@peterhoeg, thanks for your PR! By analyzing the history of the files in this pull request, we identified @vcunat, @edolstra and @FRidh to be potential reviewers.

This seems to more than double the closure size of both the bin and out outputs. Maybe it should be build without icu by default?

Good point @FRidh. So what exactly is the policy about what dependencies should be included? Any sort of best practice?

Well, the Nixpkgs manual only mentions multiple outputs, but that won't help us here.
I think the best practice is here, put in only what you need, and everything else is optional.

If someone needs xmllint with ICU support they can build a version of libxml2 that has it. Therefore I suggest you make it an optional feature.

Please rebase your work to staging.

Sure, why?

Shouldn't actually be necessary, but that was just to be sure we won't get a mass-rebuild (on master) even when ICU is off by default.

Nice catch @FRidh!

Indeed, the changed order of the buildinputs was enough to cause a massive rebuild.

I managed to massage it into submission (at least according to nox-review) so it should be OK now.

Do we have some idea about the use cases for ICU in there? /cc @fubarbaz.

@vcunat, as I understand it, some libraries will use icu while others use libiconv and to avoid ambiguity, you would want to ensure a uniform unicode conversion library all across.

But @FRidh is absolutely correct that it should be disabled by default (and now is).

Yes, at least default for use in nixpkgs, from what I know so far. The ICU data are rather large.

For installing the preference might be different. I wonder whether to provide at least an explicit attribute, to allow easier usage and provide binaries.

@vcunat we could add a libxml-icu in all-packages.nix, but, considering this is the first time it came up I think we shouldn't (yet).