On Mon, Aug 17, 2020 at 02:13:18AM +0200, Jakub Narębski wrote: > "Abhishek Kumar via GitGitGadget" <gitgitgadget@xxxxxxxxx> writes: > > > This patch series implements the corrected commit date offsets as generation > > number v2, along with other pre-requisites. > > I'm not sure if this level of detail is required in the cover letter for > the series, but generation number v2 is corrected commit date; corrected > commit date offsets is how we store this value in the commit-graph file. > > > > > Git uses topological levels in the commit-graph file for commit-graph > > traversal operations like git log --graph. Unfortunately, using topological > > levels can result in a worse performance than without them when compared > > with committer date as a heuristics. For example, git merge-base v4.8 v4.9 > > on the Linux repository walks 635,579 commits using topological levels and > > walks 167,468 using committer date. > > I would say "committer date heuristics" instead of just "committer > date", to be more exact. > > Is this data generated using https://github.com/derrickstolee/gen-test > scripts? > Yes, it is. > > > > Thus, the need for generation number v2 was born. New generation number > > needed to provide good performance, increment updates, and backward > > compatibility. Due to an unfortunate problem, we also needed a way to > > distinguish between the old and new generation number without incrementing > > graph version. > > It would be nice to have reference email (or other place with details) > for "unfortunate problem". > Will add. > > > > Various candidates were examined (https://github.com/derrickstolee/gen-test, > > https://github.com/abhishekkumar2718/git/pull/1). The proposed generation > > number v2, Corrected Commit Date with Mononotically Increasing Offsets > > performed much worse than committer date (506,577 vs. 167,468 commits walked > > for git merge-base v4.8 v4.9) and was dropped. > > > > Using Generation Data chunk (GDAT) relieves the requirement of backward > > compatibility as we would continue to store topological levels in Commit > > Data (CDAT) chunk. Thus, Corrected Commit Date was chosen as generation > > number v2. > > This is a bit of simplification, but good enough for a cover letter. > > To be more exact, from various candidates the Corrected Commit Date was > chosen. Then it turned out that old Git crashes on changed commit-graph > format version value, so if the generation number v2 was to replace v1 > it needed to be backward-compatibile: hence the idea of Corrected Commit > Date with Monotonically Increasing Offsets. But with GDAT chunk to > store generation number v2 (and for the time being leaving generation > number v1, i.e. Topological Levels, in CDAT), we are no longer > constrained by the requirement of backward-compatibility to make old Git > work with commit-graph file created by new Git. So we could go back to > Corrected Commit Date, and as you wrote above the backward-compatibile > variant performs worse. > > > The Corrected Commit Date is defined as: > > > > For a commit C, let its corrected commit date be the maximum of the commit > > date of C and the corrected commit dates of its parents. > > Actually it needs to be "corrected commit dates of its parents plus 1" > to fulfill the reachability condition for a generation number for a > commit: > > A can reach B => gen(A) < gen(B) > > Of course it can be computed in simpler way, because > > max_P (gen(P) + 1) == max_P (gen(P)) + 1 > > > > Then corrected > > commit date offset is the difference between corrected commit date of C and > > commit date of C. > > All right. > > > > > We will introduce an additional commit-graph chunk, Generation Data chunk, > > and store corrected commit date offsets in GDAT chunk while storing > > topological levels in CDAT chunk. The old versions of Git would ignore GDAT > > chunk, using topological levels from CDAT chunk. In contrast, new versions > > of Git would use corrected commit dates, falling back to topological level > > if the generation data chunk is absent in the commit-graph file. > > All right. > > However I think the cover letter should also describe what should happen > in a mixed version environment (for example new Git on command line, > copy of old Git used by GUI client), and in particular what should > happen in a mixed-chain case - both for reading and for writing the > commit-graph file. > Yes, definitely. Will add > For *writing*: because old Git would create commit-graph layers without > the GDAT chunk, to simplify the behavior and make easy to reason about > commit-graph data (the situation should be not that common, and > transient -- it should get more rare as the time goes), we want the > following behavior from new Git: > > - If top layer contains the GDAT chunk, or we are rewriting commit-graph > file (--split=replace), or we are merging layers and there are no > layers without GDAT chunk below set of layers that are merged, then > > write commit-graph file or commit-graph layer with GDAT chunk, > > otherwise > > write commit-graph layer without GDAT chunk. > > This means that there are commit-graph layers without GDAT chunk if > and only if the top layer is also without GDAT chunk. > > > For *reading* we want to use generation number v2 (corrected commit > date) if possible, and fall back to generation number v1 (topological > levels). > > - If the top layer contains the GDAT chunk (or maybe even if the topmost > layer that involves all commits in question, not necessarily the top > layer in the full commit-graph chain), then use generation number v2 > The current implementation checks the entire chain for GDAT, rather than just the topmost layer as we cannot assert that `g` would be the topmost layer of the chain. See the discussion here: https://lore.kernel.org/git/20200814045957.GA1380@Abhishek-Arch/ It's one of drawbacks of having a single member 64-bit `generation` instead of two 32-bit members `level` and `odate`. > > > - commit_graph_data->generation stores corrected commit date, > computed as sum of committer date (from CDAT) and offset (from GDAT) > > - A can reach B => gen(A) < gen(B) > > - there is no need for committer date heuristics, and no need for > limiting use of generation number to where there is a cutoff (to not > hamper performance). > > - If there are layers without GDAT chunks, which thanks to the write > behavior means simply top layer without GDAT chunk, we need to turn > off use of generation numbers or fall back to using topological levels > > - commit_graph_data->generation stores topological levels, > taken from CDAT chunk (30-bits) > > - A can reach B => gen(A) < gen(B) > > - we probably want to keep tie-breaking of sorting by generation > number via committer date, and limit use of generation number as > opposed to using committer date heuristics (with slop) to not make > performance worse. > > > > > Thanks to Dr. Stolee, Dr. Narębski, and Taylor for their reviews on the > > first version. > > > > I look forward to everyone's reviews! > > > > Thanks > > > > * Abhishek > > > > > > ---------------------------------------------------------------------------- > > > > Changes in version 3: > > > > * Reordered patches as discussed in 1 > > [https://lore.kernel.org/git/aee0ae56-3395-6848-d573-27a318d72755@xxxxxxxxx/] > > If I remember it correctly this was done to always store in GDAT chunk > corrected commit date offsets, isn't it? > Yes. > > * Split "implement corrected commit date" into two patches - one > > introducing the topo level slab and other implementing corrected commit > > dates. > > All right. > > I think it might be good idea to split off the change to tar file tests > (as a preparatory patch), to make reviews and bisecting easier. > > > * Extended split-commit-graph tests to verify at the end of test. > > Do we also test for proper merging of split commit-graph layers, not > only adding a new layer and a full rewrite (--split=replace)? > We do not, will add a test at end. Thanks for pointing this out. > > * Use topological levels as generation number if any of split commit-graph > > files do not have generation data chunk. > > That is good for performance. > > > > > Changes in version 2: > > > > * Add tests for generation data chunk. > > Good. > > > * Add an option GIT_TEST_COMMIT_GRAPH_NO_GDAT to control whether to write > > generation data chunk. > > Good, that is needed for testing mixed-version behavior. > > > * Compare commits with corrected commit dates if present in > > paint_down_to_common(). > > All right, but see the caveat. > > > * Update technical documentation. > > Always a good thing. > > > * Handle mixed graph version commit chains. > > Where by "version" you mean generation number version - the commit-graph > version number unfortunately needs to stay the same... > Yes, clarified. > > * Improve commit messages for > ^^^^^^ > Something missing in this point, the sentence ends abruptly. I didn't finish the sentence. Meant to say: - Improve commit messages for "commit-graph: fix regression when computing bloom filter", "commit-graph: consolidate fill_commit_graph_info", > > > * Revert unnecessary whitespace changes. > > Thanks. > > > * Split uint_32 -> timestamp_t change into a new commit. > > It is usually better to keep the commits small. Good. > > > Good work! > > ... > -- > Jakub Narębski
