Mill Design Principles

Principles

Dependency graph first

Mill’s most important abstraction is the dependency graph of Tasks. Constructed using the T {…​} T.task {…​} T.command {…​} syntax, these track the dependencies between steps of a build, so those steps can be executed in the correct order, queried, or parallelized.

While Mill provides helpers like ScalaModule and other things you can use to quickly instantiate a bunch of related tasks (resolve dependencies, find sources, compile, package into jar, …​) these are secondary. When Mill executes, the dependency graph is what matters: any other mode of organization (hierarchies, modules, inheritance, etc.) is only important to create this dependency graph of Tasks.

Builds are hierarchical

The syntax for running targets from the command line mill Foo.bar.baz is the same as referencing a target in Scala code, Foo.bar.baz

Everything that you can run from the command line lives in an object hierarchy in your build.sc file. Different parts of the hierarchy can have different Targets available: just add a new def foo = T {…​} somewhere and you’ll be able to run it.

Cross builds, using the Cross data structure, are just another kind of node in the object hierarchy. The only difference is syntax: from the command line you’d run something via mill core.cross[a].printIt while from code you use core.cross("a").printIt due to different restrictions in Scala/Bash syntax.

Caching by default

Every Target in a build, defined by def foo = T {…​}, is cached by default. Currently this is done using a foo.json file in the out/ folder. The Target is also provided a foo.dest/ path on the filesystem dedicated to it, for it to store output files etc.

This happens whether you want it to or not. Every Target is cached, not just the "slow" ones like compile or assembly.

Caching is keyed on the .hashCode of the returned value. For Targets returning the contents of a file/folder on disk, they return PathRef instances whose hashcode is based on the hash of the disk contents. Serialization of the returned values is done using uPickle.

Functional Purity

Mill relies heavily on build targets being "pure": they only depend on their input targets, and their only output is their return value. They do not scribble all over the filesystem, reading and writing from random places. That is what allows us to be aggressive about caching and parallelizing the evaluation of build targets during a build.

Many kinds of build steps do require files on disk, and for that Mill provides the T.dest folder. This is a folder on disk dedicated to each build target, so that it can read and write things to it without worrying about conflicts with other targets that have their own T.dest folders. In effect, this makes even file output "pure": we can know precisely where a target’s output files live when we need to invalidate them, and it allows multiple targets all reading and writing to the filesystem to do so safely even when in parallel.

Short-lived build processes

The Mill build process is meant to be run over and over, not only as a long-lived daemon/console. That means we must minimize the startup time of the process, and that a new process must be able to re-construct the in-memory data structures where a previous process left off, in order to continue the build.

Re-construction is done via the hierarchical nature of the build: each Target foo.bar.baz has a fixed position in the build hierarchy, and thus a fixed position on disk out/foo/bar/baz.json. When the old process dies and a new process starts, there will be a new instance of Target with the same implementation code and same position in the build hierarchy: this new Target can then load the out/foo/bar/baz.json file and pick up where the previous process left off.

Minimizing startup time means aggressive caching, as well as minimizing the total amount of bytecode used: Mill’s current 1-2s startup time is dominated by JVM classloading. By default Mill uses a long-lived compile server to speed things up even more, but ensuring that the "from scratch" performance remains good is a core ongoing requirement.

Static dependency graph and Applicative tasks

Tasks are Applicative, not Monadic. There is .map, .zip, but no .flatMap operation. That means that we can know the structure of the entire dependency graph before we start executing Tasks. This lets us perform all sorts of useful operations on the graph before running it:

  • Given a Target the user wants to run, pre-compute and display what targets will be evaluated ("dry run"), without running them

  • Automatically parallelize different parts of the dependency graph that do not depend on each other, perhaps even distributing it to different worker machines like Bazel/Pants can

  • Visualize the dependency graph easily, e.g. by dumping to a DOT file

  • Query the graph, e.g. "why does this thing depend on that other thing?"

  • Avoid running tasks "halfway": if a Target’s upstream Targets fail, we can skip the Target completely rather than running halfway and then bailing out with an exception

In order to avoid making people using .map and .zip all over the place when defining their Tasks, we use the T {…​}/T.task {…​}/T.command {…​} macros which allow you to use Task#apply() within the block to "extract" a value.

def test() = T.command {
  TestRunner.apply(
   "mill.UTestFramework",
   runDepClasspath().map(_.path) :+ compile().path,
   Seq(compile().path)

}

This is roughly equivalent to the following:

def test() = T.command { T.zipMap(runDepClasspath, compile, compile) {
  (runDepClasspath1, compile2, compile3) =>
  TestRunner.apply(
    "mill.UTestFramework",
    runDepClasspath1.map(_.path) :+ compile2.path,
    Seq(compile3.path)
  )
}

This is similar to SBT’s :=/.value macros, or scala-async's async/await. Like those, the T {…​} macro should let users program most of their code in a "direct" style and have it "automatically" lifted into a graph of Tasks.

How Mill aims for Simple

Why should you expect that the Mill build tool can achieve simple, easy & flexible, where other build tools in the past have failed?

Build tools inherently encompass a huge number of different concepts:

  • What "Tasks" depends on what?

  • How do I define my own tasks?

  • Where do source files come from?

  • What needs to run in what order to do what I want?

  • What can be parallelized and what can’t?

  • How do tasks pass data to each other? What data do they pass?

  • What tasks are cached? Where?

  • How are tasks run from the command line?

  • How do you deal with the repetition inherent in a build? (e.g. compile, run & test tasks for every "module")

  • What is a "Module"? How do they relate to "Tasks"?

  • How do you configure a Module to do something different?

  • How are cross-builds (across different configurations) handled?

These are a lot of questions to answer, and we haven’t even started talking about the actually compiling/running any code yet! If each such facet of a build was modelled separately, it’s easy to have an explosion of different concepts that would make a build tool hard to understand.

Before you continue, take a moment to think: how would you answer to each of those questions using an existing build tool you are familiar with? Different tools like SBT, Fake, Gradle or Grunt have very different answers.

Mill aims to provide the answer to these questions using as few, as familiar core concepts as possible. The entire Mill build is oriented around a few concepts:

  • The Object Hierarchy

  • The Call Graph

  • Instantiating Traits & Classes

These concepts are already familiar to anyone experienced in Scala (or any other programming language…), but are enough to answer all of the complicated build-related questions listed above.

The Object Hierarchy

Diagram

The module hierarchy is the graph of objects, starting from the root of the build.sc file, that extend mill.Module. At the leaves of the hierarchy are the Targets you can run.

A Target's position in the module hierarchy tells you many things. For example, a Target at position core.test.compile would:

  • Cache output metadata at out/core/test/compile.json

  • Output files to the folder out/core/test/compile.dest/

  • Source files default to a folder in core/test/, core/test/src/

  • Be runnable from the command-line via mill core.test.compile

  • Be referenced programmatically (from other Targets) via core.test.compile

From the position of any Target within the object hierarchy, you immediately know how to run it, find its output files, find any caches, or refer to it from other Targets. You know up-front where the Targets data "lives" on disk, and are sure that it will never clash with any other Targets data.

The Call Graph

Diagram

The Scala call graph of "which target references which other target" is core to how Mill operates. This graph is reified via the T {…​} macro to make it available to the Mill execution engine at runtime. The call graph tells you:

  • Which Targets depend on which other Targets

  • For a given Target to be built, what other Targets need to be run and in what order

  • Which Targets can be evaluated in parallel

  • What source files need to be watched when using --watch on a given target (by tracing the call graph up to the Sources)

  • What a given Target makes available for other Targets to depend on (via its return value)

  • Defining your own task that depends on others is as simple as def foo = T {…​}

The call graph within your Scala code is essentially a data-flow graph: by defining a snippet of code:

val b = ...
val c = ...
val d = ...
val a = f(b, c, d)

you are telling everyone that the value a depends on the values of b c and d, processed by f. A build tool needs exactly the same data structure: knowing what Target depends on what other Targets, and what processing it does on its inputs!

With Mill, you can take the Scala call graph, wrap everything in the T {…​} macro, and get a Target-dependency graph that matches exactly the call-graph you already had:

def b = T { ... }
def c = T { ... }
def d = T { ... }
def a = T { f(b(), c(), d()) }

Thus, if you are familiar with how data flows through a normal Scala program, you already know how data flows through a Mill build! The Mill build evaluation may be incremental, it may cache things, it may read and write from disk, but the fundamental syntax, and the data-flow that syntax represents, is unchanged from your normal Scala code.

Instantiating Traits & Classes

Classes and traits are a common way of re-using common data structures in Scala: if you have a bunch of fields which are related and you want to make multiple copies of those fields, you put them in a class/trait and instantiate it over and over.

In Mill, inheriting from traits is the primary way for re-using common parts of a build:

  • Scala "project"s with multiple related Targets within them, are just a Trait you instantiate

  • Replacing the default Targets within a project, making them do new things or depend on new Targets, is simply override-ing them during inheritance

  • Modifying the default Targets within a project, making use of the old value to compute the new value, is simply override`ing them and using `super.foo()

  • Required configuration parameters within a project are abstract members

  • Cross-builds are modelled as instantiating a (possibly anonymous) class multiple times, each instance with its own distinct set of Targets

In normal Scala, you bundle up common fields & functionality into a class you can instantiate over and over, and you can override the things you want to customize. Similarly, in Mill, you bundle up common parts of a build into traits you can instantiate over and over, and you can override the things you want to customize. "Subprojects", "cross-builds", and many other concepts are reduced to simply instantiating a trait over and over, with tweaks.

Prior Work

SBT

Mill is built as a substitute for SBT, whose problems are described here. Nevertheless, Mill takes on some parts of SBT (builds written in Scala, Task graph with an Applicative "idiom bracket" macro) where it makes sense.

Bazel

Mill is largely inspired by Bazel. In particular, the single-build-hierarchy, where every Target has an on-disk-cache/output-folder according to their position in the hierarchy, comes from Bazel.

Bazel is a bit odd in its own right. The underlying data model is good (hierarchy + cached dependency graph) but getting there is hell. It (like SBT) is also a 3-layer interpretation model, but layers 1 & 2 are almost exactly the same: mutable python which performs global side effects (layer 3 is the same dependency-graph evaluator as SBT/mill).

You end up having to deal with a non-trivial python codebase where everything happens via:

do_something(name="blah")

or

do_other_thing(dependencies=["blah"])

where "blah" is a global identifier that is often constructed programmatically via string concatenation and passed around. This is quite challenging.

Having the two layers be “just python” is great since people know python, but I think it’s unnecessary to have two layers ("evaluating macros" and "evaluating rule impls") that are almost exactly the same, and I think making them interact via return values rather than via a global namespace of programmatically-constructed strings would make it easier to follow.

With Mill, I’m trying to collapse Bazel’s Python layer 1 & 2 into just 1 layer of Scala, and have it define its dependency graph/hierarchy by returning values, rather than by calling global-side-effecting APIs. I’ve had trouble trying to teach people how-to-bazel at work, and am pretty sure we can make something that’s easier to use.

Scala.Rx

Mill’s "direct-style" applicative syntax is inspired by my old Scala.Rx project. While there are differences (Mill captures the dependency graph lexically using Macros, Scala.Rx captures it at runtime), they are pretty similar.

The end-goal is the same: to write code in a "direct style" and have it automatically "lifted" into a dependency graph, which you can introspect and use for incremental updates at runtime.

Scala.Rx is itself build upon the 2010 paper Deprecating the Observer Pattern.

CBT

Mill looks a lot like CBT. The inheritance based model for customizing Modules/ScalaModules comes straight from there, as does the "command line path matches Scala selector path" idea. Most other things are different though: the reified dependency graph, the execution model, the caching module all follow Bazel more than they do CBT