Add prefix to extension

[devmotion]

Extensions with non-unique name cause collisions and hence break e.g. compilation of sysimages. This PR adds a prefix to the extension to work around that issue. See JuliaMath/ChangesOfVariables.jl#13 for a longer discussion.

[jw3126]

Thanks a lot! Can you also prefix the other extensions?

[aplavin]

Unfortunate that it's needed for now. I guess it's temporary? See JuliaLang/Pkg.jl#3380 (comment).
Results in long and redundant names - eg in precompilation output:

  ✓ LogExpFunctions → LogExpFunctionsInverseFunctionsExt
  ✓ LogExpFunctions → LogExpFunctionsChainRulesCoreExt
  ✓ LogDensityProblemsAD → LogDensityProblemsADEnzymeExt
  ✓ LogDensityProblemsAD → LogDensityProblemsADForwardDiffBenchmarkToolsExt
  ✓ LogDensityProblemsAD → LogDensityProblemsADForwardDiffExt
  ✓ LogDensityProblemsAD → LogDensityProblemsADReverseDiffExt
  ✓ LogDensityProblemsAD → LogDensityProblemsADTrackerExt
  ✓ LogDensityProblemsAD → LogDensityProblemsADZygoteExt

and in Project.toml.
Some official scheme of extension naming would be nice.

[devmotion]

Well, I assume this is desired even if the name collision issues are fixed, see e.g. JuliaMath/ChangesOfVariables.jl#13 (comment).

[devmotion]

Some official scheme of extension naming would be nice.

I guess it's already established inofficially, most packages that I'm aware of use these prefixes (e.g., SciML, SpecialFunctions, and LogExpFunctions).

[aplavin]

Oh, there are LOTS of packages that don't do prefixing - just search github as https://github.com/search?q=%5Bweakdeps%5D+lang%3ATOML&type=code.

[devmotion]

Github doesn't show any results, is the link you posted correct? In any case, I'm sure there are packages which don't use prefixes but a major part of the ecosystem, including core packages, already does. And I guess it's rather these packages which form inofficial standards.

[aplavin]

Github doesn't show any results, is the link you posted correct?

Just clicked again - seems to work. It's search for [weakdeps] across all TOML-type files.

I'm not against this PR anyway.
It's just unlikely to work at scale, duplicate names will inevitably appear.
And this prefixing leads to redundancy in messages, that's not helpful.

[aplavin]

a major part of the ecosystem, including core packages, already does

I think it's mostly SciML and related, btw. Eg LoopVectorization, Plots, MPI, PGFPlotsX, PkgCacheInspector, StaticArrayInterface don't use prefixes - and these include core julia developers. This is just from the first search page :)

[devmotion]

Just clicked again - seems to work. It's search for [weakdeps] across all TOML-type files.

That's what I assumed but it doesn't work for me:

And this prefixing leads to redundancy in messages, that's not helpful.

I wonder if that might be (partly) a complaint about the standard Pkg messages rather than the names of the extensions. I imagine that if there is an error in one of the extension modules then standard logging/error messages might only point you to ChainRulesCoreExt, and it might not be immediately clear to which parent package the extension belongs.

Generally, I'm not surprised that quite a few packages don't use prefixes yet. Many packages switched to prefixes only very recently (e.g., SpecialFunctions a few days ago), now that it became apparent that there are benefits.

[aplavin]

that might be (partly) a complaint about the standard Pkg messages rather than the names of the extensions

If anything, it's natural to show both the parent package name and extension name in any messages - be it Pkg, errors, or whatever.

[jw3126]

LGTM. We should add the Accessors prefix for now, if it is not needed in future anymore, we can always drop it again. Thanks @devmotion