Summary of the problems with git pull

Hi,

There has been a lot of discussion about why `git pull` is broken for so many many workflows: [1][2][3][4][5], even as far back as [6].

Many issues has been brought up, and many proposed solutions, probably too many for most people properly digest them, so here I'll try to synthesize them.

Mainly there are two core problems with `git pull`:

 1) Many times it does a merge when it's not when people want
 2) Many times it does a merge opposite of the desired direction

The issue comes in trying to solve these problems in a way that would not affect anybody negatively.

The simplest way to fix these issues unobtrusively would be to add configurations for the new behavior. However, this wouldn't solve the problem because as it has been discussed the vast majority of people doing these "bad merges" are not advanced users, so they wouldn't activate these options.

What is needed is to change the default behavior, but in a way that is not obtrusive, and with backwards compatibility configuration.

The most concrete proposal is my patch series that puts everything in place to enable fast-forward merged by default only. However, that still doesn't solve the problem of the ordering of parents. Normally these two issues would be independent of each other, but after further analysis I think they are not.

== Two different kinds of pulls ==

It has become clear that the `git pull` command is in fact used to do two very different things:

 1) Update the current branch

Most people do `git pull` like they would do `svn update`; to update their current branch.

 2) Merge a remote branch

There are many use-cases of this.

The agreement so far is that 1) is the one that is broken, that is; a) by default only fast-forwards should be allowed, and b) when a merge happens the parents should be reverted (merge 'master' to 'origin/master').

It would be possible to differentiate these kinds by saying a `git pull` that doesn't specify the location can be assumed as 1), and a `git pull remote branch` that does as 2). Another possibility is to assume only the upstream branch corresponds to 1).

Unfortunately it's the feeling of many people that this solution is not "clean", because it's two very different behaviors for the same command. Furthermore even deeper analysis of the use-cases demonstrates there would be a need to have different configurations for the different modes (e.g. pull.updateMode and pull.integrateMode).

== git update ==

Another proposed solution is to have a new command: `git update`. This command would be similar to `git pull --ff-only` by default, but it could be configured to do merges instead, and when doing so in reverse.

An interesting side-effect of this new command is that it opens the possibility of thinking about `git update --rebase` vs. `git pull --rebase`.

For example:

  a) git update --merge

     Merge HEAD into @{upstream}

  b) git update --rebase

     Rebase HEAD onto @{upstream}

  c) git pull --merge github john

     Merge github/john into HEAD

  d) git pull --rebase github john

     Rebase github/john onto HEAD

Notice how the relationships between them are nice and consistent. Also, d) was not possible before.

This solves essentially all the issues people presented in their use-cases, except the differentiation between merging topic and upstream branches, for which we might want to add different options in `git pull`, but that's independent from the issues I mentioned at the beginning.

Nothing changes for the users of `git pull`, except that perhaps we would want --rebase to work in reverse, since the current `git pull --rebase` would already rebase HEAD onto @{upstream}.

Personally my next step would be to port the changes I did for pull.mode = merge-ff-only into a new command `git update`, which would probably be a copy-paste from `git pull` and see how that turns out. In addition, when running `git pull` without arguments we might want to add a temporary notice explaining that perhaps the user wanted to type `git update` instead.

Cheers.

[1] http://article.gmane.org/gmane.comp.version-control.git/233554 [2] http://article.gmane.org/gmane.comp.version-control.git/234295 [3] http://article.gmane.org/gmane.comp.version-control.git/225146 [4] http://article.gmane.org/gmane.comp.version-control.git/247237 [5] http://article.gmane.org/gmane.comp.version-control.git/247939 [6] http://article.gmane.org/gmane.comp.version-control.git/130819

Felipe Contreras wrote in message <5366db742d494_18f9e4b308aa@nysa.notmuch>:

Thanks for the nice summary.

As a user, I am very much in favor of adding a `git update` command. In fact I already have a ruby script that does such a thing (fast-forward all my local branches that have an upstream configured), so it would be nice to have it directly in git core.

Ping Yin

On Mon, May 5, 2014 at 8:29 AM, Felipe Contreras <felipe.contreras@gmail.com> wrote:

+1 for git update

Felipe Contreras wrote:

And here it is:

https://github.com/felipec/git/commits/fc/update

Here's the documentation:

https://github.com/felipec/git/wiki/git-update

Works exactly as expected: non-fast-forwards are rejected by default, but can be configured, and parents are reversed when merging.

I know patches from Felipe are sensitive subject and that a lot of people probably simply ignore him entirely, however I think this patch deserves more attention.

more or less in line with what Junio suggested what should be done in [1]. So the goal of this mail is just to make sure people consider it for inclusion.

Sorry if I missed a thread where it was already decided not to include it.

Felipe, please don't use this to start any non-constructive behavior (rant on who is right/wrong, "my patches are not accepted", etc).

Philippe

[1] http://article.gmane.org/gmane.comp.version-control.git/248258

Philippe Vaucher wrote:

I never sent those patches. I gave up on the Git project.