"Johannes Schindelin via GitGitGadget" [off-list ref]
writes:
From: Johannes Schindelin <redacted>
Whenever a user runs `git reflog expire --stale-fix`, the most likely
reason is that their repository is at least _somewhat_ corrupt. Which
means that it is more than just possible that some objects are missing.
If that is the case, that can currently let the command abort through
the phase where it tries to mark all reachable objects.
Instead of adding insult to injury, let's be gentle and continue as best
as we can in such a scenario, simply by ignoring the missing objects and
moving on.
Signed-off-by: Johannes Schindelin <redacted>
---
I am of two minds.
I appreciate an effort of making it looser and less likely to get in
trouble when running in a corrupt repository, but --stale-fix is a
bit special and should probably be removed.
The only reason the option was added was because we needed to have
an easy way to recover from a specific kind of reflog corruption
that would have resulted by a (then known) bug in "git reflog" in
the released version of Git that came immediately before the version
of Git that added the "fix" option, while the root cause of the
corruption got fixed.
Back when 1389d9dd (reflog expire --fix-stale, 2007-01-06) was
written, it was very useful to have a way to recover from the
corruption likely to have happened with the version of Git that came
before it. But it no longer is relevant after this many years.
There may be other ways to break the reflog entries, but --stale-fix
was never designed to deal with anything but a specific way the
reflog gets corrupted (namely, by the old version of Git that
corrupted reflog in a specific way), so keeping it would not be very
useful.
On Wed, Feb 10, 2021 at 12:30:27PM -0800, Junio C Hamano wrote:
I appreciate an effort of making it looser and less likely to get in
trouble when running in a corrupt repository, but --stale-fix is a
bit special and should probably be removed.
The only reason the option was added was because we needed to have
an easy way to recover from a specific kind of reflog corruption
that would have resulted by a (then known) bug in "git reflog" in
the released version of Git that came immediately before the version
of Git that added the "fix" option, while the root cause of the
corruption got fixed.
Back when 1389d9dd (reflog expire --fix-stale, 2007-01-06) was
written, it was very useful to have a way to recover from the
corruption likely to have happened with the version of Git that came
before it. But it no longer is relevant after this many years.
There may be other ways to break the reflog entries, but --stale-fix
was never designed to deal with anything but a specific way the
reflog gets corrupted (namely, by the old version of Git that
corrupted reflog in a specific way), so keeping it would not be very
useful.
FWIW, I have used --stale-fix for cases where a repo's reflog was
"corrupted" by its alternate pruning objects.
E.g., if you do something like this:
git clone -s orig.git new.git
you're now at risk of "git gc" in orig.git corrupting new.git, because
its reachability check doesn't know anything about those refs. You can
mitigate that by keeping a copy of new.git's refs in orig.git. Something
like:
git -C orig.git fetch ../new.git refs/*:refs/preserve/*
git -C orig.git gc
But that doesn't know about reflogs at all! It will still corrupt them
(but only sometimes; depending how often you do that fetch, you might
catch the same values in orig.git's reflog).
Possibly the correct answer here is "turn off reflogs in new.git", but
they are sometimes useful, and things _mostly_ work (for history that
doesn't rewind, or where the rewound bits are specific to new.git). So
it's useful to be able to run something like "reflog expire --stale-fix"
to clear out the occasional problem.
(A careful reader will note that objects mentioned only in the index
have a similar problem; but again, those tend to be local to new.git,
and don't exist at all in a server context).
-Peff
Hi,
On Thu, 11 Feb 2021, Jeff King wrote:
On Wed, Feb 10, 2021 at 12:30:27PM -0800, Junio C Hamano wrote:
quoted
I appreciate an effort of making it looser and less likely to get in
trouble when running in a corrupt repository, but --stale-fix is a
bit special and should probably be removed.
The only reason the option was added was because we needed to have
an easy way to recover from a specific kind of reflog corruption
that would have resulted by a (then known) bug in "git reflog" in
the released version of Git that came immediately before the version
of Git that added the "fix" option, while the root cause of the
corruption got fixed.
Back when 1389d9dd (reflog expire --fix-stale, 2007-01-06) was
written, it was very useful to have a way to recover from the
corruption likely to have happened with the version of Git that came
before it. But it no longer is relevant after this many years.
There may be other ways to break the reflog entries, but --stale-fix
was never designed to deal with anything but a specific way the
reflog gets corrupted (namely, by the old version of Git that
corrupted reflog in a specific way), so keeping it would not be very
useful.
Thank you for the original historical context.
FWIW, I have used --stale-fix for cases where a repo's reflog was
"corrupted" by its alternate pruning objects.
E.g., if you do something like this:
git clone -s orig.git new.git
you're now at risk of "git gc" in orig.git corrupting new.git, because
its reachability check doesn't know anything about those refs. You can
mitigate that by keeping a copy of new.git's refs in orig.git. Something
like:
git -C orig.git fetch ../new.git refs/*:refs/preserve/*
git -C orig.git gc
But that doesn't know about reflogs at all! It will still corrupt them
(but only sometimes; depending how often you do that fetch, you might
catch the same values in orig.git's reflog).
Possibly the correct answer here is "turn off reflogs in new.git", but
they are sometimes useful, and things _mostly_ work (for history that
doesn't rewind, or where the rewound bits are specific to new.git). So
it's useful to be able to run something like "reflog expire --stale-fix"
to clear out the occasional problem.
(A careful reader will note that objects mentioned only in the index
have a similar problem; but again, those tend to be local to new.git,
and don't exist at all in a server context).
I want to add the experience with that half year during which `git gc`
with worktrees was broken, as it would happily ignore the reflogs of the
worktree-specific `HEAD`s, all except the one from the worktree from which
`git gc` was run. That was some fun time, and `--stale-fix` was a life
saver.
With this _additional_ historical context, I would deem it wise to keep
`--stale-fix` around.
Ciao,
Dscho