Your shiny new Scala build tool!
Run unit test suite:
sbt core/test
Build a standalone executable jar:
sbt bin/test:assembly
Now you can re-build this very same project using the build.sc file, e.g. re-run core unit tests
e.g.:
./bin/target/mill Core.compile
./bin/target/mill Core.test.compile
./bin/target/mill Core.test
./bin/target/mill ScalaPlugin.assembly
There is already a watch
option that looks for changes on files, e.g.:
./bin/target/mill --watch Core.compile
Output will be generated into a the ./out
folder.
If you are repeatedly testing Mill manually by running it against the build.sc
file in the repository root, you can skip the assembly process and directly run
it via:
sbt "~bin/test:run Core.test"
sbt "~bin/test:run --repl"
Lastly, you can generate IntelliJ Scala project files using Mill via
./bin/target/mill idea
Allowing you to import a Mill project into Intellij without using SBT
Mill provides a build REPL, which lets you explore the build interactively and
run Target
s from Scala code:
lihaoyi mill$ bin/target/mill --repl
Loading...
Compiling (synthetic)/ammonite/predef/interpBridge.sc
Compiling (synthetic)/ammonite/predef/replBridge.sc
Compiling (synthetic)/ammonite/predef/DefaultPredef.sc
Compiling /Users/lihaoyi/Dropbox/Workspace/mill/build.sc
Compiling /Users/lihaoyi/Dropbox/Workspace/mill/out/run.sc
Compiling (synthetic)/ammonite/predef/CodePredef.sc
@ build
res0: build.type = build
@ build.
!= Core ScalaPlugin bridges getClass isInstanceOf |>
== MillModule asInstanceOf equals hashCode toString
@ build.Core
res1: Core.type = ammonite.predef.$up.build$Core$@5600c124
@ Core
res2: Core.type = ammonite.predef.$up.build$Core$@5600c124
@ Core.scalaV
scalaVersion
@ Core.scalaVersion
res3: define.Target[String] = ammonite.predef.^.build.MillModule#scalaVersion@4b7d16
@ Core.scalaVersion()
Running Core.scalaVersion
res4: String = "2.12.4"
@ Core.ivyDeps()
Running Core.ivyDeps
res5: Seq[scalaplugin.Dep] = List(
Scala(
Dependency(
Module("com.lihaoyi", "sourcecode", Map()),
"0.1.4",
...
@ Core.ivyDeps().foreach(println)
Running Core.ivyDeps
Scala(Dependency(com.lihaoyi:sourcecode,0.1.4,,Set(),Attributes(,),false,true))
Scala(Dependency(com.lihaoyi:pprint,0.5.3,,Set(),Attributes(,),false,true))
Point(Dependency(com.lihaoyi:ammonite,1.0.3,,Set(),Attributes(,),false,true))
Scala(Dependency(com.typesafe.play:play-json,2.6.6,,Set(),Attributes(,),false,true))
Scala(Dependency(org.scala-sbt:zinc,1.0.5,,Set(),Attributes(,),false,true))
Java(Dependency(org.scala-sbt:test-interface,1.0,,Set(),Attributes(,),false,true))
Into a build.sc
file you can define separate Module
s (e.g. ScalaModule
).
Within each Module
you can define 3 type of task:
Target
: take no argument, output is cached and should be serializable; run frombash
(e.g.def foo = T{...}
)Command
: take serializable arguments, output is not cached; run frombash
(arguments withscopt
) (e.g.def foo = T.command{...}
)Task
: take arguments, output is not cached; do not run frombash
(e.g.def foo = T.task{...}
)
You can use SBT to build a Mill executable, which itself is able to build more Mill executables that can you can use to run Mill commands:
git clean -xdf
# Build Mill executable using SBT
sbt bin/test:assembly
# Build Mill executable using the Mill executable generated by SBT
bin/target/mill devAssembly
# Build Mill executable using the Mill executable generated by Mill itself
out/devAssembly devAssembly
Eventually, as Mill stabilizes, we will get rid of the SBT build entirely and rely on previous versions of Mill to build itself.
In case of troubles with caching and/or incremental compilation, you can always
restart from scratch removing the out
directory:
rm -rf out/
A lot of mills design principles are intended to fix SBT's flaws, as described in http://www.lihaoyi.com/post/SowhatswrongwithSBT.html. Before working on Mill, read through that post to understand where it is coming from!
Mill's most important abstraction is the dependency graph of Task
s.
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, inheritence, etc.) is only important to create this
dependency graph of Task
s.
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
Target
s 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.
Every Target
in a build, defined by def foo = T{...}
, is cached by default.
Currently this is done using a foo.mill.json
file in the out/
folder. The
Target
is also provided a foo/
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 Target
s
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 tentatively done using uPickle.
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.mill.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.mill.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. In future, we may have a long lived console or nailgun/drip-based server/client models to speed up interactive usage, but we should always keep "cold" startup as fast as possible.
Task
s 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 Task
s. 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 Task
s, 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 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 Task
s.
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?
- 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 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 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 Target
s 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.mill.json
-
Output files to the folder
out/Core/test/compile/
-
Be runnable from the command-line via
mill Core.test.compile
-
Be referenced programmatically (from other
Target
s) viaCore.test.compile
From the position of any Target
within the object hierarchy, you immediately
know how to run it, find it's output files, find any caches, or refer to it from
other Target
s. You know up-front where the Target
's data "lives" on disk, and
are sure that it will never clash with any other Target
's data.
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
Target
s depend on which otherTarget
s -
For a given
Target
to be built, what otherTarget
s need to be run and in what order -
Which
Target
s 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 theSource
s) -
What a given
Target
makes available for otherTarget
s to depend on (via it's 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 Target
s, 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:
val b = T{ ... }
val c = T{ ... }
val d = T{ ... }
val 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.
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
Target
s within them, are just aTrait
you instantiate -
Replacing the default
Target
s within a project, making them do new things or depend on newTarget
s, is simplyoverride
-ing them during inheritence. -
Modifying the default
Target
s within a project, making use of the old value to compute the new value, is simplyoverride
ing them and usingsuper.foo()
-
Required configuration parameters within a
project
areabstract
members. -
Cross-builds are modelled as instantiating a (possibly anonymous) class multiple times, each instance with it's own distinct set of
Target
s
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
trait
s 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.
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.
Mill is largely inspired by Bazel. In particular, the single-build-hierarchy, where every Target has an on-disk-cache/output-directory according to their position in the hierarchy, comes from Bazel.
Bazel is a bit odd in it’s 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 unnecessary two 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 it’s 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
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.
Mill looks a lot like CBT. The inheritance based
model for customizing Module
s/ScalaModule
s 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
The end goal of the Mill project is to develop a new Scala build tool to replace SBT. Mill should satisfy most of the current use cases for SBT's functionality, but hopefully needing much fewer features and much less complexity to do so. We take inspiration from SBT, Make, Bazel, and many other existing build tools and libraries.
The immediate goal of Mill is to be feature-complete enough to:
- Sustain it's own development, without needing SBT
- Start porting over existing open-source Scala library builds from SBT to Mill
com-lihaoyi/mill#2 would kick off the process porting
com.lihaoyi:acyclic
's build to Mill, and from there we can flesh out the
missing features needed to port other builds: Scala.js support, Scala-Native
support, etc..
As the maintainer of many open-source libraries, all of the com.lihaoyi
libraries are fair game to be ported.
Once a fair number of libraries have been ported, Mill should be in good enough shape to release to the public, and we can try getting more people in the community-at-large on board trying out Mill. This should hopefully happen by the end of 2017.
Until then, let's keep Mill private. If someone wants to poke their nose in and see what's going on, we should expect them to contribute code!
The final milestone before a public release, is for Mill to be able to fully substitute SBT in developing the following projects:
-
https://github.com/lihaoyi/mill: a single-scala-version, single-platform Scala project
-
https://github.com/lihaoyi/acyclic: a Scala project cross-published against Scala {2.10, 2.11, 2.12}
-
https://github.com/playframework/play-json: a Scala project with multiple submodules partially-cross-published against Scala {2.10, 2.11, 2.12} X {JVM, JS}
-
https://github.com/lihaoyi/utest: a Scala project fully-cross-published against Scala {2.10, 2.11, 2.12} X {JVM, JS, Native}
-
https://github.com/gitbucket/gitbucket: a standalone application build using Scala (v.s. all the libraries listed above) with a simple build
Each of these are relatively simple projects. Satisfying all of their requirements (codegen, building, testing, publishing, etc.) and being happy with the Mill code necessary to do so (both in their build code, as well as Mill's own implementation). Meeting that baseline is the minimum we should support before advertising Mill to the general community