The path-restricted log problem

git log -- path/to/file must, in principle, walk every commit, diff each against its parent, and emit those that touched path/to/file. On large repos this is dominated by tree comparisons. Changed-path Bloom filters (Git 2.27+) accelerate this dramatically by storing, for each commit, a probabilistic set of paths it touched.

How Bloom filters work

A Bloom filter is a small bitmap that returns "definitely not in the set" or "probably in the set" for membership queries. False positives are possible but never false negatives. For a path query, Git first asks the filter; only if the answer is "probably yes" does it perform the expensive tree diff.

Enabling

git commit-graph write --reachable --changed-paths
git config commitGraph.changedPathsVersion 2
git maintenance run --task=commit-graph

Version 2 (Git 2.39+) uses MurmurHash3 with case-insensitive normalization and is the recommended default. Version 1 is backward-compatible.

Verifying

GIT_TRACE2_PERF=1 git log -- src/foo.c 2>&1 | grep -i bloom
git commit-graph verify

The trace shows hit/miss counts. A well-tuned filter has >95% true negatives, meaning Git skips the tree diff for the vast majority of commits.

When it shines

• Long-lived monorepo with many commits that touch only a few files each.
• git log -- src/specific/dir queries.
• git blame on individual files (which uses similar machinery).

When it does not help

• Glob/regex paths the filter cannot exact-match.
• Logs without a -- path argument — the filter only applies to path-restricted walks.
• Very small repos where the walk is already fast.

Memory and disk

Filters add roughly 8-12 bytes per commit to the commit-graph file. For a million-commit repo, expect 10MB extra. Set commitGraph.maxNewFilters (Git 2.34+) to throttle how many filters are computed per write, useful when first generating filters on a large repo.

git config commitGraph.maxNewFilters 50000

Common mistakes

Writing the graph without --changed-paths — filters are not added by default. Mixing v1 and v2 filters across teammates; let git maintenance upgrade them. Expecting filters to help unrestricted log; they do not.

Related

See "The commit-graph: faster reachability queries" and "Background maintenance: git maintenance run/start/stop".