Wednesday, August 14, 2013

Delaying Git 1.8.4 by a week

It appears that we need to revert two topics that cause regressions before the upcoming 1.8.4 release.

  • There is a corner case bug in git stash.  Suppose you have a path that is a regular file (or a symbolic link) in the committed state. You change it to a directory in your working tree, and have various new files in it. Some of them may be tracked, while others may not be. You issue git stash. The command needs to match the path to the committed state, hence it needs to remove the directory to resurrect the path. The new files in the directory you have git added will be in the stash so they are OK, but what happens to the untracked ones? They are killed. The same issue exists if you turned a tracked directory into a file and run the command without first running git add.
    An attempted fix was to ask
    git ls-files --killed to see if such a path exists that will be lost, but it turns out that this makes the command unusably slow in certain directories with very many untracked files.
  • There was an attempt to save typing four capital letters "H", "E", "A" and "D" by instead allowing you to type "@", e.g. git log @. The idea may have been a good one, but the change was executed poorly and incorrectly triggered when it shouldn't (e.g. having a branch whose name is @/foo made it into HEAD/foo or something insane).

Because we have already passed -rc3, I'd feel safer to add another rc week before the final. Updated Git Calendar is here.

Both of these changes meant well, and because we are not reverting them due to design mistakes (i.e. we are not saying that "we do not ever want to have such a feature or fix in our system"), hopefully these can be redone properly after the upcoming release is done.

Some leftover bits (I'll add more to this list later).

  • [DONE] Find out where ls-files --killed is unnecessarily wasting time, and fix it. This is a prerequisite to resurrect the stash corner case fix.
    Cf. $gmane/
  • Refactor run_hook() interface to be truly reusable by codepath other than git commit, resurrecting a "how about this" patch sent in the past.
    Cf. $gmane/192806, $gmane/212284
  • [IN PROGRESS] Extend the upload-pack protocol to tell what symbolic ref points at which other ref by resurrecting the idea outlined in 2008.
    Cf. $gmane/102039
  • [IN PROGRESS] Rethink how name-hash keeps track of names of directories and actual files to help case insensitive filesystems. Since 2092678c (name-hash.c: fix endless loop with core.ignorecase=true, 2013-02-28), there appears to be no reason why a directory name has to be registered to the hash with a trailing slash, which is the root cause why directory_exists_in_index_icase() reads past the end of the buffer.
    Cf. $gmane/232822
  • [DONE] Look into cvsserver permission bits regression between 1.8.1 and 1.8.3.
    Cf. $gmane/234476
  • Look into pathspec-limited revision traversal regression between 1.8.3 and 1.8.4.
    Cf. $gmane/234462
  • Checking out a branch X that does not have directory D (or worse, has a file D), while you are in the directory D, may want to fail.
    Cf. $gmane/234905
  • Allow extra options to "ssh" invocation made from connect.c, in a way that (ideally) does not break backward compatibility.
    Cf. $gmane/234624
  • Perhaps add a --post-service-hook to the git-daemon that can be used after a service finishes? The exit status from the service process means totally different thing from what the user of service perceives because the former has to say "successfully told the requester that the request is denied", it may not be such a useful mechanism as one na├»vely would expect, though.
    Cf. $gmane/
  • git checkout $commit -- somedir should remove somedir/file that is not in $commit but is in the original index.
    Cf. $gmane/