From: Derrick Stolee via GitGitGadget Date: Sat, 02 May 2026 14:15:47 GMT Subject: [PATCH 0/7] pack-objects: integrate --path-walk and some --filter options Message-ID: The 'git pack-objects' command has a '--path-walk' option that uses the path-walk API instead of a typical revision walk to group objects into chunks by path name instead of relying solely on name-hashes to group similar files together. (It also does a second compression pass looking for better deltas after the first pass that is focused within chunks per path.) The '--path-walk' feature was not previously integrated with the '--filter' feature, so a warning would appear and disable the path-walk API when a filter is given. This patch series integrates these together in the following ways: * --filter=blob:none updates the path-walk API options to skip blobs. * --filter=blob:limit= adds a scan to a list of blob objects to remove objects that are too large. * --filter=sparse: adds a scan to the chunks to validate that the paths match the sparse-checkout patterns. In particular, this last check is significantly faster than the previous algorithm because it can check all objects at a given path simultaneously instead of checking all sparse-checkout patterns for each discovered blob object. A subtlety must be added here, in that we must change how we mark an object as "seen" during the path-walk. We may need to add an object to multiple paths and only mark it as "seen" if it indeed matched the sparse-checkout patterns as the path is accepted for emitting to the callback. This adds a new filter that the "seen" objects must also be removed from later chunks to avoid sending the same object as grouped to multiple chunks. There's also a subtle detail here in that the path-walk API also prunes tree paths based on cone-mode sparse-checkouts, to enable 'git backfill --sparse' operating quickly for small sparse-checkout scopes. But the --filter=sparse: feature doesn't prune trees! As a future step, I do plan to recommend that we add a treesparse: setting that does allow us to trim the tree set by cone-mode sparse patterns. At the time that partial clone filters were being created, cone mode sparse-checkout didn't exist and neither did the sparse index. Those features together make a smaller tree set possible, assuming the user never needs to change their scope. This would be a significant change so it is not implemented here, though the git pack-objects integration would be quick after this series completes. Neither the sparse: or hypothetical treesparse: options are or should necessarily be supported by Git servers. It's too expensive to compute dynamically and it doesn't work well with reachability bitmaps. What becomes possible with this change is that it becomes reasonably fast to construct bundles with these filters that can bootstrap a working environment with the full history of all files within a given sparse-checkout scope. Performance Results =================== Since the '--path-walk' option is ignored in today's Git version when a '--filter' is added, the performance matches the behavior without '--path-walk'. For the tables below, you can compare the rows against each other (time and then packfile size) for the mode without and then with '--path-walk' as a representation of "before" and "after". (These tables are repeated in the commit messages as new implementations improve specific rows.) I chose a number of open source repositories of various sizes and shapes: git/git ======= Test HEAD ------------------------------------------------------------------- 5315.2: repack (no filter) 27.73 5315.3: repack size (no filter) 250.6M 5315.4: repack (no filter, --path-walk) 35.19 5315.5: repack size (no filter, --path-walk) 220.1M 5315.6: repack (blob:none) 13.42 5315.7: repack size (blob:none) 137.6M 5315.8: repack (blob:none, --path-walk) 20.98 5315.9: repack size (blob:none, --path-walk) 115.2M 5315.10: repack (sparse:oid) 72.53 5315.11: repack size (sparse:oid) 187.5M 5315.12: repack (sparse:oid, --path-walk) 29.00 5315.13: repack size (sparse:oid, --path-walk) 161.0M nodejs/node =========== Test HEAD -------------------------------------------------------------------- 5315.2: repack (no filter) 75.53 5315.3: repack size (no filter) 0.9G 5315.4: repack (no filter, --path-walk) 80.54 5315.5: repack size (no filter, --path-walk) 885.7M 5315.6: repack (blob:none) 12.65 5315.7: repack size (blob:none) 148.6M 5315.8: repack (blob:none, --path-walk) 17.60 5315.9: repack size (blob:none, --path-walk) 134.6M 5315.10: repack (sparse:oid) 518.84 5315.11: repack size (sparse:oid) 153.4M 5315.12: repack (sparse:oid, --path-walk) 27.99 5315.13: repack size (sparse:oid, --path-walk) 139.4M microsoft/fluentui ================== Test HEAD -------------------------------------------------------------------- 5315.2: repack (no filter) 146.77 5315.3: repack size (no filter) 562.1M 5315.4: repack (no filter, --path-walk) 72.82 5315.5: repack size (no filter, --path-walk) 172.6M 5315.6: repack (blob:none) 4.84 5315.7: repack size (blob:none) 62.7M 5315.8: repack (blob:none, --path-walk) 5.19 5315.9: repack size (blob:none, --path-walk) 59.9M 5315.10: repack (sparse:oid) 59.95 5315.11: repack size (sparse:oid) 85.6M 5315.12: repack (sparse:oid, --path-walk) 15.16 5315.13: repack size (sparse:oid, --path-walk) 72.5M microsoftdocs/azure-devops-docs =============================== Test HEAD --------------------------------------------------------------------- 5315.2: repack (no filter) 4.41 5315.3: repack size (no filter) 1.6G 5315.4: repack (no filter, --path-walk) 6.00 5315.5: repack size (no filter, --path-walk) 1.6G 5315.6: repack (blob:none) 1.35 5315.7: repack size (blob:none) 60.0M 5315.8: repack (blob:none, --path-walk) 1.23 5315.9: repack size (blob:none, --path-walk) 60.0M 5315.10: repack (sparse:oid) 138.24 5315.11: repack size (sparse:oid) 84.4M 5315.12: repack (sparse:oid, --path-walk) 1.86 5315.13: repack size (sparse:oid, --path-walk) 84.4M Performance conclusions ======================= As seen in earlier series around the '--path-walk' feature, the space savings can be valuable but is not always guaranteed. When the space savings doesn't happen, then the time spent is generally slower because of the two-pass mechanism. The microsoftdocs/azure-devops-docs repo demonstrates this case quite clearly. However, even in these cases the 'sparse:' filters are much faster because of the ability to check an entire set of objects against the sparse-checkout patterns only once. Thanks, -Stolee P.S. I've CC'd the folks who were on the original path-walk feature thread [1] [1] https://lore.kernel.org/git/pull.1819.git.1741571455.gitgitgadget@gmail.com/ Derrick Stolee (7): pack-objects: pass --objects with --path-walk t/perf: add pack-objects filter and path-walk benchmark path-walk: support blobless filter backfill: die on incompatible filter options path-walk: support blob size limit filter path-walk: add pl_sparse_trees to control tree pruning pack-objects: support sparse:oid filter with path-walk builtin/backfill.c | 5 + builtin/pack-objects.c | 9 +- path-walk.c | 154 ++++++++++++- path-walk.h | 21 ++ t/helper/test-path-walk.c | 17 +- t/perf/p5315-pack-objects-filter.sh | 129 +++++++++++ t/t5317-pack-objects-filter-objects.sh | 125 ++++++++++ t/t5620-backfill.sh | 8 + t/t6601-path-walk.sh | 306 +++++++++++++++++++++++++ 9 files changed, 758 insertions(+), 16 deletions(-) create mode 100755 t/perf/p5315-pack-objects-filter.sh base-commit: 94f057755b7941b321fd11fec1b2e3ca5313a4e0 Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-2101%2Fderrickstolee%2Fpath-walk-filters-v1 Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-2101/derrickstolee/path-walk-filters-v1 Pull-Request: https://github.com/gitgitgadget/git/pull/2101 -- gitgitgadget