On Sun, Sep 23, 2018 at 6:08 AM Lars Schneider <larsxschneider@xxxxxxxxx> wrote: > > Hi, > > I recently had to purge files from large Git repos (many files, many commits). > The usual recommendation is to use `git filter-branch --index-filter` to purge > files. However, this is *very* slow for large repos (e.g. it takes 45min to > remove the `builtin` directory from git core). I realized that I can remove > files *way* faster by exporting the repo, removing the file references, > and then importing the repo (see Perl script below, it takes ~30sec to remove > the `builtin` directory from git core). Do you see any problem with this > approach? It looks like others have pointed you at other tools, and you're already shifting to that route. But I think it's a useful question to answer more generally, so for those that are really curious... The basic approach is fine, though if you try to extend it much you can run into a few possible edge/corner cases (more on that below). I've been using this basic approach for years and even created a mini-python library[1] designed specifically to allow people to create "fast-filters", used as git fast-export <options> | your-fast-filter | git fast-import <options> But that library didn't really take off; even I have rarely used it, often opting for filter-branch despite its horrible performance or a simple fast-export | long-sed-command | fast-import (with some extra pre-checking to make sure the sed wouldn't unintentionally munge other data). BFG is great, as long as you're only interested in removing a few big items, but otherwise doesn't seem very useful (to be fair, it's very upfront about only wanting to solve that problem). Recently, due to continuing questions on filter-branch and folks still getting confused with it, I looked at existing tools, decided I didn't think any quite fit, and started looking into converting git_fast_filter into a filter-branch-like tool instead of just a libary. Found some bugs and missing features in fast-export along the way (and have some patches I still need to send in). But I kind of got stuck -- if the tool is in python, will that limit adoption too much? It'd be kind of nice to have this tool in core git. But I kind of like leaving open the possibility of using it as a tool _or_ as a library, the latter for the special cases where case-specific programmatic filtering is needed. But a developer-convenience library makes almost no sense unless in a higher level language, such as python. I'm still trying to make up my mind about what I want (and what others might want), and have been kind of blocking on that. (If others have opinions, I'm all ears.) Anyway, the edge/corner cases you can watch out for: - Signed tags are a problem; you may need to specify --signed-tags=strip to fast-export - References to other commits in your commit messages will now be incorrect. I think a good tool should either default to rewriting commit ids in commit messages or at least have an option to do so (BFG does this; filter-branch doesn't; fast-export format makes it really hard for a filter based on it to do so) - If the paths you remove are the only paths modified in a commit, the commit can become empty. If you're only filtering a few paths out, this might be nothing more than a minor inconvenience for you. However, if you're trying to prune directories (and perhaps several toplevel ones), then it can be extremely annoying to have a new history with the vast majority of all commits being empty. (filter-branch has an option for this; BFG does not; tools based on fast-export output can do it with sufficient effort). - If you start pruning empty commits, you have to worry about rewriting branches and tags to remaining parents. This _might_ happen for free depending on your history's structure and the fast-export stream, but to be correct in general you will have to specify the new commit for some branches or tags. - If you start pruning empty commits, you have to decide whether to allow pruning of merge commits. Your first reaction might be to not allow it, but if one parent and its entire history are all pruned, then transforming the merge commit to a normal commit and then considering whether it is empty and allowing it to be pruned is much better. - If you start pruning empty commits, you also have to worry about history topology changing, beyond the all-ancestors-empty case above. For example, the last non-empty commit in the ancestry of a merge on both sides may be the same commit, making the merge-commit have the same parent twice. Should the duplicate parent be de-duped, transforming the commit into a normal non-merge commit? (I'd say yes -- this commit is likely to be empty and prunable once you do so, but I'm not sure everyone would agree with converting a merge commit to a non-merge.) Similarly, what if the rewritten parents of a merge have one parent that is the direct ancestor of another? Can the extra unnecessary parent be removed as a parent? (And again, such a commit is likely to become empty and be prunable itself.) - If you try to avoid the extra work involved with pruning empty commits by passing path-specifiers as rev-list-args to fast-export, and use the --tag-of-filtered-object=rewrite option if needed, then depending on the topology you can hit any of three bugs: an outright die() (despite the --tag-of-filtered-object=rewrite), a branch being reset to a non-existent mark (causing fast-import to die), or find that a ref which you explicitly requested to be part of the export is silently omitted from the stream. (granted, these aren't fundamental issues; they're just bugs in fast-export that I seem to have been the first to find.) - filter-branch has a nice ability to rewrite only the last few commits using a range specifier like HEAD~10..HEAD. Trying the same with fast-export will get you a history with only 10 commits, the first of which squashes all early history together. Trying to duplicate the filter-branch behavior can be done, but it requires multiple exports with different args and usage of --export-marks and --import-marks; it's cumbersome and somewhat non-obvious. - some filters are difficult; e.g. if you want to mimick filter-branch's --parent-filter, or BFG's --strip-blobs-with-ids, you run into the issue that the fast-export stream doesn't provide the original sha1sums for commits or blobs and there's no easy way for you to associate it with the given mark. Those are the ones I know about. It's possible that there are others. Hope that helps or is at least interesting. Elijah [1] https://public-inbox.org/git/51419b2c0904072035u1182b507o836a67ac308d32b9@xxxxxxxxxxxxxx/
