Skip to main content

Dependency inference: Precise caching and concurrency, without the boilerplate

· 7 min read
Stu Hood

Scalable build tools have historically meant a significant boilerplate burden.

But it doesn't have to be that way! Pants v2 supports all of the caching, concurrency, and introspection you need to scale your repository, with significantly less boilerplate, thanks to… Dependency inference!

As discussed in our post announcing Pants v2, it's clear that Python has "grown up" by gaining facilities to help it to scale to larger projects. But as codebases grow and tool counts increase, more Python codebases need build tools. While you could write bespoke scripts to coordinate each of your tools, using Pants brings benefits like caching, concurrency, introspection, a simple and uniform user experience, and more!

Unfortunately, scalable build tools have historically meant a significant boilerplate burden: scattering BUILD files throughout your repository and then needing to edit both your code and the redundant dependency information in build definitions.

But it doesn't have to be that way! Pants v2 supports the precise caching, concurrency, and introspection that you need to scale your repository, with up to 90% less BUILD boilerplate, thanks to… Dependency inference!

Scaling up

Pants supports repos of all sizes with minimal boilerplate, but it is particularly helpful in repositories containing multiple deployable or publishable projects, each with potentially different requirements or interpreters: aka, monorepos.

Monorepos have lots of benefits (no dependency conflicts, atomic cross-project commits, easy top-to-bottom continuous integration, linear change history), but essential to making them scale are the abilities to:

  1. test, check, and deploy precisely the portion of the repository that is relevant to you
  2. cache builds and tests to avoid re-building when unrelated code has changed
  3. manage and configure the variety of tooling that users of the repository will want to use

It's critical in a monorepo to be able to test, check, and deploy exactly the relevant portion of your code, ideally with zero impact from changes in unrelated parts of the codebase. For example: if you have three libraries A, B, and C, which depend on one another in a chain like A -> B -> C, a monorepo that builds from source using Pants allows you to completely ignore versioning (and files, per-project requirements.txt, etc) while you edit library C, even if you are running the tests for library A. If you have ten other projects (or one thousand!) in your repository and only a few of them depend on C, you'd like to avoid ever running tests or mypy for the unrelated libraries while editing C.

And in those cases when you do want to take advantage of a monorepo's top-to-bottom integration testing by running "all of your dependent's tests" (or maybe just typecheck them with Mypy!), you'd like to do that as quickly as possible by taking advantage of caching, concurrency, or transparently executing them on a cluster of machines using remote execution.

To enable this scalability, monorepo build tools like Pants v1 and Bazel required that the dependencies between libraries and files were declared in BUILD files. These dependencies were then used to determine which portions of the repository needed to be built, and which files needed to be included in cache keys.

"But wait", you say! "Doesn't that mean we've traded editing and requirements.txt files for every library for editing BUILD files for every library?" In most tools, that would be the case: but not in Pants v2! Pants v2 is different.

Dependency inference

Pants 2.0.0 almost entirely eliminates the boilerplate of declaring dependencies between libraries thanks to "Dependency inference". Dependency inference is roughly what it sounds like: Pants supports discovering the dependencies between Python libraries by parsing import statements (and you can use our powerful plugin API to infer other dependencies, such as by parsing a Django settings file, YAML config, etc).

Rather than adding an import statement to your code resulting in your tests failing because you forgot to also update a BUILD file, Pants will use your newly added import statement to infer that that file now has a dependency within the repository (or outside it via your requirements.txt). When designing inference, we strove to remove boilerplate without introducing magic, so BUILD files are still used to declare any metadata your library might have (the version of the Python interpreter to use, etc), and can be used to override or extend the inferred dependencies.

Even better, Pants infers these dependencies at the file level. Rather than adding a dependency from "the library named A" to "the library named B" (or "target" in monorepo parlance), Pants tracks that "file depends on file "". Rather than staring at the content of your BUILD files to (attempt to?) understand your dependencies, you can use Pants' dependency introspection tools to easily explore them at the file level: ./pants dependencies $file.

In practice, we've found that inferred file-level dependencies can reduce the total per-file dependency count by an average of 30% (improving cache hit rates), and reduce the size of BUILD files by up to 90% (reducing boilerplate)! See our docs for a real world example.

And critically (as we'll discuss in further posts!), dependency inference is both 1) very safe, and 2) very fast. Because Pants invokes processes hermetically using SHA256 fingerprinting and strong sandboxes (your test frameworks, your linters, mypy, everything), failing to infer a dependency can never cause the wrong things to be cached. And because the core of Pants is implemented in Rust — and uses a daemon, parallelism, and very-fine-grained memoization — inference won't slow you down!


To show what it's like to use dependency inference, we'll quickly add a feature with assistance from some new first and third party dependencies (from sources and PyPI, respectively).

We'll start with a broken test that expects TOML files to be supported by a library in a different directory:

Because we're curious, we'll start by asking Pants whether the library already (directly) depends on TOML:

Nope. Only YAML. But let's add the import statement and see what happens…

Voila! This file now declares a dependency on TOML via the import statement, without any modifications to BUILD files. But we're not finished: neither the flake8 linter nor the customer will be satisfied with an unused import! Let's finish adding the feature, and then re-run the test.

Ok, TOML is clearly being used: but this time our test needs tweaking. To fix it we can import a TomlSerializer helper class from a second library (another new dependency!). We already know that we don't need to check the BUILD file to see whether this test declares a dependency on the library, so we can just focus on our code!:

Great: our test is now passing! To fix it, we edited two files and introduced two new dependencies -- but we didn't need to edit any BUILD files, despite these files living in different targets!

Finally, let's confirm the critical monorepo scalability property that edits to unrelated files don't invalidate the caching of our new test. Although this test is quick, there are a few hundred files in this repository, and plenty of other tests that could take long enough to result in coffee breaks!:

Excellent! Thanks to dependency inference's file level precision, our test is still quickly and correctly cached, even after editing neighboring files in the target! Productivity preserved; mission accomplished.


The precise, always-accurate dependencies in Pants have a lot of benefits:

  1. teams can quickly get started using Pants
  2. BUILD files (or requirements.txt/ files) can't go out of sync with the code, because you don't need to repeat all of your import statements there
  3. the cache keys for processes (and thus the number of cache hits and amount of rebuilding) are more accurate and much more fine-grained than you would ever write by hand

If you're interested in speeding-up and scaling-up your builds without the boilerplate, the Pants community would love to help!