dev meeting 2025 05 28

Notes

Attendees: @gridbugs, Chukwuma Akunyili, @Leonidas-from-XIV, @shym

Portable Lockfiles (@gridbugs)

• Dune got lockfiles/directories about a year ago
• Only guaranteed to be working on the current system
• Recently new experimental mechanism to make them portable by solving for different platforms at once
• Dune can show the depexts
• But only for the machine that you created the solution for
• Version 1 could add the depexts for the specific platform the solution was locked for and record it in the solution
• However depexts are mostly the same for any distribution, only the name of the package might differ between them (pkg-config, pkgconf etc)
• Solution in review: Depexts are taken from the OPAM files and just copied into the lockfile and evaluated at build time instead
• Problem: Usually OPAM variables like os-distribution are fixed for build commands at lock time, but not for depext commands now, which is inconsistent
• Should build commands also be allowed to be conditional on variables?
  • This would make lock files more compact as the build command does not need to be recorded for each solved platform separately
  • @rgrinberg wasn't a fan of this, @gridbugs will dig up the previous discussion
  • Maybe worth revisiting it again, now that depexts are conditional and lockfiles are meant to be portable
• Can we drop the distribution from the dependencies?
  • Depending on one OPAM package on one distribution and another one on another distribution seems very odd and possibly nobody does that
  • That would solve the issue that there's tons of Linux distributions, thus tons of solutions to be calculated (or fewer and some platforms not supported well)
  • The compiler depends on distribtion at least on Windows, to determine whether its being installed on cygwin
• Should we remove ubuntu and/or alpine from the presets?
  • Maybe, if we have mac, generic linux and some Windows distributions that could be enough?
• Does the architecture really matter for solving?
  • The compiler depends on dummy packages that contain the architecture
  • But these can also be built on other architectures, so they could be locked
• OPAM filters don't have wildcard values, undefined variables will lead to short-circuit behavior where expressions are undefined and on the top level they will be treated as false
  • We could implement proper wildcards in our solver, that would evaluate * and not * and * || <expr> correctly
  • We should however stay compatible with OPAM semantics
  • Maybe just dummy values are fine, these will always evaluate to false