locking binary files

Locking is useful to make sure that binary files (office docs, images, ...) don't get into a merge state. (If you think it's not a big deal, you have never manually merged independent changes to an ODT or something!)

When git is used in a truly distributed fashion, locking is impossible. However, in most corporate setups, there is a single central server acting as the canonical source of truth and collaboration point for all developers. In this situation it should be possible to at least prevent commits from being pushed that contains changes to files locked by someone else.

The two "lock" programs (one a command that a user uses, and one a VREF that the admin adds to a repo's access rules) together attempt to achieve this.

Of course, locking by itself is not quite enough. You may still get into merge situations if you make changes in branches. For best results you should actually keep all the binary files in their own branch, separate from the ones containing source code.

problem description

Our users are alice, bob, and carol. Our repo is foo. It has some "odt" files in the "doc/" directory. We want to make sure these odt files never get into a "merge" situation.

admin/setup

First, someone with shell access to the server must add 'lock' to the ENABLE list in the rc file.

Next, the gitolite.conf file should have something like this:

repo foo
    ...other rules...
    -   VREF/lock      =   @all

However, see below for the difference between "RW" and "RW+" from the point of view of this feature and adjust permissions accordingly.

user view

Here's a summary:

  • Any user with "W" permissions to any branch in the repo can "lock" any file. Once locked, no other user can push changes to that file, in any branch, until it is unlocked.
  • Any user with "+" permissions to any branch in the repo can "break" a lock held by someone else if needed.

For best results, everyone on the team should:

  • Switch to the branch containing the binary files when wanting to make a change.
  • Run 'git pull' or eqvt, then lock the binary file(s) before editing them.
  • Finish the editing task as quickly as possible, then commit, push, and unlock the file(s) so others are not needlessly blocked.
  • Understand that breaking a lock require additional, (out of band) communication. It is upto the team's policies what that entails.

detailed example

Alice declares her intent to work on "d1.odt":

$ git pull
$ ssh git@host lock -l foo doc/d1.odt

Similarly Bob starts on "d2.odt"

$ git pull
$ ssh git@host lock -l foo doc/d2.odt

Carol makes some changes to d2.odt (without attempting to lock the file or checking to see if it is already locked) and pushes:

$ ooffice doc/d2.odt
$ git add doc/d2.odt
$ git commit -m 'added footnotes to d2 in klingon'
$ git push
<...normal push progress output...>
remote: FATAL: W VREF/lock testing carol DENIED by VREF/lock
remote: 'doc/d2.odt' locked by 'bob'
remote: error: hook declined to update refs/heads/master
To u2:testing
 ! [remote rejected] master -> master (hook declined)
error: failed to push some refs to 'carol:foo'

Carol backs out her changes, but saves them away for a "manual merge" later.

git reset HEAD^
git stash save 'klingon changes to d2.odt saved for possible manual merge later'

Note that this still represents wasted work in some sense, because Carol would have to somehow re-apply the same changes to the new version of d2.odt after pulling it down. This is because she did not lock the file before making changes on her local repo. Educating users in doing this is important if this scheme is to help you.

She now decides to work on "d1.odt". However, she has learned her lesson and decides to follow the protocol described above:

$ git pull
$ ssh git@host lock -l foo doc/d1.odt
FATAL: 'doc/d1.odt' locked by 'alice' since Sun May 27 17:59:59 2012

Oh damn; can't work on that either.

Carol now decides to see what else there may be. Instead of checking each file to see if she can lock it, she starts with a list of what is already locked:

$ ssh git@host lock -ls foo

# locks held:

alice   doc/d1.odt      (Sun May 27 17:59:59 2012)
bob     doc/d2.odt      (Sun May 27 18:00:06 2012)

# locks broken:

Aha, looks like only d1 and d2 are locked. She picks d3.odt to work on. This time, she starts by locking it:

$ ssh git@host lock -l foo doc/d3.odt
$ ooffice doc/d3.odt
<...etc...>

Meanwhile, in a parallel universe where d3.odt doesn't exist, and Alice has gone on vacation while keeping d1.odt locked, Carol breaks the lock. Carol can do this because she has RW+ permissions for the repository itself.

However, protocol in this team requires that she get email approval from the team lead before doing this and that Alice be in CC in those emails, so she does that first, and then she breaks the lock:

$ git pull
$ ssh git@host lock --break foo doc/d1.odt

She then locks d1.odt for herself:

$ ssh git@host lock -l foo doc/d1.odt

When Alice comes back, she can tell who broke her lock and when:

$ ssh git@host lock -ls foo

# locks held:

carol   doc/d1.odt      (Sun May 27 18:17:29 2012)
bob     doc/d2.odt      (Sun May 27 18:00:06 2012)

# locks broken:

carol   doc/d1.odt      (Sun May 27 18:17:03 2012)      (locked by alice at Sun May 27 17:59:59 2012)