[[!comment format=mdwn
 username="joey"
 subject="""comment 6"""
 date="2015-02-09T21:59:31Z"
 content="""
[[forum/repair_stuck_on_ls-tree_command]] is another case of that, and I got ahold of
that repository for analysis.

In that case, there was indeed an inverse pyramid effect where each commit
added one more " (recovery from race)" to its parent commit.

The code can clearly loop 
if it keeps detecting a race and somehow fails to recover from it. Leading
to a whole stack of commits with progressively longer messages.
I don't see any way to get just one commit with a long message, which
comment #1 seems to say happened.

Apparently loops for a while and then succeeds in recovering from
the race, since it then stops looping.

I have added additional debug info to the commit message, in hopes of detecting
what might be going wrong that causes it to loop.

Seems to me there are two possibilities.

One is that something else is continually changing the git-annex
branch in a way that keeps triggering the race. If so, it might make
sense for git-annex to do a brief random sleep (a few hundredths of a
second) before recovering, to let whatever it is quiet down. I've done so.

The other is some kind of bug where it detects a race when none
occurred. Perhaps it's misparsing the commit or git cat-file is failing
to output it, and so it's not finding the expected parent refs, for example.
But in that case, why would it detect a race for many commits
in a row, and then eventually not detect a race anymore?

Also, I've made these messages no longer stack up even if it does go into a
loop, which will at least help with the object size bloat, though not with the
number of commits bloat.
"""]]