Re: [PATCH 2/2] bloom: enable multiple pathspec bloom keys

From: Lidong Yan <yldhome2d2@gmail.com>
Date: Jul 4, 2025, 12:09 UTC
Message-ID: <1BD174FD-887A-4002-955C-A67E7DBFEFCA@gmail.com>
In-Reply-To: <xmqq1pqyfvb2.fsf@gitster.g>

Junio C Hamano <gitster@pobox.com> writes:

Show 12 quoted lines

> 
> Before concluding so, we may want to double check how Bloom filters
> are built on case insensitive systems, though.  If we normalize the
> string by downcasing before murmuring the string, the resulting
> Bloom filter may have more false positives for those who want to
> (ab)use it to optimize case sensitive queries (without affecting
> correctness), but case insensitive queries would be helped.  I do
> not think we support (or want to support) a repository that spans
> across two filesystems with different case sensitivity, so those who
> worked on our changed-path Bloom filter subsystem may have already
> placed such an optimization, based on the case sensitivity recorded
> in the repository (core.ignorecase).

In bloom.c:get_or_compute_bloom_filter(), the computation of a bloom filter
looks like:
    diff_tree_oid(c’s parent or NULL, &c->object.oid, "", &diffopt);
    diffcore_std(&diffopt);
    struct hashmap path_hashmap;

    for (path : diff_queue_diff) {
        Add all parts of path to path_hashmap;
    }

    for_each(path_hashmap) {
        Add path to filter
    }

All these steps do not check config.ignoreCase, so I believe the Bloom filter we build in the commit graph is case-sensitive.

To demonstrate this assumption—and since I happen to be a Mac user (where config.ignoreCase is true by default)—I ran the following commands under the llvm-project repository:

$ git commit-graph write --split --reachable --changed-paths $ time git log -5 -t -- README.md > /dev/null real 0m0.089s user 0m0.067s sys 0m0.021s $ time git log -5 -t -- ':(icase)README.md' > /dev/null real 0m0.281s user 0m0.239s sys 0m0.041s $ time git log -5 -t -- ‘rEADME.md’ > /dev/null real 0m0.458s user 0m0.394s sys 0m0.061s

And I think it proves that changed-path Bloom filter doesn’t optimize icase pathspec item in case insensitive file system.