February 23, 2021

deps.edn and monorepos

At World Singles Networks llc we have been using a monorepo for several years and it has taken us several iterations to settle on a structure that works well with the Clojure CLI and deps.edn.

Updated April 21st, 2021 to reflect recent changes in our setup. See deps.edn and monorepos II for more details.

The Monorepo/Polylith Series

This blog post is part of an ongoing series following our experiences with our Clojure monorepo and our migration to Polylith:

  1. deps.edn and monorepos (this post)
  2. deps.edn and monorepos II
  3. deps.edn and monorepos III (Polylith)
  4. deps.edn and monorepos IV
  5. deps.edn and monorepos V (Polylith)
  6. deps.edn and monorepos VI (Polylith)
  7. deps.edn and monorepos VII (Polylith)
  8. deps.edn and monorepos VIII (Polylith)
  9. deps.edn and monorepos IX (Polylith)
  10. deps.edn and monorepos X (Polylith)
  11. deps.edn and monorepos XI (Polylith)

What does our monorepo look like?

Our main git repo has a build folder containing scripts, tooling, and configuration, and a clojure folder containing all our Clojure source and test code.

That clojure folder has over three dozen subprojects that represent either reusable "libraries" or "applications". We build just over a dozen application artifacts from this codebase for deployment (as uberjars) to production.

We have about 111,000 lines of Clojure: about 88,000 is source code and the rest, 23,000, is test code.

If you don't want to read all the back story, you can jump straight to the TL;DR and see the solution we've settled on.

Why do we use a monorepo?

Even though we can (and sometimes do) build and deploy an application artifact on its own, we tend to want to build and deploy all of our application artifacts together, tied to a single git tag and/or SHA.

Although each of our subprojects has different dependencies, we generally want to control the specific versions of many third-party library dependencies across the whole codebase so that we don't have to worry about a transitive dependency in one "library" (subproject) conflicting with a different version of the same dependency in another "library". We find value in being able to update a version in one place and have it apply to all of the subprojects. [2021-04-21: while this has been convenient, we've decided that the benefits do not outweigh the inconveniences caused by our use of the :defaults alias -- see below.]

When working in a REPL, we like to have all of the source code and all of the test code (and all of its dependencies) available in our editor, and can make changes across multiple subprojects if we wish, as well as being able to run any combination of tests from our editor.

An additional goal is to avoid, as much as possible, duplication of configuration across subprojects -- in particular, in the context of this article, to avoid duplication of dependencies and aliases across deps.edn files in these subprojects. [2021-04-21: as with pinning versions above, we have decided to accept some duplication as a trade-off for simple use of tooling -- see below.]

How does the Clojure CLI and deps.edn help us?

When we started building this codebase a decade ago, we used Leiningen because that was the only game in town. For a long time, it worked well, but it started to feel constraining as we wanted to automate more and more of our dev/test/build processes. I've talked about our switch from Leiningen to Boot before, about five years ago.

Once we switched to Boot, we decided to put our dependencies in EDN files so that we had more control over them and could manipulate them programmatically. When the Clojure CLI first appeared (2017), we saw similarities between the deps.edn approach and our own handling of dependencies and I started work on a Boot task to read the new deps.edn format with a view to providing a migration path from Boot to the Clojure CLI. I compared Leiningen, Boot, and the Clojure CLI about three years ago. We migrated completely to the Clojure CLI and deps.edn some time in 2018.

We like the simplicity and performance of the Clojure CLI: it computes a cache of dependencies and command line options only as needed and so it mostly runs just a single JVM (unlike Leiningen) with just your project's code and dependencies.

We can build any tooling we want for our dev/test/build pipeline, using regular Clojure code (much like Boot's approach, only without even the small "framework" of Boot itself). This also allows us to mix'n'match tooling as we need, much more smoothly than with either Leiningen or Boot.

We like that the Clojure CLI is official, supported tooling from Cognitect (Nubank), and gets regular updates that are carefully considered in a holistic manner alongside Clojure itself.

We like that you can choose to perform project-based CLI operations in a "reproducible" way that excludes user-level configuration in ~/.clojure/deps.edn but that you can also perform operations with the full power of your personal customizations if you want, giving our developers freedom to set up their environments however they wish.

We like that dependencies can be local -- via :local/root -- so that our subprojects can very easily depend upon each other at a source code level, mirroring how we work with all the source code in our editor already, across the whole monorepo.

We particularly like :override-deps so that we can "pin" versions of dependencies across all of our subprojects, by including a single alias when invoking the CLI. [2021-04-21: we did like it but it doesn't play all that well with a lot of CLI tooling.]

How did the Clojure CLI and deps.edn hinder us?

The CLI assumes there are three deps.edn files: the root one baked into the CLI installation (actually, into the version of org.clojure/tools.deps.alpha that underpins the CLI), the user-level one (usually in ~/.clojure/), and the project-level one.

Overriding the user-level deps.edn file

Our initial approach was to leverage the CLJ_CONFIG environment variable that the CLI supports to select a different directory for the user-level deps.edn file. This allowed us to have a single deps.edn containing all of the "control" aliases we wanted, for pinning dependencies via :override-deps, for testing tools, for building JAR files, etc, and then for each subproject to have its own project-level deps.edn file. This worked well:

$ cd monorepo/subproject
$ CLJ_CONFIG=../versions clojure -M:defaults:other:aliases -m some.tooling

We wrapped this in a build shell script so we could run the following and it would handle cd and adding the other bits of the command-line:

$ build other:aliases subproject -m some.tooling

We also had a pseudo-project called everything and a small Clojure script that merged all the subproject deps.edn files into a single deps.edn file in the everything subproject that we used as the basis for our REPL with "all code" available. [Since almost all dependency versions were specified via :override-deps in versions/deps.edn, this (generated) everything/deps.edn file only changed occasionally]

But it had the downside that developers could not leverage their user-level deps.edn because CLJ_CONFIG was overriding that with our ../versions/deps.edn. That led to that file slowly accruing an amalgam of any and all tooling that each team member wanted, and it became unwanted incidental complexity in terms of maintenance.

Generating project-level deps.edn files

Eventually, we broke down and decided to figure out a way to restore access to the user-level deps.edn file while still maintaining our "control" file. After talking to a number of other Clojurians who were also using monorepos, it seemed a common option was to programmatically generate the project-level deps.edn file, as needed, from a repo-wide template (essentially our versions/deps.edn file) and a template in each project. Since it's "just data" in EDN files, this is trivial to do in Clojure -- and we were already doing a little of this for our everything/deps.edn file. We added some code to our build script to compute hashes for the template EDN files and to automatically generate subproject deps.edn files "on demand" and for a while that worked well: we simplified versions/deps.edn to remove all the per-developer cruft and we were happy that we could customize our developer experience as much as we wanted, outside of the company repo!

This also made it easier to run the clojure command without our build script since the CLJ_CONFIG=../versions prefix was no longer required (and sometimes we'd been running it manually outside the build script even when we had to provide that environment variable override).

However, deps.edn already has a subtle "gotcha" around local dependencies (:local/root), and that is regarding what happens if a (transitive) dependency changes: if you're working inside the subproject where that dependency changes, the CLI will see the change and regenerate the cache and everything will "just work". If you're working in a different subproject, that depends locally on the one that changed, that change can't be detected "at a distance" and you need to remember -Sforce (or just blow away the .cpcache directory the CLI maintains).

Once we started generating subproject deps.edn files "on demand", it amplified that problem because using clojure -Sforce was no longer enough to pick up transitive changes and our "on demand" code wasn't always smart enough to figure it out either so running build ... -Sforce wasn't quite reliable enough.

The final straw for us was a recent change to the CLI in version 1.10.2.790 where a "warning [is issued] if :paths or :extra-paths refers to a directory outside the project root (in the future will become an error)" (emphasis mine). Our everything project depended on the other subproject's source code via :local/root but since that didn't include the tests, we used :extra-paths to provide all of the (relative) subproject paths to the tests, e.g., "../subproject/test", which falls foul of this new warning. The warning is sensible and I can understand why the Clojure team want to disable "random external paths" in projects -- but it meant we needed to rethink our "everything" setup.

Not just us

It's probably worth mentioning at this point that the Polylith team are also looking at supporting the Clojure CLI / deps.edn and their architecture is a type of monorepo -- and they are also struggling with effective ways to organize deps.edn files and how to invoke them. Monorepos come in many forms! [2021-04-21: I've spent quite a bit of time looking at Polylith since I wrote this post and that experience has influenced some of the changes we have made since then.]  

The Third Way

At this point, I pressed Alex Miller pretty hard on how he might tackle this problem if he were forced -- gun to his head -- to work on a monorepo like ours?

After a bit of back and forth with Alex in a thread on Slack and several DMs later, the path we agreed that I should explore was to create a top-level deps.edn file -- a variant of our former versions/deps.edn file -- that had an alias for every subproject that contained a :local/root dependency on the subproject itself.

Since we can't activate aliases on local dependencies' deps.edn files, I also added a :*-test alias for every subproject into the top-level deps.edn file, "lifting" the testing dependencies and paths up one level.

Finally, I added an :everything alias that had :extra-deps containing every subproject and :extra-paths for all of the subprojects' test code.

All clojure commands are now run from that top-level directory, with our :defaults alias (bringing in :override-deps), an alias for the subproject(s) you want to operate on, maybe an alias for the tests for those subprojects, and then any tooling alias(es) and arguments. We still have our build shell script to make this a little less verbose (its "API" hasn't changed at all but it uses the subproject name as an alias now instead of changing into that subdirectory). [2021-04-21: we've abandoned the :defaults alias and the use of :override-deps and accepted a small amount of duplication of dependency version declarations, in exchange for simpler tooling interactions.]

In summary, here's the structure of our monorepo now:

<repo>
|____build # our shell scripts / config / etc
...
|____clojure # our Clojure code "root"
| |____activator # a subproject
| | |____deps.edn # bare dependencies: no versions, no test
. . .    [2021-04-21: src dependencies with versions, no test]
| | |____src
| | | |____ws
| | | | |____activator.clj
| | |____test
| | | |____ws
| | | | |____activator_expectations.clj
...
| |____classes
| | |____.keep
...
| |____deps.edn # control deps.edn file
...
| |____worldsingles # another subproject
| | |____deps.edn
| | |____resources
...
| | |____src
...
| | |____test
...

Then clojure/activator/deps.edn has:

{:deps
 {worldsingles/worldsingles {:local/root "../worldsingles"}
  ;; originally:
  camel-snake-kebab/camel-snake-kebab {}
  com.stuartsierra/component {}
  seancorfield/next.jdbc {}
  ;; 2021-04-21:
  camel-snake-kebab/camel-snake-kebab {:mvn/version "0.4.2"}
  com.stuartsierra/component {:mvn/version "1.0.0"}
  seancorfield/next.jdbc {:mvn/version "1.1.646"}
  }}

And in clojure/deps.edn we have:

  ;; for each subproject, we have two aliases:
  :activator {:extra-deps {worldsingles/activator {:local/root "activator"}}}
  :activator-test
  {:extra-paths ["activator/test"]
   :extra-deps {worldsingles/worldsingles-test {:local/root "worldsingles-test"}}}

That clojure/deps.edn also has tooling aliases, such as:

  :test ; for a testing context
  {:extra-deps {com.gfredericks/test.chuck {:mvn/version "0.2.10"}
                expectations/clojure-test {:mvn/version "1.2.1"}
                org.clojure/test.check {}}
   :jvm-opts ["-Dclojure.core.async.go-checking=true"
              "-Dclojure.tools.logging.factory=clojure.tools.logging.impl/log4j2-factory"
              "-Dlogged-future=synchronous"
              "-XX:-OmitStackTraceInFastThrow"
              "--illegal-access=warn"]}

  :runner ; to run tests (test:<subproject>-test:runner) -- the <task> called test
  {:extra-deps {org.clojure/tools.namespace {}
                org.clojure/tools.reader {}
                com.cognitect/test-runner
                {:git/url "https://github.com/cognitect-labs/test-runner"
                 :sha "b6b3193fcc42659d7e46ecd1884a228993441182"}}
   :jvm-opts ["-Dlog4j2.configurationFile=log4j2-silent.properties"]
   :main-opts ["-m" "cognitect.test-runner"
               "-r" ".*[-\\.](expectations|test)(\\..*)?$"]}

So we would run:

$ clojure -M:defaults:activator:activator-test:test:runner -d activator/test
# or just
$ build test activator

All of the {} versions are specified in the :defaults alias via :override-deps[2021-04-21: this has gone away -- versions are specified directly in subprojects' deps.edn files now.]:

  ;; "pinned" versions for all cross-project dependencies
  :defaults
  {:override-deps
   {...
    camel-snake-kebab/camel-snake-kebab {:mvn/version "0.4.2"}
    clj-time/clj-time {:mvn/version "0.15.2"}
    clojure.java-time/clojure.java-time {:mvn/version "0.3.2"}
    ...
    org.clojure/tools.logging {:mvn/version "1.1.0"}
    org.clojure/tools.namespace {:mvn/version "1.0.0"}
    org.clojure/tools.reader {:mvn/version "1.3.3"}
    ...}}

When we build an uberjar:

$ clojure -X:uberjar :aliases '[:defaults :activator]' :jar '"test.jar"' :main-class ws.activator :aot true

Where the :uberjar alias is:

  :uberjar
  {:replace-deps {seancorfield/depstar {:mvn/version "2.0.187"}}
   :exec-fn hf.depstar/uberjar}

And that :everything alias?

  :everything
  {:extra-deps {worldsingles/activator {:local/root "activator"}
                ...
                worldsingles/worldsingles {:local/root "worldsingles"}
                ...}
                worldsingles/wsseogeo {:local/root "wsseogeo"}}
   :extra-paths ["activator/test"
                 ...
                 "worldsingles/test"
                 "worldsingles-test/test"
                 ...]}

Starting a REPL:

$ clj -A:defaults:everything:test
Clojure 1.10.3-rc1
user=> (require 'ws.activator)
nil
user=>

[2021-04-21: see deps.edn and monorepos II for more details about the changes made since this post was originally written.]

Got questions?

Find me on the Clojurians Slack, or just ask in the comments below.

Tags: clojure monorepo