dev meeting 2025 11 12

Agenda

• the morning meeting (@shonfeder)
• pending release (@shonfeder)

Meeting notes

Attendees: @Alizter, @rgrinberg, @shonfeder, @ElectreAAS, @Leonidas-from-XIV, @art-w

• the morning meeting
  • Should we keep having it?
• pending release (@shonfeder)
  • Get parametric lib instantiation merged
  • Get the read only promoted files change fixed or reverted
  • Shon will plan to start cutting next week
• paramteric libs (@rgrinberg)
  • needs to be scope sensitive
  • important to not
  • needs to preserve caching
  • Arthur: for inline tests, need to know for a file whether it has been preprocessed
    • b/c filenames get the pp infix
    • modules give filename, but the filename may not be correct, in case the files where preprocessed
    • so we cannot reliably look up a filename from a module without knowing if it has been preprocessed
    • Any more elegant way to solve this?
      • Rudi: maybe directories? different naming schemes? -- is just a pain to deal with now
• Portable lock dirs (@rgrinberg)
  • Portable lock dirs as feature should disappear as flag in code base
  • Instead just refer to list of platforms
  • Only outstanding question is what configurations do we solve for if user has not provided any information
  • What is the point of lock dir portability?
    • Ali: that people do not have to think about the locking at all
    • Rudi: not having to think about it may just not be an option
    • Shon: The important property is just that you can build the same locked project on different machines, even if you need to relock
    • Ali: What if a maintainer doesn't want to incorporate lock info from a new platform of some user?
      • Agreed we should have a way way to lock that appends the new platform solution, and a way to lock that only extends the platform lock locally
  • Rudi: has talked with Ryan Gibb, who has implemented an alternative solver for opam
    • based on pubgrub: https://github.com/pubgrub-rs/pubgrub (there's an ocaml implementation)
    • talks about how to solve problem of not being able to specify constraints such as a single version of package for multiple configurations
    • used in uv
    • if it is good, maybe we could swap this out for opam
  • Arthur: we could already have this behavior in the current solver if we refactor to build a single SAT problem
  • Shon: Renewing question of whether having a requirement that a single version is the right default
    • Concern about over-constraint contributing to dependency hell
    • But team confirmed that lock files are not factored into dependencies
• perf of null builds with and without dune pkg (@shonfede)
  • Could this only be an issue with portable lock dirs?
  • Ali:
    • there is a conditional in parsing
    • A second (minro) perf regression between released versions (dune to modular hashing)
    • what is responsible for the with/without?
      • thought maybe it could be digests? -- but Rudi said we should not be digesting on null builds
  • parsing, digesting, stating
  • last resort is make some assumptions about only dune writing into the _build and lock dirs
• eager watch mode not handling RPCs (@alizter)
  • #12715
  • yep, it's broken
  • Rudi: one of first things I did when joining internal dune team at JS was rewrite the RPC system
    • One of things this does is remove distinction between eager/lazy
    • The JS RPC system would not be a drop-in replacement, but could help
    • However, it won't solve all the problems anyhow
  • What about always having a build server?
    • E.g., in bazel, first launch always starts a build server if not running
    • Potential complications with opam installs