Expanding Contracting Monorepos

As some point with a monorepo approach to source control (especially with binary dependencies in the source tree), your checkouts could be bigger than your local workstation’s hard drive. Or even if the checkout is not too big for your hard drive, then it might be too much for your IDE, and you do not want to have to abandon it for Vim/Emacs. Or maybe it is not IDE that chokes is is something about the build that’s too much locally, despite command-line arguments to attempt to pare it down for a shorter elapsed time.

There is a way to intelligently expand or contract the checkout on you developer workstation, to alleviate all of the above.

Gcheckout.sh

Google’s in-house DevOps uses some simple scripting to modify the checkout on the developer’s workstation to omit the source files/packages that are not needed for the current intentions of the developer. This Blaze related technology is a shell command called ‘gcheckout’. It can modify the mappings between the multi-gigabyte HEAD revision of company-wide trunk (monorepo) and developer’s own workstation. Thus the source-control tools maintain the smallest possible subset of the monorepo on the developer’s workstation, for them to perform their daily work. Google and the industry refer to the general feature as ‘sparse checkout’.

You can run gcheckout at any time to modify your sparse checkout to be bigger or smaller (or wholly different) for different reasons. All of those are operations on your local representation of a larger trunk.

Contrived example of use

We detailed two intentions for directed graph build systems above, using a contrived application. Here is one more:

  • I now want to change TheORMweBothDependOn, because a change to MyTeamsApplication requires me to do that.

In Google, rather than feed into the backlog of the team that maintains TheORMweBothDependOn (which may exist as a part-time committee rather than a team), the developer in question would make the change themselves. Perhaps they had made it in the same commit as the first usage of it for MyTeamsApplication. In the code review cycle (Google practice common code ownership), the approvers for the TheORMweBothDependOn would see all the changes together. The larger change is all accepted or rejected (to be remediated) atomically.

So our developer was working on MyTeamsApplication, which depended on TheORMweBothDependOn (which probably transitively depended on other things). Now that developer is going to change TheORMweBothDependOn and that impacts TheirApplication too. The Blaze related checkout-modifying technology ‘gcheckout’ performs an expansion to bring in TheirApplication to the developer’s checkout. From that moment on, the developer doing update/pull/sync will bring down minute by minute changes to those three modules. For free, the build expands to make sure that the TheORMweBothDependOn changes do not break either of MyTeamsApplication or TheirApplication.

Directory structure and working copy.

If I ran gcheckout with MyTeamsApplication+TheirApplication as the parameter I would get working copy that looked like:

root/
  java/
    BUILD
    com/
      BUILD
      google/
        BUILD
        myteamsapplication/
          BUILD
          MyTeamsApplication.java  // and hundreds of other packages and source files
        theirapplication/
          BUILD
          TheirApplication.java  // and hundreds of other packages and source files
        theormwebothdependon/
          BUILD
          TheORMweBothDependOn.java  // and hundreds of other packages and source files
  java_test/
    BUILD
    com/
      BUILD
      google/
        BUILD
        myteamsapplication/
          BUILD
          MyTeamsApplicationTest.java  // and hundreds of other packages and source files
        theirapplication/
          BUILD
          TheirApplicationTest.java  // and hundreds of other packages and source files
        theormwebothdependon/
          BUILD
          TheORMweBothDependOnTest.java  // and hundreds of other packages and source files

If I ran gcheckout with MyTeamsApplication as the parameter I would get working copy that looked like:

root/
  java/
    BUILD
    com/
      BUILD
      google/
        BUILD
        myteamsapplication/
          BUILD
          MyTeamsApplication.java  // and hundreds of other packages and source files
        theormwebothdependon/
          BUILD
          TheORMweBothDependOn.java  // and hundreds of other packages and source files
  java_test/
    BUILD
    com/
      BUILD
      google/
        BUILD
        myteamsapplication/
          BUILD
          MyTeamsApplicationTest.java  // and hundreds of other packages and source files
        theormwebothdependon/
          BUILD
          TheORMweBothDependOnTest.java  // and hundreds of other packages and source files

If I ran gcheckout with TheORMweBothDependOn as the parameter I would get working copy that looked like:

root/
  java/
    BUILD
    com/
      BUILD
      google/
        BUILD
        theormwebothdependon/
          BUILD
          TheORMweBothDependOn.java  // and hundreds of other packages and source files
  java_test/
    BUILD
    com/
      BUILD
      google/
        BUILD
        theormwebothdependon/
          BUILD
          TheORMweBothDependOnTest.java  // and hundreds of other packages and source files

You can keep rerunning the gcheckout to expand or contract your working copy to meet your current goals.

Contrived example of use #2

We used ‘change the wheel on a car’, on the Branch By Abstraction page for its contrived example. It will serve us again here. The wheel is what we want to change. The other team using ‘Wheel(s)’ is making a Segway thing (two wheels and self-balancing via high-torque and very responsive motors). Here’s the procedure:

The starting position is two teams working separately, using ‘Wheel’ (4 for cars, 2 for Segways). Without any commits happening the engineer changing ‘Wheel’ for everyone, runs gcheckout (or its equivalent) to modify the source in the IDE to the union of Car and Segway (and in-house dependencies). That is marked as step 0. Let us say the change is quick/easy this time (not requiring Branch By Abstraction) step 1 shows the single commit that changes the wheel implementation for everyone. After the commit/push, running again shows the application focused team checkout - either ‘Car’ or ‘Segway’.

Git’s Sparse checkouts

Git has a ‘sparse checkout’ capability, which exactly facilitates this sort of thing. Subversion and Mercurial do too.
Perforce has a ‘client spec’ capability that is more or less the same. A team wanting to have their own gcheckout equivalent would have some scripting around sparse checkouts (or equivalent).

Using Git this way today

If you’re willing to go a ‘split history’ maneuver on your monorepo once or twice a year, Git can do the expandable and contractible monorepo setup today.

Perforce’s client-specs

Perforce has a ‘client spec’ (alternatively ‘view’) that is accessed via the client command or UI. Amongst other things, it allows a checkout to be a subset of the directories/files available within the branch. A list of globbed includes and excludes is the format. You would script this (as Google did until 2012) to have a directed graph driven expandable/contractible checkout.

PlasticSCM’s cloadked.conf

As Perforce, but via ‘cloaked.conf’ file.

Subversion’s sparse-checkouts

Subversion has a ‘sparse checkout’ capability. You do a series of checkout operations at various directory levels in order to create the mapping, so is less atomic or centrally configured than the others. They have made also made a ‘viewspec’ script to allow the reshaping of the working copy to happen in a more declarative way.