for a moment. When we list projects with
sbt projects we’ll get:
How exactly SBT determined names for those? In earlier versions we have to define them explicitly with:
but currently we can use
project macro which would lookup name of
val and use it to populate module identifier and location:
Notice, that macro requires
val here. We cannot just pass reference into some utility function and hope things to work. As such
project is useful only to initiate
Project definition, that we will customize from now on.
It is difficult to overlook that something repeats in our configuration:
That doesn’t look good and easily can lead to some errors. For instance, a moment ago I forgot to copy paste
moduleB. What happened?
B were build using different versions of Scala and as a result dependency couldn’t be resolved. This would never happen if settings common to all projects could be somehow shared, right? Let us try to create our first file within
Then we can refer to common settings with:
Settings have a signature
def settings(ss : sbt.Project.SettingsDefinition*) : sbt.Project which is the reason we have to use vararg type ascription
:_* to adjust
Another way of defining settings (better suited for things specific to a module) is… putting another
build.sbt in module’s directory. Personally I try to keep all common settings and dependencies within
project/* and use
modules/*/build.sbt for libraries used only in one module. One has also keep in mid that
project directory could be used only with root project. In case of modules it will be ignored.
One can also try to remove top level
build.sbt completely and instead create build object like this:
As a matter of the fact that way of defining modules was/is quite popular in a lot of open source projects. However newer versions of SBT deprecated them and now
build.sbt is the only option if we decide on newest versions.
While I saw some projects rely on
Common.scala approach I also saw some (more compelling) where this blob was split into something more self-explanatory, like
Settings. For instance something that I would use in my own project:
This way Scala version (and standard library), dependencies and resolvers would be kept in one place and separated from the settings. I’ve separated test dependencies from main ones to make sure that we won’t rely on unit test frameworks on production (
libraryDependencies ++= testDeps map (_ % "test")). I’ve also added some
scalac compiler options to enforce better quality of the code.
While we’re at testing we can also think about some small improvements. By default we have access to
test task which would run JUnit/Scalatest/Specs2/whatever framework is fancy at a time. But is it enough? Making CI run tests only informs us that no test got broken, it doesn’t say how much of the code is checked.
Code coverage tools are great way to figure out which part of the codebase should get special attention. When you see that some critical part of your application is severely untested you might start to worry and this should motivate you to throw some tests there. Mind, that a number itself is meaningless. What would be the point of 100% coverage of a module made entirely of POJOs or plain case classes? We should use coverage values reasonably, to decide which parts of the code needs special attention, which require more testing but requiring any level of coverage should be something that responsible programmers decide themselves. We all know that any form of coverage forced on developers against their will, would just lead up to meaningless tests that touch everything and check nothing. ;)
Ad rem. Configuring test coverage in SBT cannot be done out of the box. But it can be provided via SBT plugins. For this article I’ll use SCoverage, but there are plenty others to choose from.
First, let’s make sure that everyone running our project would use the same SBT version - similarly how Scala libraries’ packages are bound to specific Scala versions, SBT plugins are bound to SBT releases. And we would want other devs to just run build, not fight against it. We can define fixed SBT version by providing
project/build.properties file with content like:
Then we can provide plugins for SBT within
From this moment we can access SCoverage settings within our build definitions. In single module builds coverage would be enabled automatically. In multimodule however, we have to enable it in each module individually:
If we also want to enable coverage measurement by default (which I do NOT recommend, but let’s leave it for now) we can configure it with:
Now we can measure coverage by running the following command:
for a single build or
for a multiproject.
Why that way? Why not with just one command? Well, there are limitations to tools used to measure coverage. First, they are measuring coverage within files modified/recompiled since list rebuild (or at least they appear to). As a result you’ll often get coverage values that would make no sense unless you clean build prior to measurement. Second, they have to be manually instructed to start tuning - this can be worked around by settings
coverageEnabled := true as shown above, but as a side effect running application by
sbt run might cause application to fail since it will still try to load some (absent in normal runtime) coverage dependency (and that’s why I recommend against it, and so does the author of a plugin). The last one is the need to manual trigger of coverage report documentation.
After that you read reports under
target/scoveragre-report directory. You can also define minimal coverage for build to pass on CI using options like:
but as I said, first make sure that your team agrees. It is also worth knowing that coverage of some parts of code could be disabled with:
so that all kind of safe code (case classes, etc) or code that couldn’t be reasonably tested (once you finish extracting deps, you eventually end up with some place where you ultimately gather all the instances and inject them into the components) would not cause any disturbance.
I have seen few big and successful projects. What they had in common were the developers that wanted to keep quality high on each level. That means they all had their style guidelines that everyone was obligated to follow - but hardly anyone would try to learn formatting rules by heart! Instead each of those projects relied on some automatic formatter that was not subject to opinion or mistake. Simply - you work here, your code will be formatted with X, EOT. That got rid of all discussions about indentations, where spaces should go and where shouldn’t and with good tests covering large part of the projects, reviewers could actually focus on more important things: whether code makes sense, whether it will be maintainable in the future, does it leave no place for misunderstandings etc. That is why some people consider defining formatter for the project as a rule 0 of project configuration.
What I used with great success was a combination of Scalariform and Scalastyle. The former is a formatter that (by default) runs on each compilation (which means that as long as our developers commit code they actually run we have consistent codebase with no additional effort). The latter is a style checker. As those two don’t cover exactly the same elements of style guide they compliment each other. For instance, by default Scalariforms might merge some lines into one line (it doesn’t have a sense of line length limit, unfortunately), then Scalastyle might catch that and let us know that we need to handle this specific case ourselves (I admit that no line-length-limit is the greatest weakness of Scalariform).
To use them we start by adding plugins to SBT (again, in
(some versions of SBT would break if we hadn’t put those empty lines between plugins). Then Scalariform can be configured with:
Scalastyle has slightly different approach to configuration - it uses
scalastyle-config.xml file. We can generate it with
sbt scalastyleGenerateConfig command and then edit to our hearts content. Once we’re done we can check style with
If you’re as crazy for quality as I am you would like a build to fail if style is not up to standards. You can achieve that by configuring:
and marking all offending warnings as errors within
In case something break here and you don’t want it to be fixed (because you e.g. don’t agree with the tools on this particular case) you can suppress tools with:
Here we showed how we can start up (or improve) SBT project with modules that would clearly define direction of dependencies between different parts of it, highlight architecture and add several tasks that would help us keeping code quality high. We would just run
make sure that tests pass, coverage is high enough, style guidelines are followed and code reviewers would be able to focus on the important stuff, the one that no automate could check for us.