Re: [PATCH 2/2] Be more careful when determining whether a remote was configured

On Wed, Jan 18, 2017 at 05:22:40PM +0100, Johannes Schindelin wrote:

OK. Though after reading your message over several times, I think I may have confused things by raising two separate issues.

So let me re-state my issues here for clarity:

  1. I'm concerned that a setting for remote.<url>.X would fail to apply
     to a bare "git fetch <url>" if "X" is considered as "not really"
     configuring a remote.

  2. I'm concerned that splitting the remote.*.X keys into two classes:
     "really configures" and "does not really configure" creates a
     confusing interface for users. Some keys are OK to set in your
     ~/.gitconfig and others are not.

Let's talk about concern 1 first. Based on your analysis in this message, it looks like it _isn't_ a problem with your patch (because is_configured is never used for applying values, only for add/del/rename types of operations in remote.c).

So good. Thanks for addressing my concern. And that makes your "conservative" comment make more sense; the idea is that you are not breaking anything, but just loosening selectively for some values of "X".

I think there are still some weird corner cases even for "prune", though. E.g., try:

  git init
  git remote add foo whatever
  git config remote.foo.prune true
  git config remote.other.prune false

So now we have:

	[remote "foo"]
		url = whatever
		fetch = +refs/heads/*:refs/remotes/foo/*
		prune = true
	[remote "other"]
		prune = false

Now try "git remote rename foo other". With current versions of git, it's disallowed. With your patch, I get:

	[remote "other"]
		url = whatever
		prune = true
	[remote "other"]
		prune = false
		fetch = +refs/heads/*:refs/remotes/other/*

The old value of remote.other.prune is overriding the one we tried to move into place.

In your motivating example, the old value is in ~/.gitconfig, so it automatically takes lower precedence, and everything just works (I think it would also just work if "other" had been defined originally _before_ foo in the config file).

I think you could fix it by having git-remote teach the "not really" config values (like "prune") to overwrite any existing value when rewriting the config. I think this just needs the multi_replace flag set when calling git_config_set_multivar().

That raises questions about what should happen when multi-value keys like "refspec" would be set (if we were to add them to the "not really" set). Should they be overwritten, or merged? And in that sense, your patch lets you punt on those issues.

I still think my second concern is valid. It's made worse by your patch (if only because everything was disallowed before, and now some things are and some things aren't). If this is a first step towards a final state where the rules are simple, then starting conservatively makes sense. And until we get there, the answer "yes, it should, but nobody has worked out the semantics yet; care to make a patch?" is an OK one. But it sounds like you do not ever want to loosen the "refspec" case.

I don't think that's ideal, but at the very least if that's the end state, the list of "OK to use in ~/.gitconfig" keys should probably be documented.

Reading your message, though, I still wonder if we can do better...

I'd add to this that "git remote add <new>" should work in a similar way (that was the one that I think people often ran into with remote.origin.fetch refspecs).

I think _this_ is a much better way of framing the problem. It is not about which keys are set, but about _where_ they are set. IOW, a reasonable rule would be: if there is any remote.*.X in the repo config, then git-remote should consider it a configured repo. And otherwise, no matter what is in ~/.gitconfig or elsewhere, git-remote should proceed as if it doesn't exist (and repo-level config can take precedence over config defined elsewhere).

I.e., something like this: