Thread (10 messages) 10 messages, 4 authors, 2018-12-11

Re: Retrieving a file in git that was deleted and committed

From: Elijah Newren <hidden>
Date: 2018-12-10 21:33:33

On Thu, Dec 6, 2018 at 11:48 PM Jeff King [off-list ref] wrote:
On Thu, Dec 06, 2018 at 11:07:00PM -0800, biswaranjan panda wrote:
quoted
Thanks! Strangely git log --follow does work.
I suspect it would work even without --follow. When you limit a log
traversal with a pathspec, like:

  git log foo

that is not about following some continuous stream of content, but
rather just applying that pathspec to the diff of each commit, and
pruning ones where it did not change. So even if there are gaps where
the file did not exist, we continue to apply the pathspec to the older
commits.

Tools like git-blame will _not_ work, though, as they really are trying
to track the content as they walk back through history. And Once all of
the content seems to appear from nowhere in your new commit, that seems
like a dead end.

In theory there could be some machine-readable annotation in the commit
object (or in a note created after the fact) to say "even though 'foo'
is a new file here, it came from $commit:foo".  And then git-blame could
keep following the content there. But such a feature does not yet exist.

-Peff
Hmm...sure, if the file is deleted on the only relevant branch through
history...but what if there were another branch where it weren't
deleted?  What does git blame do then?

In other words, do NOT restore the file as biswaranjan suggested, but
instead restore it this way[1]:

git checkout -b keep-foo $REVISION_BEFORE_FOO_DELETED
git commit --allow-empty -m "We want to keep foo"
git checkout A
git merge --no-commit keep-foo
git checkout keep-foo -- foo.txt
git commit


Now, when you run
  git blame foo.txt

blame should notice that foo.txt didn't exist in the first parent
history on A, so it won't bother walking it to find that at some point
foo.txt did exist there.  Instead, it'll walk down the second parent
and follow its history, where it should keep walking back and show all
the old changes...right?  Or did I mess up my testcase and
misunderstand something somehow?


Elijah

[1] Sidenote: it seems like those commands I gave should have simplified down to

git merge --no-commit --no-ff $REVISION_BEFORE_FOO_DELETED
git checkout $REVISION_BEFORE_FOO_DELETED -- foo.txt
git commit

But that seems to error out with "Already up-to-date"...even when
testing with older versions of git, so it wasn't any of my changes.
Not sure why the --no-ff flag doesn't work.
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help