This RFC includes a potential direction to make the sparse-checkout more user-friendly. While there, I also present a way to use a limited set of patterns to gain a significant performance boost in very large repositories. Sparse-checkout is only documented as a subsection of the read-tree docs [1], which makes the feature hard to discover. Users have trouble navigating the feature, especially at clone time [2], and have even resorted to creating their own helper tools [3]. This RFC attempts to solve these problems using a new builtin. Here is a sample workflow to give a feeling for how it can work: In an existing repo: $ git sparse-checkout init $ ls myFile1.txt myFile2.txt $ git sparse-checkout add /myFolder/* ^D $ ls myFile1.txt myFile2.txt myFolder $ ls myFolder a.c a.h $ git sparse-checkout disable $ ls hiddenFolder myFile1.txt myFile2.txt myFolder At clone time: $ git clone --sparse origin repo $ cd repo $ ls myFile1.txt myFile2.txt $ git sparse-checkout add /myFolder/* ^D $ ls myFile1.txt myFile2.txt myFolder Here are some more specific details: * git sparse-checkout init enables core.sparseCheckout and populates the sparse-checkout file with patterns that match only the files at root. * git clone learns the --sparse argument to run git sparse-checkout init before the first checkout. * git sparse-checkout add reads patterns from stdin, one per line, then adds them to the sparse-checkout file and refreshes the working directory. * git sparse-checkout disable removes the patterns from the sparse-checkout file, disables core.sparseCheckout, and refills the working directory. * git sparse-checkout list lists the contents of the sparse-checkout file. The documentation for the sparse-checkout feature can now live primarily with the git-sparse-checkout documentation. Cone Mode ========= What really got me interested in this area is a performance problem. If we have N patterns in the sparse-checkout file and M entries in the index, then we can perform up to O(N * M) pattern checks in clear_ce_flags(). This quadratic growth is not sustainable in a repo with 1,000+ patterns and 1,000,000+ index entries. To solve this problem, I propose a new, more restrictive mode to sparse-checkout: "cone mode". In this mode, all patterns are based on prefix matches at a directory level. This can then use hashsets for fast performance -- O(M) instead of O(N*M). My hashset implementation is based on the virtual filesystem hook in the VFS for Git custom code [4]. In cone mode, a user specifies a list of folders which the user wants every file inside. In addition, the cone adds all blobs that are siblings of the folders in the directory path to that folder. This makes the directories look "hydrated" as a user drills down to those recursively-closed folders. These directories are called "parent" folders, as a file matches them only if the file's immediate parent is that directory. When building a prototype of this feature, I used a separate file to contain the list of recursively-closed folders and built the hashsets dynamically based on that file. In this implementation, I tried to maximize the amount of backwards-compatibility by storing all data in the sparse-checkout file using patterns recognized by earlier Git versions. For example, if we add A/B/C as a recursive folder, then we add the following patterns to the sparse-checkout file: /* !/*/* /A/* !/A/*/* /A/B/* !/A/B/*/* /A/B/C/* The alternating positive/negative patterns say "include everything in this folder, but exclude everything another level deeper". The final pattern has no matching negation, so is a recursively closed pattern. Note that I have some basic warnings to try and check that the sparse-checkout file doesn't match what would be written by a cone-mode add. In such a case, Git writes a warning to stderr and continues with the old pattern matching algorithm. These checks are currently very barebones, and would need to be updated with more robust checks for things like regex characters in the middle of the pattern. As review moves forward (and if we don't change the data storage) then we could spend more time on this. Thanks, -Stolee [1] https://git-scm.com/docs/git-read-tree#_sparse_checkoutSparse-checkout documentation in git-read-tree. [2] https://stackoverflow.com/a/4909267/127088Is it possible to do a sparse checkout without checking out the whole repository first? [3] http://www.marcoyuen.com/articles/2016/06/07/git-sparse.htmlA blog post of a user's extra "git-sparse" helper. [4] https://github.com/git/git/compare/fc5fd706ff733392053e6180086a4d7f96acc2af...01204f24c5349aa2fb0c474546d768946d315dab The virtual filesystem hook in microsoft/git. Derrick Stolee (8): sparse-checkout: create builtin with 'list' subcommand sparse-checkout: create 'init' subcommand clone: add --sparse mode sparse-checkout: 'add' subcommand sparse-checkout: create 'disable' subcommand sparse-checkout: add 'cone' mode sparse-checkout: use hashmaps for cone patterns sparse-checkout: init and add in cone mode Jeff Hostetler (1): trace2:experiment: clear_ce_flags_1 .gitignore | 1 + Documentation/config/core.txt | 7 +- Documentation/git-clone.txt | 8 +- Documentation/git-read-tree.txt | 2 +- Documentation/git-sparse-checkout.txt | 146 ++++++++++ Makefile | 1 + builtin.h | 1 + builtin/clone.c | 27 ++ builtin/sparse-checkout.c | 389 ++++++++++++++++++++++++++ cache.h | 8 +- config.c | 10 +- dir.c | 154 +++++++++- dir.h | 27 ++ environment.c | 2 +- git.c | 1 + t/t1091-sparse-checkout-builtin.sh | 195 +++++++++++++ unpack-trees.c | 12 +- 17 files changed, 976 insertions(+), 15 deletions(-) create mode 100644 Documentation/git-sparse-checkout.txt create mode 100644 builtin/sparse-checkout.c create mode 100755 t/t1091-sparse-checkout-builtin.sh base-commit: 5fa0f5238b0cd46cfe7f6fa76c3f526ea98148d9 Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-316%2Fderrickstolee%2Fsparse-checkout%2Fupstream-v1 Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-316/derrickstolee/sparse-checkout/upstream-v1 Pull-Request: https://github.com/gitgitgadget/git/pull/316 -- gitgitgadget
