Semi-automated decomposition of large compilation units (jars only for now).

Usage: jdeps -v path/to/app.jar | bazel run hairball [options] [command] [command options]
  Options:
    -h, --help

      Default: false
  Commands:
    ProposeClusters      attempt to break input jar into n acyclic clusters.
            writes a .dot graph of the proposed clusters to stdout.
      Usage: ProposeClusters [options]
        Options:
          --assign-labels
            pass custom assign_labels arg to sklearn.spectral_clustering. See https://scikit-learn.org/stable/modules/generated/sklearn.cluster.spectral_clustering.html.
            Default: kmeans
            Possible Values: [kmeans, discretize, cluster_qr]
        * --n-clusters
            target number of clusters to create. the actual number of clusters
            may be smaller; see README.md for discussion.

    MinimalClusters      break input jar into minimal acyclic clusters
            (=strongly connected components). writes a .dot graph to stdout.
      Usage: MinimalClusters

Large compilation units are a common headache in software maintenance. They cause many problems, including long builds, large artifacts, and difficulty sharing code with other software.

The traditional way to shrink a large compilation unit is to manually move some of its sources into other compilation units. This requires knowledge of the application's structure and good judgment. In the case of monolithic compilation units that have grown over time, this manual process often runs into circular dependencies: it can feel like every source file imports every other source file, making it impossible to extract a small subset of sources.

hairball partially automates this process. When you point jdeps at a jar file and pipe its output to hairball, it computes a way to partition the classes in the jar into clusters, such that there are no circular dependencies between clusters. The output is a .dot file describing a graph structure. You can then use graphviz (not included) to visualize the proposed structure (e.g. dot -Tsvg -o app.svg app.dot).

The examples directory contains SVGs of hairball runs on some prominent open-source projects.

hairball is built with bazel. Once you have bazel installed, clone the repo and build the hairball binary:

$ git clone https://github.com/Ubehebe/hairball
$ cd hairball
$ bazel build hairball

This will automatically download and configure the toolchains and dependencies (java, kotlin, python) needed to run hairball, without disturbing versions that may already be installed on your system.

hairball doesn't inspect jars or source code directly. Instead, it expects to read the output of jdeps -v on standard input. jdeps has many options to restrict the scope of analysis, so you can invoke jdeps how you like and pipe it to hairball. For example:

$ jdeps -v --regex 'some\.package\.path.*' path/to/app.jar |
    bazel run hairball -- ProposeClusters --n-clusters 100 |
    dot -Tsvg -o app.svg /dev/stdin

For meaningful results, the input jar should represent a single compilation unit (gradle module, bazel target, etc.). jars that include their transitive dependencies (shadow jars, shaded jars, deploy jars) already represent a graph of compilation units; analyzing them as though they represent a single compilation unit is pointless.

Once you're satisfied with the clustering proposed by hairball, it's up to you to implement it by making the appropriate edits to your project's source code.

hairball first condenses the source graph into strongly connected components. This is the most granular possible way to decompose a compilation unit: into sets of sources, each member of which imports all the others.

An extremely fine-grained dependency graph is rarely desirable, because each compilation unit requires some boilerplate to set up (build.gradle.kts, etc.); you don't want one of these files for each and every source file. Usually, what you want is a "reasonable" number of compilation units: maybe 20 or 50, depending on the size of the application. To achieve this, hairball uses spectral clustering to partition the connected components into a user-specified number of clusters (--n-clusters). Because the clustering may introduce new circular dependencies, hairball re-condenses the clusters into strongly connected components before writing the final graph, to remove the circular dependencies. This means that the number of clusters in the output may be less than --n-clusters.

• hairball doesn't try to break circular dependencies. Large compilation units often contain a subset of sources that all transitively import each other. hairball will warn about them, like so:

  [main] INFO hairball.Main - condensed 298 classes into 245 strongly connected components. the
  largest component contains 48 classes that transitively depend on one another. you will have to
  break these dependencies manually.
  

  Circular dependencies are the fundamental reason why shrinking a compilation unit is hard. hairball helps by extracting the dependencies that can be extracted, letting you focus on the manual work of breaking circular dependencies.

• hairball outputs a proposed dependency structure; it doesn't rearrange your source tree to implement that structure. (hairball has no access to source code.) For the time being, you have to manually translate hairball's output into edits to your build.gradle.kts files, BUILD files, etc., depending on your build system.

• hairball only works on the output of jdeps -v at the moment, though its underlying algorithms should work for any kind of dependency structure. Feel free to send a pull request to teach hairball about other languages or build systems.

• hairball doesn't understand what the different parts of your application do. It's common to decompose applications into compilation units based on function: all the database code in one unit, all the HTTP code in another unit, etc. hairball often clusters semantically related code together based purely on their imports, but it is not guaranteed to.

• hairball purposely ignores all kinds of structure inside the jar other than class->class dependencies, including packages and java modules. This allows hairball to propose a structure different from (and possibly better than) those structures.