Re: [ros-dev] Microsoft switched to Git

[Alex Ionescu]

I too have no idea where this "history limit" comes into play. I work on
and have repos spaning 50-100 developers working on dozens of thousands of
branches over two dozen repos with almost 8 years of history. Some people
rebase, but most people don't.

It is extremely easy to have a clean history due to the fact we enforce the
"[COMPONENT] Description" commit message (and can use a git hook to make
sure it looks this way). It is then trivial to write a script that only
looks for those commits. Heck you can even auto-squash/rebase anything that
doesn't match.

That being said, that type of "dirty history" only happens if you heavily
work with branches. That's not how reactos developers work -- we don't open
PRs and separate branches for every checkin.

These ALL sound like manufactured problems or poor/strange use of git.

On Thu, Feb 16, 2017 at 12:02 AM David Quintana (gigaherz) <
gigaherz(a)gmail.com&gt; wrote:

>   If git has an upper limit on history, it's the
> first time I hear about it,
>  and it would be an issue I was simply not aware of. So far as I was aware,
>  git just keeps a parent hash on each commit, and the GC just deletes
>  commits without any reference. In fact, we use Git at my current job, and
>  the history goes back many years, and it seems perfectly capable of
>  tracking it. And no, we don't rebase or squash, ever. We are not allowed to
>  do any operation that changes commit history. So it's quite a mess of merge
>  commits, but even then it's manageable.
>
>  I disagree on needing a single person responsible for merges. The the PR
>  workflow we have at my current job is that the developer submits a PR for
>  review, and adds to the PR two other developers, who have to approve the
>  changes or request fixes. Once it's approved, the original developer can
>  click the merge button themselves. For external PRs done by contributors,
>  then one of the reviewers is the responsible for clicking the merge button.
>
>  And even if you do choose to require a commit-master, with the way
>  github's merge button works, you can choose to have a merge commit, or to
>  squash & rebase implicitly, and the history ends up with one single commit
>  attributed to all the developers that were involved in the merge. As an
>  example, I have a hobby writing Minecraft mods, and they use Forge to work:
>  https://github.com/MinecraftForge/MinecraftForge/commits/1.11.x -- you'll
>  see many "<someone> commited with LexManos". That's because
> there's an
>  author, and a commiter, in the git info.
>
>  I do agree, git history can get messy, and ugly, and hard to track -- if
>  you abuse merge commits. I, unlike Alex, am a fan of manual rebases, and
>  leaving a clean history before you submit the changes for review. But I
>  understand that constantly rewriting history means you end up with a
>  fabricated history at the end. So the policy choices should probably either
>  be no merges at all, or no rebasing at all. But that's where "git blame"
>  comes in. With a proper tool, you can see exactly who changed a file, and
>  when, and you can use it to track back the history of that file, or that
>  specific line of code.
>
>
>
>  On 16 February 2017 at 03:45, Zachary Gorden &lt;drakekaizer666(a)gmail.com&gt;
>  wrote:
>
>  I've used both git and svn in work environments. If all you do is git pull
>  and git push, you end up with lots of noise in the commit log with git
>  tracking every single merge because you don't rebase. Combined with the
>  fact that git has an upper limit on how much history it can track and the
>  solution literally being to purge history, I'm not exactly sure why all of
>  you are so enthused about it. Unless the team wants to adopt having a
>  single person being responsible for all commits going into the canonical
>  master repo to avoid all of the problems with how git tracks history, the
>  commit log is going to be next to useless for actual tracking of history.
>  If you don't care about the commit history, then sure, go ahead, but I
>  personally would like to be able to easily track changes back cleanly. We
>  get that basically for free with svn right now. With git, the usage
>  patterns that those of you pushing for git are promoting actively works
>  against keeping a clean history.
>
>  On Wed, Feb 15, 2017 at 7:32 PM, Alex Ionescu &lt;ionucu(a)videotron.ca&gt; wrote:
>
>  Sure, I didn't count git add because you can do it with git commit -a.
>  git status/log are the same as the svn equivalents. just like git
>  diff/svn diff. I was mainly referring to regular workflow.
>
>  In fact, I think outside of stash (which is an optional, but awesome,
>  feature) fetch and rebase (which I refuse to learn), all commands map
>  1:1 with svn. That's why I don't get this whole "it takes way more
>  commands/steps in git".
>
>  git commit -a -m "[BOOTLIB] Fix yet another bug]"
>  git push
>
>  Done.
>
>  Best regards,
>  Alex Ionescu
>
>
>  On Wed, Feb 15, 2017 at 3:48 PM, David Quintana (gigaherz)
>  &lt;gigaherz(a)gmail.com&gt; wrote:
> >   My command set is a bit more extended:
> >
> >  git clone -- similar to svn checkout into a new folder
> >  git checkout -- for changing the current branch
> >  git pull -- effectively the same as "svn update", xcept it gets the 
> entire
> >   change history, not just the latest commit data
> >  git push [--force] -- for sending changes into the repository
> >  git fetch -- downloads stuff but doesn't apply it to the checkout copy
> >  git merge -- can be used to merge the remote data (in which case it's 
>  like
> >   svn update), or to merge from another branch
> >  git branch
> >  git add
> >  git commit
> >  git stash save/pop -- can be used to temporarily undo some changes, and 
>  be
> >   able to recover them afterward
> >  git status, git log, ... -- for getting info about the state of the
> >  repository and the uncommited changes
> >  ... and more I that I use less often
> >
> >  I do agree that it is a bit annoying that git has so much trouble pulling
> >  with local changes, and that is the one area where svn just simply works
> >  better. In every other aspect, I have come to like the "git way" more.
> >
> >  That said, I avoid commandline git as much as possible. I prefer to use
> >  TortoiseGit (in Windows, at home), or SourceTree (at work, where I use a
> >  mac, and SourceTree is probably the least shitty frontend for git).
> >
> >  I like to say, that for someone who knows Subversion, learning git 
>  starts by
> >   realizing that all the usual svn concepts, apply
> > to git, just NOT with 
>  the
> >   remote repository. The svn-like commands work
> > with the local repository
> >  clone, and then it has a separate command set for interacting with 
>  remotes.
> >   Of course it's not a 1:1 match, but it's
> > a good starting point. If you 
>  are
> >   able to "catch" that, then learning how
> > to work with git becomes a LOT
> >  easier.
> >
> >
> >
> >  On 16 February 2017 at 00:31, Alex Ionescu &lt;ionucu(a)videotron.ca&gt; wrote:
> > >  
> > >  On Wed, Feb 15, 2017 at 9:01 AM, Zachary Gorden
> > >  &lt;drakekaizer666(a)gmail.com&gt; wrote:
> > > >   Why is there a need for anything beyond "git
> > > > commit" or "git push" or
> > > >  "git
> > > >  pull" to do anything? 
> > >  Good question. I've never used any other git command other than those
> > >  (except git checkout). Oh, that's lie, I've also used "git branch",
> > >  just like on svn, to create a branch.
> > >
> > >  Sounds like you've never actually used git? I've never rebased in my
> > >  life, and I don't know what other commands even exist.
> > >
> > >  Best regards,
> > >  Alex Ionescu
> > >
> > >  _______________________________________________
> > >  Ros-dev mailing list
> > >  Ros-dev(a)reactos.org
> > >  http://www.reactos.org/mailman/listinfo/ros-dev 
> >
> >
> >  _______________________________________________
> >  Ros-dev mailing list
> >  Ros-dev(a)reactos.org
> >  http://www.reactos.org/mailman/listinfo/ros-dev
> >  
>  _______________________________________________
>  Ros-dev mailing list
>  Ros-dev(a)reactos.org
>  http://www.reactos.org/mailman/listinfo/ros-dev
>
>
>  _______________________________________________
>  Ros-dev mailing list
>  Ros-dev(a)reactos.org
>  http://www.reactos.org/mailman/listinfo/ros-dev
>
>  _______________________________________________
>  Ros-dev mailing list
>  Ros-dev(a)reactos.org
>  http://www.reactos.org/mailman/listinfo/ros-dev
>  
-- 
Best regards,
Alex Ionescu

[Colin Finck]

Am 16.02.2017 um 14:40 schrieb Alex Ionescu:
>   That being said, that type of "dirty
> history" only happens if you
>  heavily work with branches. That's not how reactos developers work -- we
>  don't open PRs and separate branches for every checkin.
>  
>  These ALL sound like manufactured problems or poor/strange use of git. 
That merge hell is easily reproducible using my default Git setup:

1) Change something in your clone of master and do a "git commit".
2) Let someone else change something in his clone of master and let him
"git commit" and "git push" it.
3) Try to "git push" your commit, won't work because of the commit of
the other person.
4) Do a "git pull" to fix the problem in 3) -> Bam! Git will do an
automatic merge of both masters and pollute your history.

You see, not a single extra branch is involved and yet we get two
parallel streams of history.

Rafal's mentioned "pull.rebase" option sounds promising, but can we
enforce that somehow?


- Colin

[Zachary Gorden]

The project that I worked with using git has history going back to 1988.
They certainly didn't start with git, nor did they necessarily start with
any revision control at the beginning, but after they migrated to it they
discovered the history problem.

On Thu, Feb 16, 2017 at 10:42 AM, Colin Finck <colin(a)reactos.org> wrote:

>   Am 16.02.2017 um 14:40 schrieb Alex Ionescu:
> >   That being said, that type of "dirty
> > history" only happens if you
> >  heavily work with branches. That's not how reactos developers work -- we
> >  don't open PRs and separate branches for every checkin.
> >
> >  These ALL sound like manufactured problems or poor/strange use of git. 
>  That merge hell is easily reproducible using my default Git setup:
>
>  1) Change something in your clone of master and do a "git commit".
>  2) Let someone else change something in his clone of master and let him
>  "git commit" and "git push" it.
>  3) Try to "git push" your commit, won't work because of the commit of
>  the other person.
>  4) Do a "git pull" to fix the problem in 3) -> Bam! Git will do an
>  automatic merge of both masters and pollute your history.
>
>  You see, not a single extra branch is involved and yet we get two
>  parallel streams of history.
>
>  Rafal's mentioned "pull.rebase" option sounds promising, but can we
>  enforce that somehow?
>
>
>  - Colin
>
>  _______________________________________________
>  Ros-dev mailing list
>  Ros-dev(a)reactos.org
>  http://www.reactos.org/mailman/listinfo/ros-dev
>

[Magnus Johnsson]

You could enforce it via.. uh.. communication? :) (not trying to be snarky,
just thinking out loud)

Like, maintainers are the only ones getting push access to the master,
right? So you make sure anyone getting access to push things to the master
has that option used. Anyone not having access to it can only submit pull
requests, and you can easily see in the pull request if it will pollute the
history? It feels about right.
On github you have a setting on a repo, that lets you set a template for
pull request descriptions. In it, you can ask people to verify that they
have rebased. :).