dev meeting 2025 07 09

Agenda

• Merging strategy (@maiste)
• CI broken on main (@Leonidas-from-XIV)
• Signal to use to enable dune pkg (@Leonidas-from-XIV)
• Caching ocaml-toolchain (@Alizter)
• Lock dir representation (@Leonidas-from-XIV)

Meeting notes

Attendees: @Alizter, @maiste, @Leonidas-from-XIV, @OlivierNicole @shonfeder @ElectreAAS @aguluman @rgrinberg @Sudha247

Merging strategy (@maiste)

• The history is hard to read because of merging from main into branches and merging the branches back to main
• Rebasing takes a while as we need to wait for the result before merging
  • This is @Leonidas-from-XIV workflow, nobody enjoys this
• Merge queue was tested and didn't work well
• Automerging?
  • @Leonidas-from-XIV tested it and it merged immediately, so it didn't work well
  • @shonfeder points out that the tests are not configured as required, thus it didn't wait for them to complete
• We should try automerging in the future

CI failures (@Leonidas-from-XIV)

• CI was failing on main a lot lately
• Often times this is due to Melange or Jsoo, so please make sure they don't break the build
• It makes contributing hard because PRs branched of from broken main inherit the brokenness and need to be rebased after main is fixed, creating unnecessary friction for contributors
• Some test lock up forever
  • Rudi was able to reproduce one locally and disabled it
• Ali wants to implement a timeout for cram tests, so they fail on tests that get stuck

Signal to enable Dune package management (@Leonidas-from-XIV)

• If we autolock into _build without promotion into the source tree we can't use the presence of a lock dir in a project as a signal to enable package management
• New flag in dune-workspace
  • That requires everyone who wants to use it create or edit their dune-workspace
• Additionally a flag in dune config, so it can be enabled globally for all user's projects
• It will still allow installation via OPAM without package management as -p disables lock directories.

Caching the toolchain (@Alizter)

• OCaml toolchain is getting invalidated when you delete _build
  • One option is to not do that
  • However, what about CI setups?
    • Why not share _build as well?
    • Not great if the same CI setup is building multiple projects that should all reference the same toolchain compiler
• Make the toolchain rule put the result in the shared cache instead of recompiling it
• The compiler can't be put in the shared cache itself because it is not relocatable and restoring it from the cache leads to a non-working compiler
• If we put the cookie in the cache it is possible that the compiler it points to gets deleted, thus it would break the cache
• We'd need to store it in the cache and make sure that what it references still exists
• How do we know a file in the cache is a cookie?
  • The cache is content-addressable so we can't ask for the file name
  • The cookies need to register their existence somewhere from where they can be deleted
• We could write a command that will clean up toolchains and fix up broken cookies
• When is the relocatable compiler coming?
  • @rgrinberg: The feature is basically ready and according to @dra27 going to be backported
  • @OlivierNicole: it's scheduled for OCaml 5.5 by the end of the year, no idea about the backporting to older compiler

Autolocking API (@Leonidas-from-XIV)

• Currently it is all Path.Source.t but if we write/load in _build, should it be Path.t or Path.Local.t?
• @rgrinberg At best it would be Path.Build.t since that's where they will live
• It the lock dir is in the source directory, it can be copied to _build anyway, like any other source file for building so we can just read it from there
• Maybe we should make people store just inputs to the solver
• This avoids issues with portable lock files as the solver will find a configuration for the current machine
• @maiste: What should be our principal, recommended workflow for dune pkg?
  • We don't really know yet
• We could in theory support he opam lockfile format
  • There's some issues with that as the lockfile format still has the filters on and our lock format resolves filters but it is pretty close

Native cflags (@OlivierNicole)

• https://github.com/ocaml/dune/issues/11923
• Blocks Semgrep
• @rgrinberg does not have this on his roadmap, unsure about the Tarides roadmap for this
• @voodoos worked on building the stubs concurrently
• @shonfeder: What is needed to solve this?
  • @OlivierNicole: Dune should do the right thing by default, because currently Dune users need to configure it manually to work properly, but it can be worked around
• @rgrinberg: It's definitely a bug and Dune should do the right thing by default especially as @voodoos has implemented the necessary support already to make it work properly