[QUERY] Why do we have git-completion.zsh?

Hi,

I realize that we maintain a bash completion script, and a thin wrapper around it for ZSH. However, I don't understand why we maintain it, because there's a comprehensive first-class completer in ZSH core [1] which I use all the time. Shouldn't the completion folks be contributing to this instead?

Thanks.

[1]: https://github.com/zsh-users/zsh/blob/master/Completion/Unix/Command/_git

Hi,

Ramkumar Ramachandra wrote:

Only if they want to.

Kind regards, Jonathan

http://thread.gmane.org/gmane.comp.version-control.git/210022/focus=210024

Jonathan Nieder wrote:

[...]

Sorry, wrong link.  Here's a better one:

http://www.zsh.org/mla/workers/2011/msg00497.html

More fundamentally, your question seems to assume some kind of shared plan regarding what people work on, rather than each person choosing to work on what they consider valuable without having to justify it to others. I am very happy with Felipe's work on git's completion code and don't think it does any harm to efforts elsewhere.

Thanks for your interest, Jonathan

On Sat, Apr 20, 2013 at 1:53 PM, Jonathan Nieder <jrnieder@gmail.com> wrote:

To complement the reason; the zsh folks (or perhaps it's only one; who works on the git stuff), absolutely prioritize "correctness" over speed, that means if it takes ten seconds to list all the possible files to complete, grouped nicely, that's exactly what they'll do, instead of the way the bash completion does, which lists the most likely files, which takes milliseconds.

I found zsh's completion unbearably slow, which is ironic; git is the fastest DSCM on Earth, and they blow the performance away by making the completion dead slow, to the point where the completion is several orders of magnitude slower than the actual command. To me, the whole point of the completion is to make the work faster, and if the completion takes longer than me typing it, what's the point of it?

And I'm not the only one, when I contributed this stuff to oh-my-zsh[1] I got a very positive response.

I couldn't convince them of this obvious fact, so I choose git's bash fast completion.

I have patches to improve the zsh thin wrapper to give more information, zsh style, in a way that is superior to zsh's git completion, but I'm still not happy with them, I want the user to be able to configure the output so it's not totally bloated like in zsh, and I'm still working on it. In the mean time, I want my completion to be blazingly fast, and at least as good as git's bash completion.

If they wanted my contributions, that would be awesome, but that doesn't appear to be the case. And I'm kind of relieved, because zsh's completion is monstrous code, and the zsh development practices are not good (e.g. all my logically independent patches get squashed into one commit).

Cheers.

[1] https://github.com/robbyrussell/oh-my-zsh

On Sat, Apr 20, 2013 at 5:39 PM, Felipe Contreras <felipe.contreras@gmail.com> wrote:

Actually, it's mostly there now: https://github.com/felipec/git/blob/fc/zsh/ng/contrib/completion/git-completion.zsh

This greatly improves 'git <tab>' over the plain bash version, and it's even better than the official zsh version, because you can do something like:

  zstyle ':completion:*:*:git:*' tag-order 'common-commands porcelain-commands'

To show only the command groups you are interested on. With official zsh, you get the whole enchilada of all the commands git has. Hardly user-friendly.

Cheers.

Felipe Contreras wrote:

I see.  Yes, I was looking for a reason like this.

True. I've noticed that my shell just hangs when I attempt to tab-complete too early/ something wrong, and this would be nice to fix. Does it have to be a hard trade-off between correctness and speed? Is it not possible to reach a fair compromise?

Now, I don't know anything about zsh's git completer versus the gitfast completer. From a glance, it looks like zsh's git completer is much larger and stuffed with features that the gitfast completer doesn't have. Although I agree that speed isn't the only parameter, can you shed some light on how these two compare on other parameters?

Or you could help me figure all this out myself. How do I profile/debug shell code? I'm feeling a little lost without gdb and valgrind.

Found it: https://github.com/robbyrussell/oh-my-zsh/commit/5a11228

I see. I've just started poking them with patches. I'll see what happens for myself.

Cheers.

On Sat, Apr 20, 2013 at 11:36 PM, Ramkumar Ramachandra <artagnon@gmail.com> wrote:

It doesn't matter, the zsh folks don't want *any* compromise: http://article.gmane.org/gmane.comp.shells.zsh.devel/22475

First of all, 'gitfast' is just the name I gave to the oh-my-zsh plugin that uses git.git's completion stuff. The zsh support in git's bash completion has been working for years, I just copied the stuff to oh-my-zsh so those guys can use it easily. I also added my thin zsh wrapper, because of issues with zsh's bash completion emulation, but that's not specific to oh-my-zsh in any way.

Secondly, I don't see what "features" zsh's git completer might have that we don't. Yes, some specific argument behaviors are nice, like adding a '=' after --git-dir, and then complete only directories, and completion descriptions, along with tag groupings, but all those things are cosmetic, and they could be added relatively easy to my thin wrapper (which wouldn't be so thin after all). It's mostly grunt work, not something that requires a great mind.

Functionally I don't see any value.

For example, a difference is that when you do 'git show v<tab>', git.git bash's completion will show refs, and only if there aren't any that could match, show files, any files. zsh's version will show both files and refs at the same time, and if you have grouping enabled they would show that way, in the git.git repo I see them grouped in commit tag, tag, 'cached file' (although showing tags twice is probably a bug), and only the files present in the repository.

Another one is that when you do 'git add <tab>', you see the results grouped for example in 'modified file', 'untracked file', whereas in bash they can't be grouped.

Are these minor features worth all this slowness and complicated code? I don't think so.

Also, the grouping requires a configuration like: zstyle ':completion:*:descriptions' format '%B%d%b' zstyle ':completion:*' group-name ''

And I bet most people don't have such configuration. So all this complicated code achieves nothing for the majority of users.

The reason why I prefer git.git's bash completion is that it has taken years to develop, and using good development practices, borrowed from the mainline git process. Many more people use them, have debugged them through the years, and optimized them. It's relatively small (compared to zsh's version), much more readable, and it even has tests (which I helped to start), and it's much less buggy. It's basically rock solid.

Just now I was reading about a bug in zsh-workers when you do 'git show v3<tab>' in the Linux kernel git tree, and you can count the seconds before it completes. This is the sort of thing that just doesn't happen with git's bash completion.

For git.git's bash completion I just run a script that exercises the function I'm interested in with perf. Fortunately nothing there is really slow, so I don't have to track offenders (usually).

Unfortunately zsh's completion stuff can't be scripted, or at least I haven't managed to find a way to either profile or write tests for any of that code.

Yeah, I did contribute patches myself too. As long as they are superficial, like adding a missing argument or so, they are fine. But don't try messing with their precious "completeness".

Cheers.

Felipe Contreras wrote:

Yeah, I know.  I just used the name.

Moved to using git.git's completer, and I see that you're absolutely right about the "minor features" missing part. I just assumed that zsh's completer must be more complete because it's so much larger than git.git's bash/zsh completer. Working backwards from zsh's completer would've been a nightmare.

Great! I'll stop working on zsh's completer immediately, and try to redirect my attention on improving git.git's completer instead.

Thanks for the information.