Git Data Recovery

Lost commits survive until gc prunes (default: 14 days dangling, 90 days reflog-reachable). Recovery is a race against gc.

Path 1: Reflog (for commits you were on)

git reflog                     # HEAD history
git reflog show <branch>       # per-branch
git log -g --oneline           # same as log

git branch rescue <sha>
git reset --hard HEAD@{5}
git cherry-pick HEAD@{1}

Reflog is local, per-ref. A fresh clone has none.

Path 2: fsck (when reflog is gone)

git fsck --full --no-reflogs --unreachable --lost-found
# dangling commit abc...

--lost-found writes objects to .git/lost-found/{commit,other}/.

Inspect candidates:

for sha in $(git fsck --no-reflog --unreachable | awk '/commit/ {print $3}'); do
  git log -1 --format='%h %ci %s' "$sha"
done
git branch rescue <sha>

Path 3: One File from a Commit

git show <sha>:path > recovered
git checkout <sha> -- path

Protect Before Risky Ops

git tag rescue/$(date +%s) HEAD
git update-ref refs/rescue/before-rebase HEAD

Any ref under refs/ keeps objects reachable indefinitely.

gc / Prune

• Auto-triggers at ~7,000 loose objects or 50+ packfiles.
• git gc --prune=now immediately deletes unreachable objects — do not run during recovery.
• git reflog expire --expire-unreachable=now --all && git gc --prune=now destroys recovery paths.

Find Large Objects

git rev-list --objects --all \
  | git cat-file --batch-check='%(objectname) %(objecttype) %(objectsize) %(rest)' \
  | awk '$2=="blob"{print $3,$4}' | sort -n | tail -20

Purge: git filter-repo --path <file> --invert-paths.

Pitfalls

• Debug on the original working copy; a fresh clone has no reflog.
• Don't "clean up" before verifying recovery — many one-liners silently destroy it.