Why Use Mill?

Overall across our benchmarks, Mill is 4-10x faster than Maven for clean compiles, both parallel and sequential, and for many modules or for a single module:

First, let’s look at Parallel Clean Compile All. This benchmark involves running clean to delete all generated files and re-compiling everything in parallel. Mill sees a significant ~6x speedup over Maven for this benchmark. You can click on the link above to see a more detailed discussion of how this benchmark was run.

The second benchmark worth noting is Incremental Compile Single Module. This benchmark involves making a single edit to a single already-compiled file in common - adding a single newline to the end of the file - and re-compiling common and common.test. Mill sees a huge ~12x speedup for this benchmark, because Mill’s incremental compiler (Zinc) is able to detect that only one file in one module has changed, and that the change is small enough to not require other files to re-compile. In contrast, Maven re-compiles all files in both modules, even though only one file was touched and the change was trivial.

The comparison with Gradle is less stark, but still significant. Mill is 2-4x faster than Gradle across the various workflows:

Mill’s various "clean compile" workflows 3-4x faster than Gradle’s, while it’s incremental and no-op compile workflows are 2x faster. Both Gradle and Mill appear to do a good job limiting the compilation to only the changed file, but Mill has less fixed overhead than Gradle does, finishing in about ~0.5s rather than ~1.5 seconds.

In general, these benchmarks don’t show Mill doing anything that Maven or Gradle do not: these are equivalent builds for the same projects (Netty and Mockito respectively), compiling the same number of files using the same Java compiler, in the same module structure and passing the same suite of tests. Rather, what we are seeing is Mill simply having less build-tool overhead than Maven or Gradle, so the performance of the underlying JVM and Java compiler (which is actually pretty fast!) can really shine through.

All Mill runs generate some debugging metadata files in out/mill-*. One of these is out/mill-chrome-profile.json, which is a file following the Chrome Profiling format. It can be loaded into any Chrome browser’s built in chrome://tracing UI, to let you interactively explore what Mill was doing during its last run. e.g. when performing a clean compile on the Netty codebase, the profile ends up looking like this:

The Chrome profile shows what task each Mill thread was executing throughout the run. The Chrome profiling UI is interactive, so you can zoom in and out, or click on individual tasks to show the exact duration and other metadata.

But the real benefit of the Chrome profile isn’t the low-level data it provides, but the high-level view:

This waiting is likely due to dependencies between them, and they take long enough that all the other tasks depending on them get held up. For example, when codec.compile finishes above, we can see a number of downstream tasks immediately start running.

This understanding of your build’s performance profile is not just an academic exercise, but provides actionable information:

Most build tools do provide some way of analyzing build performance, but none of them provide it as easily as Mill does: any Mill run generates a profile automatically, and any computer with Chrome on it is able to load and let you explore that profile. That is a powerful tool to help engineers understand what the build is doing: any engineer who felt a build was slow can trivially load it into their Chrome browser to analyze and figure out what.

Apart from the Mill Chrome Profile, Mill also provides the ./mill visualize command, which is useful to show the logical dependency graph between tasks. For example, we can use ./mill visualize __.compile (double underscore means wildcard) to show the dependency graph between the modules of the Netty build below:

(Right-click open-image-in-new-tab to see full size)

In this graph, we can clearly see that common.compile, buffer.compile, transport.compile, and codec.compile depend on each other in a linear fashion. This explains why they each must wait for the prior task to complete before starting, and cannot run in parallel with one another. Furthermore, we can again confirm that many of the codec-*.compile tasks depend on codec.compile, which is in the profile why we saw them waiting for the upstream task to complete before starting.

Although these are things we could have guessed from looking at the Chrome Profile above, ./mill visualize gives you a separate angle from which to look at your build. Together these tools can help give greater understanding of what your build is doing and why it is doing that: something that can be hard to come by with build tools that are often considered confusing and inscrutable.

One area that Mill does better than Gradle is providing a seamless IDE experience. For example, consider the snippet below where we are using Gradle to configure the javac compiler options. Due to .gradle files being untyped Groovy, the autocomplete and code-assist experience working with these files is hit-or-miss. In the example below, we can see that IntelliJ is able to identify that compileArgs exists and has the type List<String>:

But if you try to jump to definition or find out anything else about it you hit a wall:

Often working with build configurations feels like hitting dead ends: if you don’t have options.compilerArgs memorized in your head, there is literally nothing you can do in your editor to make progress to figure out what it is or what it is used for. That leaves you googling for answers, which can be a frustrating experience that distracts you from the task at hand.

The fundamental problem with tools like Gradle is that the code you write does not actually perform the build: rather, you are just setting up some data structure that is used to configure the real build engine that runs later. Thus when you explore the Gradle build in an IDE, the IDE can only explore the configuration logic (the getCompilerArgs method above) and is unable to explore the actual build logic (how getCompilerArgs actually gets used in Gradle)

In comparison, Mill’s .mill files are all statically typed, and as a result IntelliJ is easily able to pull up the documentation for def javacOptions, even though it doesn’t have any special support for Mill built into the IDE:

Apart from static typing, the way Mill builds are structured also helps the IDE: Mill code actually performs your build, rather than configuring some opaque build engine. While that sounds academic, one concrete consequence is that IntelliJ is able to take your def javacOptions override and find the original definitions that were overridden, and show you where they are defined:

You can jump to any of the overriden `def`s quickly and precisely:

Furthermore, because task dependencies in Mill are just normal method calls, IntelliJ is able to find usages, showing you where the task is used. Below, we can see the method call in the def compile task, which uses javacOptions() along with a number of other tasks:

From there, if you are curious about any of the other tasks used alongside javacOptions, it’s easy for you to pull up their documentation, jump to their definition, or find their usages. For example we can pull up the docs of compileClasspath() below, jump to its implementation, and continue interactively exploring your build logic from there:

Unlike most other build tools, Mill builds can be explored interactively in your IDE. If you do not know what something does, it’s documentation, definition, or usages is always one click away in IntelliJ or VSCode. This isn’t a new experience for Java developers, as it is what you would be used to day-to-day in your application code! But Mill brings that same polished experience to your build system - traditionally something that has been opaque and hard to understand - and does so in a way that no other build tool does.

The following Mill build is a minimal Java module foo. It contains no custom configuration, and so inherits all the defaults from mill.javalib.JavaModule: default source folder layout, default assembly configuration, default compiler flags, and so on.

If you want to add a custom task, this is as simple as defining a method e.g. def lineCount = Task { …​ }. The body of Task performs the action we want, and can depend on other tasks such as allSourceFiles() below:

Once we define a new task, we can immediately begin using it in our build. lineCount is not used by any existing JavaModule tasks, but we can still show its value via the Mill command line to force it to evaluate:

Note that as lineCount is a Task, we get automatic caching, invalidation, and parallelization: these are things that every Task gets for free, without the task author to do anything. And although we wrote the lineCount logic in the main build.mill file for this example, if it grows complex enough to get messy it is easy to move it to your own custom plugins

To wire up lineCount into our main JavaModule compile/test/run tasks, one way is to take the line count value and write it to a file in def resources. This file can then be read at runtime as a JVM resource. We do that below by overriding def resources and making it depend on lineCount, in addition to its existing value super.resources():

Because our def resources overrides the existing resources method inherited from JavaModule, the downstream tasks automatically now use the new override instead, as that is how overrides work. That means if you call mill foo.run, it will automatically pick up the new resources including the generated line-count.txt file and make it available to the application code to use e.g. to print it out at runtime:

Next, we’ll look at a more realistic example, which includes usage of third-party libraries in the build.

Earlier on we discussed possibly pre-rendering HTML pages in the build so they can be served at runtime. The use case for this are obvious: if a page never changes, rendering it on every request is wasteful, and even rendering it once and then caching it can impact your application startup time. Thus, you may want to move some HTML rendering to build-time, but with traditional build tools such a move is sufficiently inconvenient and complicated that people do not do it.

With Mill, pre-rendering HTML at build time is really easy, even if you need a third-party library. Mill does not ship with a bundled HTML templating engine, but you can use the import $ivy syntax to include one such as Thymeleaf, which would immediately make the Thymeleaf classes available for you to import and use in your build as below:

Once we have run import $ivy, we can import TemplateEngine, Context, and replace our def lineCount with a def htmlSnippet task that uses Thymeleaf to render HTML. Again, we get full IDE support for working with the Thymeleaf Java API, the new htmlSnippet task is inspectable from the Mill command line via show, and we wire it up into def resources so it can be inspected and used at runtime by the application (in this case just printed out):

Rendering HTML using the Thymeleaf templating engine is not rocket science, but what is interesting here is what we did not need to do:

Instead, we could simply import Thymeleaf directly from Maven Central and use it just like we would use it in any Java application, with IDE support, typechecking, and automatic parallelism and caching.

Most real projects require some kind of ad-hoc build tasks: you may be pre-processing static assets for web deployment, embedding build metadata for runtime debugging, or pre-rendering HTML pages to optimize performance at runtime. With most build tools, you often needed to pull in some poorly-maintained plugin off of Github, write your own using a complicated plugin framework, or even wrap your build tool in ad-hoc shell scripts. With most other build tools, caching and parallelism are things that the build author needs to use manually, meaning nobody gets it right and your build performance is never as good as it could be.

In contrast, Mill makes it easy it is to write concise type-checked code to perform ad-hoc tasks to do whatever you need to do. You get full IDE support, automatic caching and parallelism, and access to the huge JVM library ecosystem on Maven Central. Rather than grabbing unmaintained plugins off of Github or augmenting your build with fragile shell scripts, Mill allows your own custom logic to be implemented in a way that is flexible, performant, and safe, such that anyone can configure their build correctly and achieve maximum performance even without being a build tool expert.