git-filter-repo output isn't deterministic

[barneygale]

First, thanks very much for git-filter-repo - it's been a massive help.

The docs say:

Running the same git-filter-repo commands, even if identical, can result in them getting new hashes for commits that are different than your new hashes, and you’ll end up with a mess involving two or more copies of every commit.

Would you mind expanding on what causes these differing results? If it's something I'm able to fix I'll try to put a PR together.

For context: I'm trying to co-ordinate a git rewrite across two organisations that share a git repository. Each organisation has its own primary git remote, and only a subset of branches are shared, so no single party has the complete history. I had hoped that by matching python / git / git-filter-repo versions we'd be able to rewrite in a "compatible" way from both sides, but unfortunately we're hitting the non-determinism that (I realise too late!) is noted in the docs.

[nfogh]

I think I have hit the same snag :(

One thing I have noticed, is that the fixups of git SHAs in commit messages can be different depending on when the source repo was cloned.

In my case, I have two clones of the same repo. One cloned a few months earlier than the other. One would expect that running the same git-filter-repo command on the two repos would at least give the same history in the resulting repos (up until the "newer" repo had new commits). But it does not, as git-filter-repo sometimes fails to fixup some git SHAs in commit messages.

[byronluk]

I just ran into this as well. Not sure exactly why. I have the same repo, with the same commit history, just initialized in two different ways. One is a full clone and another is a minimal initialization with only one commit and it's lineage. Running the same filtering through these two identical repos (other than adjacent refs) resulted in different commit SHAs.

[nfogh]

The most reliable way I have found is to add flags to avoid rewriting git SHAs in commit messages, and not purging empty or degenerate commits.