From: Kristofer Karlsson via GitGitGadget Date: Wed, 07 Oct 2026 14:22:55 GMT Subject: [PATCH v3 0/2] fetch: write commit-graph using updated refs only Message-ID: In-Reply-To: When fetch.writeCommitGraph is enabled, the commit-graph is rebuilt from all reachable refs after every fetch. This is unnecessarily expensive on repositories with many refs, since add_ref_to_set() validates each ref against the odb. This series optimizes the commit-graph write by using only the newly updated refs as seeds instead of scanning all refs. Since fetch writes the commit-graph in split mode, the newly fetched history is added as a new layer on top of the existing chain. A three-mode enum (REACHABLE / TIPS / SKIP) makes the policy explicit: * No-op fetch: skip the commit-graph write entirely * Updated refs + existing graph: write incrementally from updated tips only * No existing graph or multi-remote fetch: fall back to full reachable scan Patch 1 adds a commit-info subcommand to test-tool read-graph for verifying graph contents in tests. Patch 2 implements the optimization in builtin/fetch.c with tests covering the incremental, unrelated-commit, no-op, fallback, and shallow-rejected cases. Benchmark on a synthetic setup: git.git with 200K extra packed refs (~206K total), a local file:// remote, an existing split commit-graph and a warm page cache. Times are the median of 9 runs of the trace2 region fetch/write-commit-graph: scenario before after no-op fetch 380 ms (skipped) 1 ref updated 357 ms 9.3 ms 10 refs updated 359 ms 8.9 ms Changes since v2: * Trim the commit message: inline the commit references, keep the explanation of why the incremental write relies on split mode, and drop the paragraphs that only restated the diff (collecting the tips, auto-followed tags, skipping shallow-rejected refs). * Reword the comment on skipping shallow-rejected refs without the reference to store_updated_refs(). * Simplify the t5537 test by using "test_commit -C ... --no-tag" instead of subshells. Changes since v1: * Explain in the commit message that fetch writes in split mode, so the new tips are added as a new layer on top of the existing chain, and that the incremental path relies on this (a non-split write would replace the graph with only the closure of the seeds). * Extend the incremental test to check that a local-only commit, which was in the graph before the fetch but is not reachable from the fetched tips, is still in the graph afterwards. * Add benchmark numbers to the commit message. * Add a comment explaining why shallow-rejected refs are skipped when collecting the updated tips (like store_updated_refs(), since their history is incomplete), and mention it in the commit message. * Add a test in t5537 for a fetch with fetch.writeCommitGraph where a ref is rejected because it would require changes to .git/shallow. Without the check, the commit-graph write dies on the missing parent. Kristofer Karlsson (2): test-tool read-graph: add commit-info subcommand fetch: write commit-graph using updated refs only builtin/fetch.c | 69 +++++++++++++++++++++++++++++++++----- commit-graph.c | 2 +- commit-graph.h | 1 + t/helper/test-read-graph.c | 23 ++++++++++++- t/t5510-fetch.sh | 59 ++++++++++++++++++++++++++++++++ t/t5537-fetch-shallow.sh | 22 ++++++++++++ 6 files changed, 165 insertions(+), 11 deletions(-) base-commit: 0f8e75abebff0877cae681a3d5ff31ac47f54220 Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-2239%2Fspkrka%2Fkrka%2Fincremental-commit-graph-v3 Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-2239/spkrka/krka/incremental-commit-graph-v3 Pull-Request: https://github.com/gitgitgadget/git/pull/2239 Range-diff vs v2: 1: 36cc3ff3b3 = 1: 36cc3ff3b3 test-tool read-graph: add commit-info subcommand 2: 7507354cc9 ! 2: 01da9857bc fetch: write commit-graph using updated refs only @@ Metadata ## Commit message ## fetch: write commit-graph using updated refs only - When fetch.writeCommitGraph was introduced in - - 50f26bd035 (fetch: add fetch.writeCommitGraph config - setting, 2019-09-02), - - the stated goal was to stay updated with the latest commits after - fetching new objects. The implementation used - write_commit_graph_reachable() because it was the only API available, - but two things have changed since then: + When fetch.writeCommitGraph was introduced in 50f26bd035 (fetch: add + fetch.writeCommitGraph config setting, 2019-09-02), the stated goal + was to stay updated with the latest commits after fetching new + objects. The implementation used write_commit_graph_reachable() + because it was the only API available, but two things have changed + since then: 1. write_commit_graph() was added, and it accepts an explicit set of commits as seeds, enabling more targeted commit-graph updates. 2. The ref-scanning callback add_ref_to_set() became more expensive - in - 630cd5194e (commit-graph.c: peel refs in 'add_ref_to_set', - 2020-07-22) - when it started to validate the refs against the odb + in 630cd5194e (commit-graph.c: peel refs in 'add_ref_to_set', + 2020-07-22) when it started to validate the refs against the odb for correctness. On a repository with many refs, this makes the full reachable scan unnecessarily costly for a targeted fetch. @@ Commit message (since that would require propagating the set of refs across process boundaries). - Since do_fetch() already knows which refs were updated, collect them - into an oidset and then pass them directly to write_commit_graph(). - fetch always writes the commit-graph in split mode, so this adds a - new layer on top of the existing chain rather than replacing it: - close_reachable() walks from the updated tips and stops at commits - already present in the graph, so the new layer only contains the - newly fetched history, and commits covered by the existing layers - remain covered. This relies on split mode; a non-split write would - replace the graph with just the closure of the seeds. - - The reachability closure also covers auto-followed tags, since their - targets are reachable from the fetched tips that caused them to be - auto-followed. - - Refs that are rejected because they would require changes to - .git/shallow are skipped, just like store_updated_refs() does. Their - objects are received but their history is incomplete, so walking from - them would make the commit-graph write fail. + This relies on the commit-graph write being additive, keeping the + commits that are already in the graph. fetch already operates in + this mode (COMMIT_GRAPH_WRITE_SPLIT) and now that becomes + required for correctness. Without that mode, the write would + replace the commit-graph and lose other commits. After fetch_one() returns, call prepare_commit_graph() (which is made non-static by this commit) to determine the graph-write mode: @@ builtin/fetch.c: out: + for (rm = ref_map; rm; rm = rm->next) { + struct commit *commit; + /* -+ * Like store_updated_refs(), skip shallow-rejected refs: -+ * they are not stored, and their history is incomplete. ++ * Shallow-rejected refs are not stored and their history ++ * is incomplete, so skip them. + */ + if (rm->status == REF_STATUS_REJECT_SHALLOW) + continue; @@ t/t5537-fetch-shallow.sh: test_expect_success 'fetch that requires changes in .g +test_expect_success 'fetch.writeCommitGraph skips refs that require changes in .git/shallow' ' + git clone --no-local --depth=2 .git shallow-graph && -+ ( -+ cd shallow-graph && -+ git checkout --orphan no-shallow && -+ commit no-shallow -+ ) && ++ git -C shallow-graph checkout --orphan no-shallow && ++ test_commit -C shallow-graph --no-tag no-shallow && + git init notshallow-graph && + git -C notshallow-graph -c fetch.writeCommitGraph=true \ + fetch ../shallow-graph/.git "refs/heads/*:refs/remotes/shallow/*" && -+ ( -+ cd shallow-graph && -+ commit no-shallow-2 -+ ) && ++ test_commit -C shallow-graph --no-tag no-shallow-2 && + rejected=$(git -C shallow-graph rev-parse main) && + ( + cd notshallow-graph && -- gitgitgadget