Author here! Hope you take a look at the project and find it cool. There's a lot of interesting stuff here. In particular, the Video linked on the landing page is a great intros from a Java developer point of view, and the following video is a great intro from a Build Tool Architect point of view:
For those of you who want to learn more about the design principles and architecture of Mill, and what makes it unique, you should check out the page on Design Principles which has links to videos and blog posts where I elaborate on what exactly makes Mill so different from Maven, Gradle, SBT, Bazel, and so on:
I've mentioned this in a few places, but the comparisons with other build tools are best-effort. I have no doubt they can be made more accurate, and welcome feedback so I can go back and refine them. Please take them with a grain of salt
I'm also trying to get the community involved, so it's not just me writing code and running the show. To that end, I have set up a bounty program, so pay out significant sums of money (500-2000USD a piece) for people who make non-trivial contributions. It's already paid out about 10kUSD and has another 20kUSD on the table, so if anyone wants to get involved and make a little cash, feel free to take a shot at one of the bounties! https://github.com/orgs/com-lihaoyi/discussions/6
How possible is it to make your tool "zero-config" by default? I see a lot of comments in this thread and elsewhere on twitter asking for essentially `go build`, `go fmt`, `go test` for Java/JVM. I think the language has quite strong convention around directory layout and file naming already, so do you think it would be possible for mill or a mill wrapper to offer the same kind of standardized zero config workflow? I think a JVM tool that gets that right - takes it as far as possible to the golang model - would have a lot of happy users.
Scala CLI has replaced the default runner since Scala 3.5, so you can effectively do `scala run`, `scala fmt`, and so on. On the Java side, I believe JBang provides a very similar developer experience.
Fundamentally it's hard to reconcile both worlds though. Building non-trivial multi-module projects on the JVM is inherently complex especially when you throw in multiple build targets, multiple toolchains, multiple platforms...
With simpler build tools (like in Go or Rust) you shift this complexity elsewhere, typically in a Makefile and/or a Docker/OCI based build pipeline, and these can get pretty complex too. Let alone distributed build tools like Bazel.
There's scala-cli, which has become the default Scala run command since Scala 3.5 but is also available separately. It has all those bells and whistles and allows scripts to grow organically. And no matter the name, it handles Java code too.
With scala-cli, there's not even a need to download a Java runtime or a language distribution. You can let the runner do its thing, or pass options to choose the JVM and the language version to use, or even write those options into special headers in the code files. You can also write tests, format code... it's all built-in. And in cases the code outgrows the tool and there's a need to migrate to a different build tool, there's even a feature to export the build to Sbt or Mill.
Not the author, but this is unfortunately a bit more difficult than it sounds. Like for example, where do you get the name of the jar file to build? I guess you could use the name of the root directory, but that may not be ideal.
How do you figure out dependencies? Import statements in .java files give you the packages to import, but those package names could be provided by one or more .jar files and, regardless, the package names need not bear any relation to the jar name or its group/artifact IDs (if pulling from e.g. a maven-style repository, which basically everyone does).
For multi-module projects, how do you figure out the dependencies between the modules, even? Sure, you could probably figure that out by parsing all the .java files in all modules and figuring out what they provide and import, but that would be slower than maven, probably.
You could certainly do this for small, dependency-free programs, but it would be such a niche use case that I don't think it would be worth the time.
Dependency resolution uses Coursier, which is one of the open source JVM dependency resolvers. SBT uses it too, and my last company used it with Bazel
The "ivy" thing is legacy haha. Mill used to use Apache Ivy to resolve dependencies, years ago. Coursier was a better/faster replacement, but names have a tendency to stick around
I’d be interested to read a comparison with Bazel (which you already mention as one of the influences).
For somebody looking to escape from Gradle, Bazel seems like one of the most promising alternatives, as it’s built on sane and sound fundamentals. Although in practice it has plenty of rough edges and annoyances, so maybe there are areas where Mill can do better.
Traditionally I've labelled my OSS projects 1.0 when they've stabilized and the rate of change has greatly reduced. Right now Mill is not there yet, but maybe if at end-2025 we realize no breaking changes have been needed since end-2024, we can call it 1.0
A build tool that is not only fast but configured in a type safe way sounds great. I really like this quote from the "Why use scala" part of the documentation:
> Most developers using a build tool are not build tool experts, and have no desire to become build tool experts. They will forever be cargo-culting examples they find online, copy-pasting from other parts of the codebase, or blindly fumbling their customizations. It is in this context that Mill’s static typing really shines: what such "perpetual beginners" need most is help understanding/navigating the build logic, and help checking their proposed changes for dumb mistakes. And there will be dumb mistakes, because most people are not and will never be build-tool experts or enthusiasts
I gave Mill a try earlier this year. My hope was to escape the nightmare that is Gradle, which I've been using for many years. Mill sounds great in theory (except for the Scala DSL). Unfortunately, I couldn't get a basic Java build to work in half a day, even though I have (admittedly rusty) working knowledge of Scala. It was one obscure error after another. My conclusion was that Java support isn't ready. There was also very little documentation on how to build Java.
In my opinion, using a GPL as the build language of a polyglot build tool is a dead end, both for technical/usability reasons and because the ensuing language wars can't be won. I'm looking forward to the day when a build tool embraces a modern config language such as CUE or Pkl.
Depending on when you tried, it could be worth trying again: support for Java has improved greatly in the last few months, as have the documentation. Come by our discord channel if you get stuck and i can help unblock you
I've had the same gripe about having to keep up with a second language just for the build tool for a while. Try taking a look at JeKa https://jeka.dev/
Mill looks interesting, but, _from a Java development perspective_, it has the same fundamental challenge as Gradle (and most other build systems), which is that its config language _is something other than Java_. That means there's a significant cognitive burden to understand and manage something that one hopes to not have to think about very often.
I find that the pain I experience with Gradle isn't usually about how to do something clever or customized etc, but instead it's when I haven't thought about Gradle syntax in the last 3 months since everything has been silently working, but now I need to figure out some small thing, and that means I need to go re-learn basic Gradle stuff - whether it's groovy, Kotlin, or some aspect of the build DSL - since my mind has unloaded everything about Gradle in the meantime.
Simplifying the semantic complexity of a general purpose build system will always help, but the most useful thing for me would be if the configuration for a Java build were to natively use the Java language directly.
It's intended as a replacement for _scala_ builds. Having a build definition in the native language that doesn't require a different syntax (like a declarative syntax such as maven xml or toml) makes task customization easier for the maintainer of a given project. Unfortunately, it also means that you have to know the language and read the documentation for the build system.
If you want something declarative, there's also bleep[1] in the scala ecosystem. And for single module builds there's scala-cli[2]. It's also possible to use gradle and maven for scala projects, but for an java-only shop I wouldn't be using mill or bleep because there's no need to introduce a new language just to manage the build. For scala/java/kotlin hybrid projects though, gradle or mill or sbt would be my recommended tool because of how tightly they are coupled with the cross-platform build matrix nature of scala library and build system plugin ecosystems. For larger builds, it's mill or bazel because there s a performance cliff in sbt and gradle, and bleep is too new to have all the standard plugins ported. We use mill at writer.
The intention has changed, Mill now explicity targets Java and Kotlin as well. It now has dedicated Java/Kotlin docsite
sections and examples, and has grown integrations with Palantir-Format, Checkstyle, Errorprone, Jacoco, and all their Kotlin equivalents (ktfmt, ktlint, kover).
Java and Scala (and Kotlin) are remarkably similar from a tooling perspective, so Mill tries to target both using the same shared infrastructure
> It's intended as a replacement for _scala_ builds. Having a build definition in the native language [...] makes task customization easier for the maintainer
Totally agree! But the title of the post says "Mill: A fast JVM build tool for Java and Scala" :) - it certainly looks like better tool for the Scala community.
For projects that are primarily building Java sources, it'd be nice to have a build system that uses Java code to describe the build. I don't think this exists at the moment.
Even in Java, because the language is relatively verbose, many frameworks fall back to an "inner platform" of magic annotations which have the same problem: e.g. just because you know how Java works doesnt mean your mind hasn't unloaded all the SpringBoot annotation semantics! But despite that it is worth it, because conciseness does matter.
Mill using Scala syntax is like that, but with the added advantage that even if you forget how Scala works, your IDE does not. You can really lean on Intellij or VScode to help you understand and navigate around a Mill build in a way that is beyond what is possible for most build tools: You can autocomplete things, peek at docs, navigate the build graph and module tree, etc. and learn what you need to learn without needing to reach for Google/ChatGPT. I use this ability heavily, and I hope others will enjoy these benefits as well
> Until you need to fix a 3 year old build that has some insane wizardry going on.
My experience with Gradle is that it's the "3 year old build" that is almost certainly a death knell more than the insane wizardry part. My experience:
git clone .../ancient-codebase.git
cd ancient-codebase
./gradlew # <-- oh, the wrapper, so it will download the version it wants, hazzah!
for _ in $(seq 1 infinity); do echo gradle vomit you have to sift through; done
echo 'BUILD FAILED' >&2
exit 1
Having worked with Maven and Gradle, I'd say Gradle was worse in the average case, but better in the worst case. There are way more Gradle projects with unnecessary custom build code because Gradle makes it easy to do.
On the other hand, when builds are specified in a limited-power build config language, like POM, then when someone needs to do something custom, they have to extend or modify the build tool itself, which in my experience causes way more pain than custom code in a build file. Custom logic in Maven means building and publishing an extension; it can't be local to the project. You may encounter projects that depend on extensions from long-lost open source projects, or long-lost internal projects. On one occasion, I was lucky to find a source jar for the extension in the Maven repository. It can be a nightmare.
The same could happen with Gradle, since a build can depend on arbitrary libraries, but I never saw it in the wild. People depended on major open-source extensions and added their own custom code inside the build.
My problem with gradle is that they keep making breaking changes for low value things like naming of options, so I have to chase deprecation warnings, and can never rely on a distro supplied gradle version
Gradle devs, please get over yourself and stay backward compatible.
Not to defend Gradle too much, but Groovy is a superset of Java. So if you want, you can just use the regular Groovy dialect and then write Java in your build scripts, it should work.
This is not entirely a solution though, because Gradle's APIs are fairly complicated and change regularly.
(Nitpick, but it’s just a “superficial” superset. The biggest difference is probably doing “multi-methods”, aka the runtime type of an argument deciding which method implementation to call vs java’s static overload resolution.)
The thing that’s great about maven is its declarative nature. You can declare goals and profiles for whatever you need the build system to do.
The main appeal that I can see from mill over maven is the power of dynamic programming over static xml files. Maybe good lsp/ide support will make managing a build system like this bearable?
Yes, IDE support in Mill is key. Without IntelliJ or VSCode, Mill would not be nearly as pleasant to use as it is today.
Mill and Maven both let you declare goals for what you want to do. One does it in XML and one does it in typechecked code. While XML does work, doing things in code with typechecking and full IDE support turns out to be pretty nice as well!
The comparision with Gradle is not up to date. There is stated that you would end up in an untyped mess of Groovy build files, but statically typed Kotlin files are the default for quite some time now in Gradle!
https://mill-build.org/mill/0.12.1/comparisons/gradle.html
Author here. Unfortunately this is because my own experience with Gradle is not up to date; I've only lived in the Gradle Groovy world! If anyone is interested in helping out, I have a 1500USD bounty on porting a gradle.kts build to Mill, so we can do a fair up-to-date comparison https://github.com/com-lihaoyi/mill/issues/3670
I never got why people thought Kotlin would help Gradle. It absolutely doesn't.
Groovy was never the problem (Groovy has types, always had, you could use them if you wanted).
Think about it: what do you do with a build tool? You write a little recipe, then you run it. Does that remind something? Yes, it reminds scripts, like bash scripts you run all the time in your terminal. And why are scripts almost universally written without types... and no typed alternative, of which there are many, has caught on? Because if you're just going to modify a script and immediately run it, while keeping it short enough so you can know what it does without reading a book, how does adding types help you? Quite to the contrary, scripts (including build scripts) should be small, and having types all over the place make it far more verbose than it should be, likely pushing it out of your comfortable local memory in your brain, at which point you need something akin to a "real" programming language and a compiled program, not a script. Larger programs benefit from types because you don't just run the program, make changes, and run them again, like you do with scripts. You write them, test them, compile them, package them and finally you distribute them to your users who hopefully only need to configure them, not modify their internals. If your build is that complex, that's exactly what you should be doing instead of trying to shoehorn types into your scripts and expecting them to look like real programs.
Also, the Kotlin DSL just doesn't assist in the most problematic aspect of Gradle: its total lack of discoverability. Try doing something on your Kotlin Gradle file using a plugin you're not familiar with (which is all of them for most of us). It's completely impossible unless you know the DSL of the plugin, just like it was the case with Groovy... Once you know the DSL, it's fairly easy, but even in Groovy you will get auto-completion once you've got to the DSL "entry point", no need for Kotlin. I've been saying this since before they introduced the Kotlin DSL, and now i feel completely vindicated. I've never met anyone who told me "Gradle is so much easier now with Kotlin". But it did mess up plugins I wrote in Kotlin as now Gradle has a dependency on a very particular version of the Kotlin compiler, and God help you if your plugin was written with a different version in mind.
I disagree with the static vs. dynamic typing part. Modern statically typed languages (like Kotlin, Scala, Rust etc.) are concise and readable. In the case of the Groovy DSL for Gradle it was sometimes hard to get code right or to find a bug. Even IntelliJ struggled at times with this mess of a DSL. So, in my opinion Kotlin is definitely an improvement here!
However, I agree with your second part, the DSL as such. The syntax is arbitrary in many cases and just not easy to remember or to make sense of. It looks like a DSL for the sake of a DSL. Take a look at this example (https://docs.gradle.org/current/userguide/plugins.html):
plugins {
application // by name
java // by name
id("java") // by id - recommended
id("org.jetbrains.kotlin.jvm") version "1.9.0" // by id - recommended
}
Why are there two ways to reference a plug-in? Why is the version written without parenthesis? Why is version an infix operator? Why not something as simple and consistent as this:
Yes the lack of discoverability, plus the unfamiliar syntax of Groovy, plus names changing between versions, I started with Gradle thinking it would be easier but in the end I'd love to go back to Ant. That was awful to write but at least you could understand it.
What tends to be complex about build requirements that necessitates special purpose tools? Golang seems to be doing fine with just go build and go test. What else are people doing with gradle/maven that requires static typing, DAGs, plugins etc.?
`go build` and `go test` do work, at limited scale and complexity. In Scala there's Scala-CLI which is excellent. If they work for you, you probably aren't the target market for these build tools. Once you start layering on bash scripts, layering on make, layering on Python scripts, layering on manual steps written down in a readme.md somewhere, that's the time when you should consider a proper build tool. And if that has never happened to you in your career, count your blessings :)
Why not just write some boring, pure code to handle the build? Why not write my build system in vanilla LYAH Haskell? It turns out that builds do have some specific requirements that most programs do not need to care about: caching, parallelization, introspection, and so on. Check out the following blog/talk for more details:
Thus "naively" building your project "directly with code" ends up not working, so you do need some additional support. While most build tools end up constructing a complete bespoke programming environment from scratch, Mill tries to leverage the Scala language and JVM as much as possible, so you can re-use all your expertise and tooling (e.g. IntelliJ, VSCode, Maven Central, etc.) almost verbatim while getting all the necessary build-tool stuff (parallelism, caching, introspection) for free. Check out those two links if you want to learn more!
Thanks, the post actually answers my question on build requirements. It's a very good write up overall!
If your USP is to solve those layering cases, then why not target other ecosystems like golang/rust as well? Your design philosophy certainly seems to be language agnostic. By calling yourself a build tool for Java and Scala, it gives an impression that this is solving problems specific to those environments, and your adoption also indicates as such. Is it that these communities do not like to adopt such tools or is there something about the JVM ecosystem that tends towards having complex build requirements?
The issue is that most people in jvm land are on a closed bubble and haven't seen anything else.
This is true for build systems as is for non OO design for example.
Most simply don't know better and the rest of us are simply stuck.
Ant and then Maven started simple enough but people always find a way to justify adding more stuff.
Gradle already started complex enough and they keep adding more stuff...
The big thing is that many Java builds are not just blobs of binaries crammed together, but also have structure and metadata, sometimes generated on the fly.
Not all Java builds are simply compiles. There are several projects that rely on processing steps during the Java build. EAR files are jars within jars.
Then, of course, there's all of the dependencies.
The modern Maven based repository based dependency manager is a blessing and a curse. Drag and dropping an artifact into your project that inevitably downloads the entirety of the internet. Now you may wish to cull your dependency tree, so that needs to be expressible as well.
The primary benefit of Maven and the pom.xml file is that for a vast majority of applications it just work. Even better, its become a universal "project" format that many IDEs directly support. It well handles "dependency hell" in a cross tool way.
I wish Maven were a bit faster, but, simply, it's as fast as it can be for what it does. A good Ant build just flies, but Ant "doesn't do anything". It's just a bag of steps that it follows (for good and ill), in contrast to Mavens declarative style (for good and ill).
I have no experience with Gradle other than I've never run into enough problems with Maven to justify trying something else. On its surface, it doesn't really appeal to me. I was comfortable with Ant (I have no problem with XML), I'm mostly comfortable with Maven. I've not been unhappy enough with Maven to try and jump back to Ant w/Ivy.
To be blunt, if anything it might be you who live in a closed bubble.
Builds that require ad-hoc functionality is the default. It’s extremely rare that everything fits nicely into “cargo build” or other single language build tools’ model. And while these often have escape hatches, at that point you have to write imperative code with no caching and parallelization that is literally the job of a build tool.
You have to keep in mind that Gradle Inc. earns money by providing consulting for complex builds. An easy build tool would destroy this business model ;-)
In JVM world, the de factor equivalent to `go build` and `go test` are `mvn compile` and `mvn test`, which works 99% percent of the time.
Other build tools and plugins just compete/fill in for:
* improved build speed / test speed: using background daemon to reduce strtup speed, intelligent caching / task reordering to avoid redoing, etc..
* extra functionalities like code generation, publishing or deployments. As code generation is really big in JVM world, and there are many ways to deploy an application: jar + libs in a zip file, uber jars, container image, etc...
Curmudgeon here: this was true for a relatively brief period of time. Nowadays I'd say that gradle has (inexplicably to me) taken the lead - and everyone adds custom crap to their gradle build making them far less predictable than maven builds used to be.
I guess it's better than the nightmare over in the front-enders' world...
My typical mvn session after a month of not touching maven:
% mvn
[ERROR] No goals have been specified for this build.
Uh.
% mvn build
[ERROR] Unknown lifecycle phase "build".
Uh, nope, that wasn't it....
% mvn compile
BUILD SUCCESS
Now trying to run the app... Error: app jar not found.
Reading the README.md. Aha! So I need to install it!
% mvn install
16:59:35,075 [INFO] Building <blahblahblah> [1/32]
...
16:59:38,483 [INFO] -------------------------------------------------------
16:59:38,483 [INFO] T E S T S
16:59:38,483 [INFO] -------------------------------------------------------
^C
// me searching on google how to not run tests
% mvn install -DskipTests=true
// now good...
Ok, let's run some tests. The FlakyTest broke again in CI. Let me run it locally:
% mvn test FlakyTest
17:02:19,481 [ERROR] Unknown lifecycle phase "FlakyTest".
Aaargh, ok, googling it again:
% mvn test -Dtest=FlakyTest
17:03:16,102 [ERROR] Failed to execute goal org.apache.maven.plugins:maven-surefire-plugin:3.3.0:test (default-test) on project blah blah: No tests matching pattern "FlakyTest" were executed! (Set -Dsurefire.failIfNoSpecifiedTests=false to ignore this error.)
The original idea behind maven is nice.
But the defaults are so bad, that the whole experience of convention over configuration has been ruined.
Cargo and go build systems took the original maven philosophy and implemented them right, with good UX.
Gradle, SBT and friends took a step back to the times before maven, and went fully the Ant way, doubling down on "configuration over convention". Where "configuration" is actually "programming in a DSL on top of another language on top of Java".
Try to build a Go project that uses Cgo and non-trivial C/C++ libraries. Throw in cross-compilation for more fun. You'll end up with an external build system that invokes Go build as one of the steps.
Go projects normally just tend to be self-contained server-like software that doesn't need a lot of external libraries. But once you step away from that, you're on your own.
I guess my problem with Gradle is that app building should be way simpler than it is. Apps are not something niche anymore, but the tooling is still similar to the embedded software for microcontrollers.
I'm working on a project that encompasses both JVM (Gradle, Kotlin) and Golang.
My hot take: JVM build tools, especially Gradle, are a soup of unnecessary complexity, and people working in that ecosystem have Stockholm Syndrome.
In Golang, I spend about 99% of my time dealing with code.
In JVM land, I'm spending 30% just dealing with the build system. It's actually insane, and the community at large thinks this is normal. The amount of time it takes to publish a multi-platform Kotlin library for the first time can be measured in days. I published my first Golang library in minutes, by comparison.
You speak from my soul! I'm in the Java world for a really long time now and I'm wondering for years why the build tools need to be so complicated an annoying. I know Go, Node.js and bit of Rust and all have more pleasant easier to use build tools! The JVM (or GraalVM) as an ecosystem is just fine and probably one of the best, but build tools might be achille's heel. Maybe it would be a good idea for Oracle to invest into that area ...
I'm in JVM land. I spend very little time dealing with the build system. It is actually insane how well it works.
Also, why does it matter how long it takes to publish a library for the first time? It sounds like a non-issue to me. I have written dozens of libraries and published them to a local artifactory instance because it simply doesn't matter if your company specific code is accessible to the world or not.
One note from having worked with both that I don’t see mentioned: Golang dependencies are sources you basically pull and compile with your own code. In JVM-land dependencies are precompliled packages (jars). This adds one little step.
> The amount of time it takes to publish a multi-platform Kotlin library for the first time can be measured in days. I published my first Golang library in minutes, by comparison.
It's a bit Apple & Orange comparison: publishing a JVM only Kotlin library is quite easy, it's the multiplatform part that takes time.
I've been working on Java-based systems for about 20 years now, and I fully relate to that. Same experience.
This is so annoying that I prefer to use Rust over Java even in areas where things like better performance or better type system don't matter. But being able to start a fresh project with one `cargo init` and a few `cargo add` invocations to add any dependencies... well, this is priceless.
Interesting. I spend nearly zero time with my maven setup and almost all the time is in coding. I am genuinely curious to know where that 30% time goes? Is it waiting for builds?
There are quite a few cases: the moment you touch another language, resources that require a compile-step (e.g. xml schemas to code dtos, protonbuf, all that kind of stuff), sometimes even the code itself requires generation.
Build tools sit in an unhappy corner of the design space where they provide features not found in the core of regular programming languages, and which are so generally useful that there's a temptation to make them very abstract, but then they often lack some of the features that let regular programs scale well.
The key feature that justifies their existence is parallel and incremental execution of DAGs of world-mutating tasks. This is an awkward fit with most programming languages, hence the prevalence of DSLs. But people don't want their build system to become a general purpose programming language, because they don't want to think about build systems at all and because programmers don't buy programming languages anymore, so, this causes a big design tension between generality (people want to use build systems to automate many things) and deliberately limiting the expressive power to try and constrain the design space and thus tooling investment required.
Java is in an awkward place because the JDK was born in the 90s on UNIX, by people who thought make is a sufficiently good solution. You still see remnants of this belief in the official Java tutorials, in JEPs, and in the fact that OpenJDK itself is compiled using an autotools based build system! (fortunately it's one of the nice make based build systems out there).
The problem with make is twofold:
1. It assumes a CLI that's both powerful and standardized provided by the host OS. Windows violates this assumption but Java is meant to be portable to Windows.
2. "Plugins" are CLI tools or scripts and so make implicitly assumes that subprocess creation is cheap. But process creation on Windows is expensive, and starting up JVM programs is also expensive due to the JIT compiling.
Therefore make just doesn't work well in the JVM ecosystem. At the same time, the Java project wasn't providing any competing solution, so the wider open source community was left to fill in the gaps. These days language developers provide build tooling out of the box as part of the base toolset along with the compiler, but Java still doesn't.
So - you ask, what are people doing with Gradle/Maven that requires all those features. The answer is: everything! Gradle builds frequently orchestrate dozens of different tools as part of a build pipeline, build documentation websites, do upload and deployment, download and manage dependencies, run security scanners and license compliance checkers, analyze dependency graphs, modify compiler behaviors, and so on.
Additionally Gradle isn't specific to Java, or even JVM apps. It can also be used to compile C/C++ programs, run native code compilers like Kotlin/Native, and it abstracts the underlying platform so Gradle builds aren't tied to UNIX.
Readit News
* https://www.youtube.com/watch?v=UsXgCeU-ovI
While Mill is focusing on JVM for now, it is very extensible and I have a strawman demo of adding a Javascript toolchain in ~100 lines of code https://mill-build.org/mill/0.12.1/extending/new-language.ht...
For those of you who want to learn more about the design principles and architecture of Mill, and what makes it unique, you should check out the page on Design Principles which has links to videos and blog posts where I elaborate on what exactly makes Mill so different from Maven, Gradle, SBT, Bazel, and so on:
* Mill Design Principles https://mill-build.org/mill/0.12.1/depth/design-principles.h...
I've mentioned this in a few places, but the comparisons with other build tools are best-effort. I have no doubt they can be made more accurate, and welcome feedback so I can go back and refine them. Please take them with a grain of salt
I'm also trying to get the community involved, so it's not just me writing code and running the show. To that end, I have set up a bounty program, so pay out significant sums of money (500-2000USD a piece) for people who make non-trivial contributions. It's already paid out about 10kUSD and has another 20kUSD on the table, so if anyone wants to get involved and make a little cash, feel free to take a shot at one of the bounties! https://github.com/orgs/com-lihaoyi/discussions/6
Fundamentally it's hard to reconcile both worlds though. Building non-trivial multi-module projects on the JVM is inherently complex especially when you throw in multiple build targets, multiple toolchains, multiple platforms...
With simpler build tools (like in Go or Rust) you shift this complexity elsewhere, typically in a Makefile and/or a Docker/OCI based build pipeline, and these can get pretty complex too. Let alone distributed build tools like Bazel.
- https://scala-cli.virtuslab.org
- https://www.jbang.dev
With scala-cli, there's not even a need to download a Java runtime or a language distribution. You can let the runner do its thing, or pass options to choose the JVM and the language version to use, or even write those options into special headers in the code files. You can also write tests, format code... it's all built-in. And in cases the code outgrows the tool and there's a need to migrate to a different build tool, there's even a feature to export the build to Sbt or Mill.
How do you figure out dependencies? Import statements in .java files give you the packages to import, but those package names could be provided by one or more .jar files and, regardless, the package names need not bear any relation to the jar name or its group/artifact IDs (if pulling from e.g. a maven-style repository, which basically everyone does).
For multi-module projects, how do you figure out the dependencies between the modules, even? Sure, you could probably figure that out by parsing all the .java files in all modules and figuring out what they provide and import, but that would be slower than maven, probably.
You could certainly do this for small, dependency-free programs, but it would be such a niche use case that I don't think it would be worth the time.
Also, what’s with the “ivy” on https://mill-build.org/mill/0.12.1/comparisons/maven.html ? Any relation to Apache Ivy?
The "ivy" thing is legacy haha. Mill used to use Apache Ivy to resolve dependencies, years ago. Coursier was a better/faster replacement, but names have a tendency to stick around
Good luck for your project!
For somebody looking to escape from Gradle, Bazel seems like one of the most promising alternatives, as it’s built on sane and sound fundamentals. Although in practice it has plenty of rough edges and annoyances, so maybe there are areas where Mill can do better.
What's required for v1.0?
He has written multiple useful libraries. Out of many JSON libraries, his one was the most intuitive and practial.
His book is excellent too. I bought it when it came out. It is worthy of a plug: https://www.handsonscala.com/
I miss working on Scala projects. Sadly I rarely see new ones these days.
Does IntelliJ plugin finally work on Scala 3? About 2 years ago it was half broken.
> Most developers using a build tool are not build tool experts, and have no desire to become build tool experts. They will forever be cargo-culting examples they find online, copy-pasting from other parts of the codebase, or blindly fumbling their customizations. It is in this context that Mill’s static typing really shines: what such "perpetual beginners" need most is help understanding/navigating the build logic, and help checking their proposed changes for dumb mistakes. And there will be dumb mistakes, because most people are not and will never be build-tool experts or enthusiasts
In my opinion, using a GPL as the build language of a polyglot build tool is a dead end, both for technical/usability reasons and because the ensuing language wars can't be won. I'm looking forward to the day when a build tool embraces a modern config language such as CUE or Pkl.
I find that the pain I experience with Gradle isn't usually about how to do something clever or customized etc, but instead it's when I haven't thought about Gradle syntax in the last 3 months since everything has been silently working, but now I need to figure out some small thing, and that means I need to go re-learn basic Gradle stuff - whether it's groovy, Kotlin, or some aspect of the build DSL - since my mind has unloaded everything about Gradle in the meantime.
Simplifying the semantic complexity of a general purpose build system will always help, but the most useful thing for me would be if the configuration for a Java build were to natively use the Java language directly.
If you want something declarative, there's also bleep[1] in the scala ecosystem. And for single module builds there's scala-cli[2]. It's also possible to use gradle and maven for scala projects, but for an java-only shop I wouldn't be using mill or bleep because there's no need to introduce a new language just to manage the build. For scala/java/kotlin hybrid projects though, gradle or mill or sbt would be my recommended tool because of how tightly they are coupled with the cross-platform build matrix nature of scala library and build system plugin ecosystems. For larger builds, it's mill or bazel because there s a performance cliff in sbt and gradle, and bleep is too new to have all the standard plugins ported. We use mill at writer.
1. https://bleep.build/docs/
2. https://scala-cli.virtuslab.org/
Java and Scala (and Kotlin) are remarkably similar from a tooling perspective, so Mill tries to target both using the same shared infrastructure
Totally agree! But the title of the post says "Mill: A fast JVM build tool for Java and Scala" :) - it certainly looks like better tool for the Scala community.
For projects that are primarily building Java sources, it'd be nice to have a build system that uses Java code to describe the build. I don't think this exists at the moment.
Mill using Scala syntax is like that, but with the added advantage that even if you forget how Scala works, your IDE does not. You can really lean on Intellij or VScode to help you understand and navigate around a Mill build in a way that is beyond what is possible for most build tools: You can autocomplete things, peek at docs, navigate the build graph and module tree, etc. and learn what you need to learn without needing to reach for Google/ChatGPT. I use this ability heavily, and I hope others will enjoy these benefits as well
Sounds amazing in practice. And it is. Until you need to fix a 3 year old build that has some insane wizardry going on.
My experience with Gradle is that it's the "3 year old build" that is almost certainly a death knell more than the insane wizardry part. My experience:
Contrast that with https://github.com/apache/maven-app-engine (just to pick on something sorted by earliest push date, some 10 years ago):On the other hand, when builds are specified in a limited-power build config language, like POM, then when someone needs to do something custom, they have to extend or modify the build tool itself, which in my experience causes way more pain than custom code in a build file. Custom logic in Maven means building and publishing an extension; it can't be local to the project. You may encounter projects that depend on extensions from long-lost open source projects, or long-lost internal projects. On one occasion, I was lucky to find a source jar for the extension in the Maven repository. It can be a nightmare.
The same could happen with Gradle, since a build can depend on arbitrary libraries, but I never saw it in the wild. People depended on major open-source extensions and added their own custom code inside the build.
Gradle devs, please get over yourself and stay backward compatible.
This is not entirely a solution though, because Gradle's APIs are fairly complicated and change regularly.
The main appeal that I can see from mill over maven is the power of dynamic programming over static xml files. Maybe good lsp/ide support will make managing a build system like this bearable?
Mill and Maven both let you declare goals for what you want to do. One does it in XML and one does it in typechecked code. While XML does work, doing things in code with typechecking and full IDE support turns out to be pretty nice as well!
Or, I believe you can submit a PR to linguist to make it globally registered: https://github.com/github-linguist/linguist/blob/v8.0.1/CONT...
Groovy was never the problem (Groovy has types, always had, you could use them if you wanted).
Think about it: what do you do with a build tool? You write a little recipe, then you run it. Does that remind something? Yes, it reminds scripts, like bash scripts you run all the time in your terminal. And why are scripts almost universally written without types... and no typed alternative, of which there are many, has caught on? Because if you're just going to modify a script and immediately run it, while keeping it short enough so you can know what it does without reading a book, how does adding types help you? Quite to the contrary, scripts (including build scripts) should be small, and having types all over the place make it far more verbose than it should be, likely pushing it out of your comfortable local memory in your brain, at which point you need something akin to a "real" programming language and a compiled program, not a script. Larger programs benefit from types because you don't just run the program, make changes, and run them again, like you do with scripts. You write them, test them, compile them, package them and finally you distribute them to your users who hopefully only need to configure them, not modify their internals. If your build is that complex, that's exactly what you should be doing instead of trying to shoehorn types into your scripts and expecting them to look like real programs.
Also, the Kotlin DSL just doesn't assist in the most problematic aspect of Gradle: its total lack of discoverability. Try doing something on your Kotlin Gradle file using a plugin you're not familiar with (which is all of them for most of us). It's completely impossible unless you know the DSL of the plugin, just like it was the case with Groovy... Once you know the DSL, it's fairly easy, but even in Groovy you will get auto-completion once you've got to the DSL "entry point", no need for Kotlin. I've been saying this since before they introduced the Kotlin DSL, and now i feel completely vindicated. I've never met anyone who told me "Gradle is so much easier now with Kotlin". But it did mess up plugins I wrote in Kotlin as now Gradle has a dependency on a very particular version of the Kotlin compiler, and God help you if your plugin was written with a different version in mind.
However, I agree with your second part, the DSL as such. The syntax is arbitrary in many cases and just not easy to remember or to make sense of. It looks like a DSL for the sake of a DSL. Take a look at this example (https://docs.gradle.org/current/userguide/plugins.html):
Why are there two ways to reference a plug-in? Why is the version written without parenthesis? Why is version an infix operator? Why not something as simple and consistent as this: How does the DSL help here? Is it more readable? Easier to lear or to remember?Just look at Guice how nice a DSL can look like with pure Java:
I'd really whish Java had build tools with better developer experience. I whish Mill the best luck!`go build` and `go test` do work, at limited scale and complexity. In Scala there's Scala-CLI which is excellent. If they work for you, you probably aren't the target market for these build tools. Once you start layering on bash scripts, layering on make, layering on Python scripts, layering on manual steps written down in a readme.md somewhere, that's the time when you should consider a proper build tool. And if that has never happened to you in your career, count your blessings :)
Why not just write some boring, pure code to handle the build? Why not write my build system in vanilla LYAH Haskell? It turns out that builds do have some specific requirements that most programs do not need to care about: caching, parallelization, introspection, and so on. Check out the following blog/talk for more details:
* Blog Post: Build Tools as Pure Functional Programs https://www.lihaoyi.com/post/BuildToolsasPureFunctionalProgr...
* Video: Mill: a Build Tool based on Pure Functional Programming https://www.youtube.com/watch?v=j6uThGxx-18&list=PLBqWQH1Miw...
Thus "naively" building your project "directly with code" ends up not working, so you do need some additional support. While most build tools end up constructing a complete bespoke programming environment from scratch, Mill tries to leverage the Scala language and JVM as much as possible, so you can re-use all your expertise and tooling (e.g. IntelliJ, VSCode, Maven Central, etc.) almost verbatim while getting all the necessary build-tool stuff (parallelism, caching, introspection) for free. Check out those two links if you want to learn more!
If your USP is to solve those layering cases, then why not target other ecosystems like golang/rust as well? Your design philosophy certainly seems to be language agnostic. By calling yourself a build tool for Java and Scala, it gives an impression that this is solving problems specific to those environments, and your adoption also indicates as such. Is it that these communities do not like to adopt such tools or is there something about the JVM ecosystem that tends towards having complex build requirements?
The issue is that most people in jvm land are on a closed bubble and haven't seen anything else. This is true for build systems as is for non OO design for example. Most simply don't know better and the rest of us are simply stuck.
Ant and then Maven started simple enough but people always find a way to justify adding more stuff. Gradle already started complex enough and they keep adding more stuff...
Not all Java builds are simply compiles. There are several projects that rely on processing steps during the Java build. EAR files are jars within jars.
Then, of course, there's all of the dependencies.
The modern Maven based repository based dependency manager is a blessing and a curse. Drag and dropping an artifact into your project that inevitably downloads the entirety of the internet. Now you may wish to cull your dependency tree, so that needs to be expressible as well.
The primary benefit of Maven and the pom.xml file is that for a vast majority of applications it just work. Even better, its become a universal "project" format that many IDEs directly support. It well handles "dependency hell" in a cross tool way.
I wish Maven were a bit faster, but, simply, it's as fast as it can be for what it does. A good Ant build just flies, but Ant "doesn't do anything". It's just a bag of steps that it follows (for good and ill), in contrast to Mavens declarative style (for good and ill).
I have no experience with Gradle other than I've never run into enough problems with Maven to justify trying something else. On its surface, it doesn't really appeal to me. I was comfortable with Ant (I have no problem with XML), I'm mostly comfortable with Maven. I've not been unhappy enough with Maven to try and jump back to Ant w/Ivy.
Builds that require ad-hoc functionality is the default. It’s extremely rare that everything fits nicely into “cargo build” or other single language build tools’ model. And while these often have escape hatches, at that point you have to write imperative code with no caching and parallelization that is literally the job of a build tool.
Other build tools and plugins just compete/fill in for:
* improved build speed / test speed: using background daemon to reduce strtup speed, intelligent caching / task reordering to avoid redoing, etc..
* extra functionalities like code generation, publishing or deployments. As code generation is really big in JVM world, and there are many ways to deploy an application: jar + libs in a zip file, uber jars, container image, etc...
I guess it's better than the nightmare over in the front-enders' world...
Cargo and go build systems took the original maven philosophy and implemented them right, with good UX.
Gradle, SBT and friends took a step back to the times before maven, and went fully the Ant way, doubling down on "configuration over convention". Where "configuration" is actually "programming in a DSL on top of another language on top of Java".
Go projects normally just tend to be self-contained server-like software that doesn't need a lot of external libraries. But once you step away from that, you're on your own.
I guess my problem with Gradle is that app building should be way simpler than it is. Apps are not something niche anymore, but the tooling is still similar to the embedded software for microcontrollers.
My hot take: JVM build tools, especially Gradle, are a soup of unnecessary complexity, and people working in that ecosystem have Stockholm Syndrome.
In Golang, I spend about 99% of my time dealing with code.
In JVM land, I'm spending 30% just dealing with the build system. It's actually insane, and the community at large thinks this is normal. The amount of time it takes to publish a multi-platform Kotlin library for the first time can be measured in days. I published my first Golang library in minutes, by comparison.
Also, why does it matter how long it takes to publish a library for the first time? It sounds like a non-issue to me. I have written dozens of libraries and published them to a local artifactory instance because it simply doesn't matter if your company specific code is accessible to the world or not.
It's a bit Apple & Orange comparison: publishing a JVM only Kotlin library is quite easy, it's the multiplatform part that takes time.
This is so annoying that I prefer to use Rust over Java even in areas where things like better performance or better type system don't matter. But being able to start a fresh project with one `cargo init` and a few `cargo add` invocations to add any dependencies... well, this is priceless.
Half are ignorant. Other half are like me and just stuck with no options.
But the tooling ecosystem on the JVM truly is horrific.
For what platform(s)?
Or did you really just push the source code?
The key feature that justifies their existence is parallel and incremental execution of DAGs of world-mutating tasks. This is an awkward fit with most programming languages, hence the prevalence of DSLs. But people don't want their build system to become a general purpose programming language, because they don't want to think about build systems at all and because programmers don't buy programming languages anymore, so, this causes a big design tension between generality (people want to use build systems to automate many things) and deliberately limiting the expressive power to try and constrain the design space and thus tooling investment required.
Java is in an awkward place because the JDK was born in the 90s on UNIX, by people who thought make is a sufficiently good solution. You still see remnants of this belief in the official Java tutorials, in JEPs, and in the fact that OpenJDK itself is compiled using an autotools based build system! (fortunately it's one of the nice make based build systems out there).
The problem with make is twofold:
1. It assumes a CLI that's both powerful and standardized provided by the host OS. Windows violates this assumption but Java is meant to be portable to Windows.
2. "Plugins" are CLI tools or scripts and so make implicitly assumes that subprocess creation is cheap. But process creation on Windows is expensive, and starting up JVM programs is also expensive due to the JIT compiling.
Therefore make just doesn't work well in the JVM ecosystem. At the same time, the Java project wasn't providing any competing solution, so the wider open source community was left to fill in the gaps. These days language developers provide build tooling out of the box as part of the base toolset along with the compiler, but Java still doesn't.
So - you ask, what are people doing with Gradle/Maven that requires all those features. The answer is: everything! Gradle builds frequently orchestrate dozens of different tools as part of a build pipeline, build documentation websites, do upload and deployment, download and manage dependencies, run security scanners and license compliance checkers, analyze dependency graphs, modify compiler behaviors, and so on.
Additionally Gradle isn't specific to Java, or even JVM apps. It can also be used to compile C/C++ programs, run native code compilers like Kotlin/Native, and it abstracts the underlying platform so Gradle builds aren't tied to UNIX.
That's why it's so complicated.
Configuration and workflow execution.