I'm mirroring and reworking a large Git repository with git filter-branch (conversion ETA: 20h), and I was wondering how to use --state-branch which is supposed to speed-up later updates, or split a large conversion in several updates.

The documentation is pretty terse, the option can produce weird results (like an identity mapping that breaks all later updates, or calling the expensive tree-filter but discarding the results), wrappers are convoluted, but I got something to work so I'll share :)

The main point is: run the initial script and the later updates in the same configuration, which means the target branch needs to be reset to the upstream branch each time, before it's rewritten again by filter-branch. In other words, don't re-run it on the rewritten branch, nor attempt some complex merge/cherry-pick.

git fetch
git branch --no-track -f myrewrite origin/master
git filter-branch \
  --xxx-filter ... \
  --xxx-filter ... \
  --state-branch refs/heads/filter-branch/myrewrite \
  -d /dev/shm/filter-branch/$$ -f \
  myrewrite

Updates restart from scratch but only take a few seconds to skim through all the already-rewritten commits, and maintain a stable history.

Note that if the process is interrupted, the state-branch isn't modified, so it's not a stop/resume feature. If you want to split a lenghty conversion, you could simulate multiple upstream updates by checking out successive points in history (e.g. per year using $(git rev-list -1 --before='2020-01-01 00:00:00Z')).

--state-branch isn't meant to rewrite in reverse chronological order either, because all commit ids would constantly change. Still, you can rewrite only the recent history for a quick discardable test.

Be cautious when using/deleting rewritten branches, especially during early tests, because Git tends to save them to multiple places which may desync (e.g. .git/refs/heads/, .git/logs/refs/, .git/packed-refs). Also remember to delete the state-branch between different tests. Last, note the unique temporary directory -d to avoid ruining concurrent tests ^_^'