Tracking parent branches in Git

Hello,

We would like to track parent branches so that creating pull requests can automatically determine the correct branch to merge against.  I understand that this would require tracking more information than is currently available right now in git.  Also, it seems that if some cases, it is not possible to determine a parent branch, in which case it would just be empty/null.  If I made a change to track the parent branch for each branch, would this feature be accepted/welcomed as part of git, even if it off by default?

Thanks, Eric

:(){:|:&};:

Eric Kulcyk <Eric.kulcyk@microsoft.com> writes:

[Overly long lines are not appreciated around here.]

Do you mean by "parent branch" what people usually call "upstream branch" (i.e. when that branch on the other side gains more commits independent from what you have been working on, then you would want to rebase your work on top of the updated state of that branch on the other side) around here? Perhaps "git help glossary", look for "upstream branch" and start from there? The entry mentions the configuration variables used to keep track of that information, which are described in "git help config", I think.

Regardless of what is being proposed, this is often not a very useful question. Something worth doing for yourself is worth doing whether others also find it useful ;-) And others usually do not have enough information to judge if such a change is welcome until seeing it in a bit more concrete form.

Thanks.

On Mon, Jul 1, 2019 at 12:35 PM Junio C Hamano <gitster@pobox.com> wrote:

I suspect the question is in regards to "What branch did I create my local branch from?", especially given the pull request reference.

In other words, when I locally do: git checkout --no-track -b bturner-some-bugfix origin/release/5.16

release/5.16 is the "parent branch" of my bugfix branch and, when I push my branch and try to open a pull request, release/5.16 is a _likely_ target for where I'd want to merge it. There may be a remote in the name, a la "origin" in my example, or it might be created on top of some other local branch. It's a common feature request for Bitbucket Server[1], for example, to automatically select the "right" target branch for a new pull request based on the ancestry of the branch in question--except branches have no ancestry. (This sort of metadata could potentially offer some benefits for building commit graphs (referring to UI treatments for visualizing the DAG, rather than Git's "commit-graph" functionality), depending on how it was implemented, since it would make branch points more stable.)

Since branches are ephemeral names and have no intrinsic metadata of their own (unlike, say, annotated tags or commits), I suspect implementing something like this may be more complicated than it might initially appear, especially if said metadata needs to be communicated to remote repositories (which implies it might require changes to the wire protocol as well).

Best regards, Bryan Turner

[1] https://jira.atlassian.com/browse/BSERV-7116

On July 1, 2019 3:48 PM, Bryan Turner wrote: On Mon, Jul 1, 2019 at 12:35 PM Junio C Hamano <gitster@pobox.com> wrote:

I suspect the question is in regards to "What branch did I create my local branch from?", especially given the pull request reference.

In other words, when I locally do: git checkout --no-track -b bturner-some-bugfix origin/release/5.16

release/5.16 is the "parent branch" of my bugfix branch and, when I push my branch and try to open a pull request, release/5.16 is a _likely_ target for where I'd want to merge it. There may be a remote in the name, a la "origin" in my example, or it might be created on top of some other local branch. It's a common feature request for Bitbucket Server[1], for example, to automatically select the "right" target branch for a new pull request based on the ancestry of the branch in question--except branches have no ancestry. (This sort of metadata could potentially offer some benefits for building commit graphs (referring to UI treatments for visualizing the DAG, rather than Git's "commit-graph" functionality), depending on how it was implemented, since it would make branch points more stable.)

Since branches are ephemeral names and have no intrinsic metadata of their own (unlike, say, annotated tags or commits), I suspect implementing something like this may be more complicated than it might initially appear, especially if said metadata needs to be communicated to remote repositories (which implies it might require changes to the wire protocol as well).

Best regards, Bryan Turner

[1] https://jira.atlassian.com/browse/BSERV-7116

Was there not, at some point in recent history (2019), a discussion about storing extra arbitrary data associated with a branch or other objects? My thought for satisfying what Eric was originally proposing is to store the root commit associated with the original branch HEAD when checkout -b/branch was done to create the branch. Presumably another datum could store the branch that the branch HEAD was on, but that may not be unique - which is a root part of the problem with this request, although it might be something that the user could select/specify - not sure how - at branch creation. 

But aside from that both of the above are transient relative to the new branch and by the time you wanted to create a Pull Request, the information you originally wanted could irrelevant - at least to git. If I was the product manager on this, I would suggest going to GitLab, GitHub, or BitBucket and asking for some augmented capability of branch creation, that stores the data for future Pull Request management - instead of doing this in core git because of the transient nature of the relationship between a branch and a commit.

My $0.02. Randall

On 01/07/2019 21:12, rsbecker@nexbridge.com wrote:

 From the Git side, maybe one could simply populate the branch 
description with the commit oid and branched-from name at the time of 
branch creation (no doubt set as a core.option).
The field is already there and almost never used - there's no easy way 
(via git command) to populate the description anyway. Plus its a local 
field, keeping Git distributed.
--
Philip

Philip Oakley <philipoakley@iee.org> writes:

I do not think you want branch.description to get mixed-up in this.

In this whole thread, I have been wondering if I am missing something crucial, but now I am deeply puzzled why after many people made comments, nobody raises a question about the "--no-track" thing in the early message in the thread.

If you do not add that, i.e.

	$ git checout -t -b bturner-some-bugfix origin/release/5.16
        
(note that I added '-t' for illustration, but it should be on by
default when starting from origin/<whatever>), then you'd get in
your configuration file these recorded:

	$ git config --get-regexp 'branch\.bturner-some-bugfix\..*'
	branch.bturner-some-bugfix.remote origin
	branch.bturner-some-bugfix.merge refs/heads/release/5.16

You created 'bturner-some-bugfix' branch out of the 'release/5.16' branch taken from the remote whose name is 'origin'.

Is that different from the answer to the question being sought? What am I missing???

On Tue, Jul 2, 2019 at 10:36 AM Junio C Hamano <gitster@pobox.com> wrote:

Sorry, I should have clarified my "--no-track" in my original message when I provided the example. I did "--no-track" because if I push "bturner-some-bugfix" to a server, I'm likely going to do something like "git push -u origin bturner-some-bugfix" so that my local "bturner-some-bugfix" branch will track the remote version of itself. At that point, the remote-tracking information would change from "release/5.16" to "bturner-some-bugfix" (without any sort of warning, for whatever that's worth), effectively "losing" the ancestry.

The other issue is that my local remote-tracking information doesn't help the server I'm talking to; it's not shareable. Assuming I could use remote-tracking to track ancestry, there's still no way to communicate that to the server so that it could know, when I go to create a pull request for "bturner-some-bugfix", that it's tracking "release/5.16" in my local repository.

I could certainly be misunderstanding the request, but I think it's asking for something less ephemeral--and more shareable--than remote-tracking, and it seems logical to want to be able to retain ancestry while still using remote-tracking setup such that local branches still track the remote version of themselves, rather than some other (albeit related) branch.

Best regards, Bryan Turner

On Tue, Jul 02, 2019 at 11:52:35AM -0700, Bryan Turner wrote:

I think the real problem with all of this feature request is that it's all presuming a particular workflow, and git is currently *not* strongly opinionated about the workflow.

So for example, if I create a local branch against origin/release/5.16 with a series of fixes, and I'm pushing to a gerrit server for review, I might need to do something like this:

  git push gerrit HEAD:refs/for/release/5.16%r=reviewer@example.com,%cc=kernel-reviewers@example.com

So the fact that we've recorded the information about the parent branch when doing a "git checkout -b bugfix origin/release/5.16" is part of the puzzle, but the other part of the puzzle is knowing what the destination server is going to want or need. Internally at $WORK we have our own internal script that does this automatically, so I just have to run something like "kdt mail -r reviewer@example.com", but it's always going to be very workflow-independent.

In order for it to be shareable, we have to make some very strong assumptions about workflow. It gets worse when people pull from more than one tree, or there is a series of subtrees.

For example, suppose the 2nd-level lieutenant has a "net" tree, and the 3rd-level lieutenants have a "sctp" and a "ipsec" tree. Now let's assume that there is some common infrastructure branch that both 3rd-level lieutenants need to build off of that comes from some other 2nd-level tree, and so for that development cycle, they first pull from the top-level tree, and then both 3rd level trees merge in a branch from the "mm" tree, but then they will be both be sending pull request emails to the "net" tree.

There will be a whole series of git pulls and git merges that will be in flight, and just for yucks, let's suppose some trees are do patch review via email, and some trees are doing patch review via gerrit, and yet other trees have their maintainer do a quick cursory review at "git pull" time after the tree sends a git pull request.

What sort of shareable context should git --- which is ignorant of all of these arrangements --- be preserving? And presenting to whom? And for what goal/purpose?

						- Ted

On 02/07/2019 20:24, Theodore Ts'o wrote:

I'd suggest that git does have a clear preference for a workflow that is based on the "upstream" view. (e.g. see 'rebase')

While Git tries to avoid being opinionated, it does develop a bias of avoiding various common workflows, to the point that they are not even well named.

In particular, we have a number of variants of the triangular workflow, such as having a personal github fork, and a then also a maintainer's repo to complete the triangle [which is even worse for those supporting Git-for-Windows because of two level maintenance cascade, with different compilation and OS requirements..]. Having three 'origins' is no fun.

There was an attempt to document a triangular flow a while back, but it wasn't actually that one. e.g. [1])

The other preference aspect is the tendency to expect users to 'support the machine' (e.g. the assume-unchanged file flag, which is commonly misunderstood), rather than having a clarity about when Git will support the user. Here we (they) are looking for the human readable name of the branch that they forked from (and is likely to have been extended since), rather than the oid hash of the fork point, though that is clearly useful and should be recorded with the branch name as it is immutable.

I am cautious about support for Gerrit on the basis that it could accidentally reinforce a centralised workflow (to the detriment of distributed operation), which should be avoided strongly as a deliberate bias.

However the desire to _locally_ record the branch name that the current branch was created from is something I would support (which is why I suggested the unused branch description...).

Philip

[1] https://public-inbox.org/git/5A8F8EE0162B49818813DAEFD68F61DA@PhilipOakley/

(sorry for any delays, I'm still in catch-up mode)

On Mon, Jul 01, 2019 at 12:48:16PM -0700, Bryan Turner wrote:

You can right now give meta data of your choice with --push-option to the push command. The Gerrit system makes use of that already. However, this would not be intrinsic to Git, but the serve needs to react on those options. And it should be in good company with suitable client tools.

Take care, Eckhard

Thanks for the feedback, is there an email client or tool that can format the lines correctly?

Yes, this is what I meant.

@Eckhard, is that documented somewhere? I don't see it on https://git-scm.com/docs/git-push/1.6.4.1

I wanted to get a some thoughts before I worked out the details. The feature is of little use if no one is using the version of git with it since most everyone uses standard distros of git.

Multiple branches can have the same commit, but only one branch can be checked out at once, right? Then the parent branch would be the branch you were on when you ran checkout -b. That branch might change and no longer be of use, but it seems like in practice that is not the case. New features are usually created off of a local branch and then PR'd back to the upstream version of that branch.

Thanks! Eric

________________________________________
From: Eckhard Maaß <eckhard.s.maass@googlemail.com>
Sent: Monday, July 1, 2019 1:40 PM
To: Bryan Turner
Cc: gitster@pobox.com; Eric Kulcyk; git@vger.kernel.org
Subject: Re: Tracking parent branches in Git

On Mon, Jul 01, 2019 at 12:48:16PM -0700, Bryan Turner wrote:

You can right now give meta data of your choice with --push-option to the push command. The Gerrit system makes use of that already. However, this would not be intrinsic to Git, but the serve needs to react on those options. And it should be in good company with suitable client tools.

Take care, Eckhard

Interestingly a Google search for git --push-option turns up old documentation first, I found the newer docs. Don't you still need a way to store this data locally? For instance, a wrapper around git checkout could record the branch. However, you would be unable to use vanilla git check out if you want to do rely upon the data.

From: Eric Kulcyk

Sent: Monday, July 1, 2019 1:58 PM

To: Eckhard Maaß; Bryan Turner

Cc: gitster@pobox.com; git@vger.kernel.org

Subject: Re: Tracking parent branches in Git

Thanks for the feedback, is there an email client or tool

that can format the lines correctly?

Yes, this is what I meant.

 be created on top of some other local branch. It's a common feature request for Bitbucket Server[1], for example, to automatically select the "right"

 DAG, rather than Git's "commit-graph" functionality), depending on how it was implemented, since it would make branch points more stable.)

@Eckhard, is that documented somewhere?  I don't see it on

https://git-scm.com/docs/git-push/1.6.4.1

https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fjira.atlassian.com%2Fbrowse%2FBSERV-7116&data=02%7C01%7CEric.kulcyk%40microsoft.com%7Ca780a740f5894ba5ce6508d6fe607317%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C636976088301830933&sdata=1KrFM%2BLcKoa3FCc1F86jIVZBrIUGi%2B6ad7CuA8ekAUY%3D&reserved=0

I wanted to get a some thoughts before I worked out the details.

The feature is of little use if no one is using the version of git with it

since most everyone uses standard distros of git.

 original branch HEAD when checkout -b/branch was done to create the branch. Presumably another datum could store the branch that the branch HEAD was on, but that may not be unique - which is a root part of the problem with this request, although it might be
 something that the user could select/specify - not sure how - at branch creation.

 going to GitLab, GitHub, or BitBucket and asking for some augmented capability of branch creation, that stores the data for future Pull Request management - instead of doing this in core git because of the transient nature of the relationship between a branch
 and a commit.

Multiple branches can have the same commit, but only one branch can

be checked out at once, right?  Then the parent branch would be the

branch you were on when you ran checkout -b.  That branch might change

and no longer be of use, but it seems like in practice that is not the case.

New features are usually created off of a local branch and then PR'd

back to the upstream version of that branch.

Thanks!

Eric

________________________________________

From: Eckhard Maaß <eckhard.s.maass@googlemail.com>

Sent: Monday, July 1, 2019 1:40 PM

To: Bryan Turner

Cc: gitster@pobox.com; Eric Kulcyk; git@vger.kernel.org

Subject: Re: Tracking parent branches in Git

On Mon, Jul 01, 2019 at 12:48:16PM -0700, Bryan Turner wrote:

You can right now give meta data of your choice with --push-option to

the push command. The Gerrit system makes use of that already. However,

this would not be intrinsic to Git, but the serve needs to react on

those options. And it should be in good company with suitable client

tools.

Take care,

Eckhard

On Mon, 01 Jul 2019 12:48:16 +0000, Bryan Turner wrote: ...

We have simply conventionalized this - the parent relation is in the branch names. Your bugfix branch would be release/5.16/bturner-some-bugfix (in the central repo; we don't care how you name it locally), and, because ref storage, the parent would be release/5.16/master.

You'd just do

  git create-br release/5.16/bturner-some-bugfix

and it would be branched off the corresponding /master, be checked out, and tracking already been set up. Likewise we have a

  git update

which looks at the upstream name, deduces the parent, and pulls that in, with a suitable commit message. Finally,

  git mkpullreq

creates a pull request from the current to the parent branch.

(I would like to have a way to make bitbucket server use the same convention for the default pull request target.)

- Andreas