Conservative updates #4676

First shot adding new update CLI options and passing those through Definition into Resolver where they're needed. I opted for three separate options to reduce typing, vs. a --level [patch|minor|major]. It's a little painful merely extending parameter lists in places, but bigger refactorings prolly be more painful.

Putting all this code into Resolver itself, when it's almost completely isolated, didn't make sense. UpdateOptions was the best place for it, prompting me to think that class needs renaming and see if the current search_for method could be moved into it as well. Except for `index_for` it's a cinch. `install_spec` and `update_spec` are still passing with this addition, first new spec for conservative update isn't yet.

Prior commit added in the code but it only worked in the default :major case by staying out of the way, so it at least didn't break existing functionality. But the GemsToPatch class hadn't been brought over from bundler-patch yet, so none of the patch/minor code could work as it was referencing an ivar that didn't exist yet. Bundler-patch not only will handle an update to whatever is the latest patch/minor version, but will also handle an update to a version specified in an advisory from ruby-advisory-db. At this point I don't believe we'll be adding the advisory behavior to Bundler proper, so this commit can bring over simply the array of gem names being updated without the GemsToPatch class or the extra sorting code that takes a specific version number into account. With these simplifications, the starter update_spec can run without exception, though for some reason it's behaving as if the :minor level were specified, instead of :patch. Rather than keep debugging from an integration level, it's time to start bringing over resolver tests.

I stared at that code 2 or 3 times before seeing the obvious bug. Sigh. Anyway - needed to get these resolver specs going sooner or later. Both the first new resolver spec is passing as is the first update_spec spec (plus the pre-existing ones).

Ported over resolver specs from bundler-patch README (which don't match the actual specs for no particular reason other than dev lolz) and expanded on some of the cases after a few weird cases resolved ... a little unexpectedly. They're passing, but some discussion to be had on some of the cases perhaps. See inline spec comments.

UpdateOptions which was then renamed to DependencySearch is now called GemVersionPromoter, cuz I can't name this damn class. It's in its own file now, so there's that. I took a shot at moving Resolver#search_for into it, but had naively overlooked a few instance variables and such and it just didn't make as much sense as I'd first envisioned. Probably some other smaller classes in between perhaps. GemVersionPromoter class now caching its results, too, and I moved out the return from it back into Resolver as it made more sense there. As a standalone class, it may make sense to have this actually implement :major sorting, but maybe later.

Another bit brought over from bundler-patch, this code in Definition ensures the GemVersionPromoter has the locked specs it needs in the 'unlock all' case. Everywhere else in Definition, empty @locked_specs means unlock all, but doing conservative updates requires knowing the current locked version so it always needs this list.

A couple of confusing cases that should be clearer now and the differences called out. Details in the spec code and comments. Plus some MODO additions and other pending spec additions.

This shouldn't ever surface to the user, would be the result of a coding mistake.

In the process I was able to simplify some of the code inside GemVersionPromoter dealing with SpecGroups. I also attempted to implement the :major behavior into GemVersionPromoter as that would eliminate the logic to skip it in `#search_for`, however I ran into some test failures that I need to investigate further, though unit specs are working so far.

If GVP handles the default :major case, it now passes all the specs, but that's a lot of existing functionality to hand off to it at this stage, so I kept in the conditional to just roll with existing results if :major. Got rid of a couple of superfluous begin/end I'd included to make RubyMine auto-format the code in a way that made Rubocop happy.

...now that its output is inside the Resolver's cache.

Getting caught up on missing update_specs, realized `--strict` flagged wasn't being passed through. Fixed.

When I ported over specs from bundler-patch, I had a TODO on a combination that had no specs, so that's taken care of now.

@segiddins

This is the first behavior change from bundler-patch. Used to be older versions would never be an option, but Bundler proper has always supported this (if necessary to resolve the dependency tree) and there can be some legit cases for doing this. The `--strict` flag could be used to override this behavior, but I'm running into a Molinillo behavior that I'm not sure is correct, so the specs involving the strict option are failing right now. I'm going to push so @segiddins and I can discuss.

Also some pending specs cleanup. These may be added as new issues and addressed later.

segiddins and I are agreed there appears to be an issue with Molinillo in these cases, but they are unusual and for now we're looking to do an undocumented release of the new conservative updates, so we can start to get feedback. We'll revisit these cases.

Not a common case, but glitch is triggered by new plugin integration tests when there's no Gemfile or .bundle directory. A GemfileNotFound exception is raised deeper down the call stack trying to access the cache_path when executed in a non-bundler dir. That case apparently is legit for plugins.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Conservative updates #4676

Conservative updates #4676

Commits on Jul 9, 2016