Partial package updates

[carlobaldassi]

This introduces the possibility to do Pkg.update(pkgs...) and only update a subset of the packages.

The result is that only the packages explicitly provided and their dependencies will be updated, everything else is kept fixed.

For example:

julia> Pkg.update("Colors")
INFO: Updating METADATA...
INFO: Computing changes...
INFO: Upgrading Colors: v0.6.4 => v0.6.5
INFO: Upgrading Compat: v0.8.1 => v0.8.2

Also, the packages explicitly provided will never get downgraded (which is a rare occurrence, but in principle may happen with a full Pkg.update()).

In case the update fails to install the latest version of the packages explicitly provided, an info message is emitted, and some analysis is attempted to suggest possible causes (a full analysis is basically out of the question, and unlikely to be of any practical relevance IMO). For example, in a case in which I had artificially capped the version of Colors.jl:

julia> Pkg.update("Colors")
INFO: Updating METADATA...
INFO: Computing changes...
INFO: Package Colors was set to version 0.6.4, but a higher version 0.6.5 exists.
      The update is prevented by explicit requirements constraints. Edit your REQUIRE file to change this.
INFO: No packages to install, update or remove

Still missing: tests, perhaps extra docs (? Someone please advise me on this, I only updated the docstring and it works in the repl, should I do something else?), probably some note in NEWS.

We may also want to add a keyword argument to truly restrict the update to the packages given as arguments, not their dependencies (e.g. dependencies=false? Suggestions welcome). This however would require some additional work, especially to perform the analysis afterwards. Maybe it can be done as a second PR.

I tried to test this functionality in a variety of cases (registered packages, unregistered ones, dirty ones, checked out, etc.) but it would be best if other people could put this to test as well.

[edit by tkelman: closes #13487]

[toivoh]

If packages B and C depend on package A and you do

Pkg.update("B")

could this update A to a version that is incompatible with the currently installed version of C?

[carlobaldassi]

@toivoh: no, dependency resolution is done the same way as always, just with additional constraints on packages which shouldn't be updated. If it happens, it's a bug.

[tkelman]

I think these should be separate docstrings for the docsystem to handle them properly - the new signature should be added to the rst version of the docs (kinda messy at the moment) so make docs populates the rest of it

[carlobaldassi]

Added tests (kind of messy, but the best I could come up with), docs, item in NEWS, and rebased.
This should now be ready to go.

[tkelman]

what is the Bool in this Tuple?

[carlobaldassi]

It's the fixed tag as returned by Read.installed, see here.

I tend to spell out the type signatures in full in these functions (rather than just writing instd::Dict) mostly for "documentation" purposes.

[carlobaldassi]

I have addressed all the comments (as always, thanks @tkelman!) and rebased.

[tkelman]

❤️ the comments

[tkelman]

This is great, glad you had a more productive JuliaCon hacking session than I did. Kind of sneaking in a feature, but I say we merge it.

A side-effect of having this feature is that too-loose version bounds in packages on their dependencies will now be more visible. We may have to work out an automated cross-package multi-version testing system to identify that kind of issue.

[StefanKarpinski]

I'm also in favor of merging this. People will be a lot less unhappy with the current package manager if they don't have to wait until Pkg3 for this nice functionality :)