| Message ID | pull.1005.git.1630359290.gitgitgadget@gmail.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | Upstreaming the Scalar command | expand |
On 8/30/21 5:34 PM, Johannes Schindelin via GitGitGadget wrote: > tl;dr: This series contributes the Scalar command to the Git project. This > command provides an opinionated way to create and configure repositories > with a focus on very large repositories. I want to give Johannes a big thanks for organizing this RFC. As you can see from the authorship of the patches, this was an amazingly collaborative effort, but Johannes led the way by creating a base that the rest of us could work with, then finally he brought in all of the gritty details to finish the effort. > Background > ========== ... > The Scalar project > was created to make that separation, refine the key concepts, and then > extract those features into the new Scalar command. When people have asked me how Scalar fits with the core Git client, I point them to our "Philosophy of Scalar" document [1]. The most concise summary of our goals since starting Scalar has been that Scalar aligns with features already within Git that enable scale. I've said several times that we are constantly making Scalar do less by making Git do more. [1] https://github.com/microsoft/git/blob/HEAD/contrib/scalar/docs/philosophy.md Here is an example: when our large, internal customer told us that they required Linux support for Scalar, we looked at what it would take. We could have done the necessary platform-specific things to convince .NET Core to create a long-running process that launched Git maintenance tasks at different intervals, creating a similar mechanism to the Windows and macOS services that did those operations. But we also knew that the existing system was stuck with architectural decisions from VFS for Git that were not actually in service of how Scalar worked. Instead, we decided to build background maintenance into Git itself and had our Linux port of Scalar run "git maintenance start". Once the Linux port was proven out with Git's background maintenance, we realized that the window where a user actually interacts with Scalar instead of Git is extremely narrow: users run "scalar clone" or "scalar register" and otherwise only run Git commands. The Scalar process does not need to exist outside of that. (There are some other helpers that can be used in a pinch to diagnose and fix problems, but they are rarely used. These commands, such as 'scalar diagnose' can be contributed separately.) It became clear that for our own needs it would be easier to ship one installer that included the microsoft/git fork and the Scalar CLI, and it would be simple to rewrite the Scalar CLI with all of the Git helper APIs. We organized the code in a way that we thought would be amenable to an upstream contribution (by placing in contrib/ and using Git code style). The thing about these commands is that they are _opinionated_. We rely on these opinions for important internal users, but we realize that they are not necessarily optimal for all users. Hence, we did not think it wise to push those opinions onto the 'git' executable. Having 'scalar' continue to live as a separate executable made sense to us. I believe that by contributing Scalar to the full community, that we create opportunities for Git in the future. For one, users and Git distributors can opt into compiling Scalar so it is more available to users who are interested. Another hopeful idea is that maybe this reinvigorates ideas of how to streamline Git clones for large repos without users needing to learn each and every knob to twist to get things working. Since the Scalar CLI is contributed in the full license of the Git project, pieces of it can be adapted into Git proper as needed. I look forward to hearing your thoughts. Thanks, -Stolee
On Mon, Aug 30, 2021 at 5:52 PM Derrick Stolee <stolee@gmail.com> wrote: > > On 8/30/21 5:34 PM, Johannes Schindelin via GitGitGadget wrote: > > tl;dr: This series contributes the Scalar command to the Git project. This > > command provides an opinionated way to create and configure repositories > > with a focus on very large repositories. > > I want to give Johannes a big thanks for organizing this RFC. As you > can see from the authorship of the patches, this was an amazingly > collaborative effort, but Johannes led the way by creating a base that > the rest of us could work with, then finally he brought in all of the > gritty details to finish the effort. > > > Background > > ========== > > ... > > > The Scalar project > > was created to make that separation, refine the key concepts, and then > > extract those features into the new Scalar command. > > When people have asked me how Scalar fits with the core Git client, I > point them to our "Philosophy of Scalar" document [1]. The most concise > summary of our goals since starting Scalar has been that Scalar aligns > with features already within Git that enable scale. I've said several > times that we are constantly making Scalar do less by making Git do more. > > [1] https://github.com/microsoft/git/blob/HEAD/contrib/scalar/docs/philosophy.md > > Here is an example: when our large, internal customer told us that they > required Linux support for Scalar, we looked at what it would take. We > could have done the necessary platform-specific things to convince .NET > Core to create a long-running process that launched Git maintenance tasks > at different intervals, creating a similar mechanism to the Windows and > macOS services that did those operations. But we also knew that the > existing system was stuck with architectural decisions from VFS for Git > that were not actually in service of how Scalar worked. Instead, we > decided to build background maintenance into Git itself and had our Linux > port of Scalar run "git maintenance start". > > Once the Linux port was proven out with Git's background maintenance, we > realized that the window where a user actually interacts with Scalar instead > of Git is extremely narrow: users run "scalar clone" or "scalar register" > and otherwise only run Git commands. The Scalar process does not need to > exist outside of that. (There are some other helpers that can be used in > a pinch to diagnose and fix problems, but they are rarely used. These > commands, such as 'scalar diagnose' can be contributed separately.) > > It became clear that for our own needs it would be easier to ship one > installer that included the microsoft/git fork and the Scalar CLI, and > it would be simple to rewrite the Scalar CLI with all of the Git helper > APIs. We organized the code in a way that we thought would be amenable > to an upstream contribution (by placing in contrib/ and using Git code > style). > > The thing about these commands is that they are _opinionated_. We rely > on these opinions for important internal users, but we realize that they > are not necessarily optimal for all users. Hence, we did not think it > wise to push those opinions onto the 'git' executable. Having 'scalar' > continue to live as a separate executable made sense to us. > > I believe that by contributing Scalar to the full community, that we > create opportunities for Git in the future. For one, users and Git > distributors can opt into compiling Scalar so it is more available > to users who are interested. Another hopeful idea is that maybe this > reinvigorates ideas of how to streamline Git clones for large repos > without users needing to learn each and every knob to twist to get > things working. Since the Scalar CLI is contributed in the full > license of the Git project, pieces of it can be adapted into Git > proper as needed. > > I look forward to hearing your thoughts. > > Thanks, > -Stolee Looks like exciting stuff, you two. I'm behind on review as it is; I still need to get back to Stolee's sparse-index add/rm/mv series, but I'll try to circle back and take a look.
tl;dr: This series contributes the Scalar command to the Git project. This command provides an opinionated way to create and configure repositories with a focus on very large repositories. Background ========== Years ago, Microsoft wanted to move the source code of the Windows operating system to Git. The challenge there was to prove that Git could scale to massive monorepos. The VFS for Git (formerly GVFS) project was born to take up that challenge. The final solution included a virtual filesystem (with both user-mode and kernel components) and a customized fork of Git for Windows. This solution contained several key concepts, such as only populating a portion of the working directory, demand-fetching blobs, and performing periodic repo maintenance in the background. However, the required kernel drivers made it difficult to port the solution to other platforms. But it was realized that many of these key concepts were independent of the actual VFS and its projection of the working directory. The Scalar project was created to make that separation, refine the key concepts, and then extract those features into the new Scalar command. The present =========== The Scalar project provides a completely functional non-virtual experience for monorepos. But why stop there. The Scalar project was designed to be a self-destructing vehicle to allow those key concepts to be moved into core Git itself for the benefit of all. For example, partial clone, sparse-checkout, and background maintenance have already been upstreamed and removed from Scalar proper. This patch series provides a C-based implementation of the final remaining portions of the Scalar command. This will make it easier for users to experiment with the Scalar command. It will also make it substantially easier to experiment with moving functionality from Scalar into core Git, while maintaining backwards-compatibility for existing Scalar users. The C-based Scalar has been shipped to Scalar users, and can be tested by any interested reader: https://github.com/microsoft/git/releases/tag/v2.33.0.vfs.0.0 (it offers a Git for Windows installer, a macOS package and an Ubuntu package). Opportunities ============= Apart from providing the Scalar command, this contribution is intended to serve as a basis for further mailing list discussions on moving (some of) these key concepts into the main Git commands. For example, we previously discussed the idea of a "git big-clone" that does much of what "scalar clone" is doing. This patch series is a step to make such functionality exist in the Git code base while we simmer on what such a "git big-clone" command-line interface would look like. This is one of many possible ways to do this. Creating a 'git big-clone' could lock Git into backwards compatibility concerns so it is necessary to approach such an endeavor with caution. As a discussion starter, the scalar clone <url> command does roughly this: 1. git clone --sparse --filter=blob:none /src 2. git -C /src sparse-checkout init --cone 3. git -C /src config (many times) 4. git -C /src maintenance start It is my hope inspire discussions about what parts of Scalar could go into core Git, and where, and in which form. While we wish to maintain backwards-compatibility of Scalar's command-line interface (because it is already in use), by having the Scalar code in the same code base as Git's, it will be much easier to move functionality without having to maintain loose version coupling between independently-versioned Scalar and Git. The tight version-coupling, along with having access to libgit.a also allows the C-based implementation of Scalar to be much smaller than the original .NET version. For example, we might choose in the future to implement, say, git clone --scale=partial,cone to initialize a partial clone with a cone-sparse checkout, that would not only be totally doable, and not only would we already have precedent and data to prove that this actually makes engineers happy who have to work on ginormous repositories, but we could then also implement it by moving parts of contrib/scalar/ to builtin/ (where contrib/scalar/ would then call the built-ins accordingly rather than hard-coding the defaults itself). We now also have the opportunity to discuss the merits of Scalar's clone caching, which is not actually part of this patch series because it is a bit coupled with the GVFS parts of microsoft/git for the moment, where clones automatically get registered with a populated alternate repository that is identified by the URL, meaning: subsequent clones of the same repository are vastly faster than the first one because they do not actually download the already-received objects again, they access the cache instead. Another thing that I could imagine to be discussed at length is the distinction between enlistment and worktree (where the latter is the actual Git worktree and usually lives in the src/ subdirectory of the former). This encourages untracked and ignored files to be placed outside the worktree, making Git's job much easier. This idea, too, might find its way in one way or another into Git proper. These are just a few concepts in Scalar that do not yet have equivalents in Git. By putting this initial implementation into contrib/, we create a foundation for future discussions of these concepts. We plan on updating the recommended config settings in scalar register as new Git features are available (such as builtin FSMonitor and sparse-index, when ready). To facilitate upgrading existing Scalar enlistments, their paths are automatically added to the [scalar] section of the global Git config, and the scalar reconfigure --all command will process all of them. Epilogue ======== Now, to address some questions that I imagine every reader has who made it this far: * Why not put the Scalar functionality directly into a built-in? Creating a Git builtin requires scrutiny over every aspect of the feature, which is difficult to do while also maintaining the command-line interface contract and expected behavior of the Scalar command (there are existing users, after all). By having the Scalar command in contrib/, we present a simple option for users to have these features in the short term while the Git contributor community decides which bits to absorb into Git built-ins. * Why implement the Scalar command in the Git codebase? We ported Scalar to the microsoft/git fork for several reasons. First, we realized it was possible now that the core features exist inside Git itself. Second, compiling Scalar directly within a version of Git allows us to remove a version compatibility check from each config option that might or might not apply based on the installed Git version. Finally, this new location has greatly simplified our release process and the installation process for users. We now have ways to install Scalar with microsoft/git via winget, brew, and apt-get. This has been the case since we shipped v2.32.0 to our users, read: this setup has served us well already. * Why contribute Scalar to the Git project? We are biased, of course, yet we do have evidence that the Scalar command is a helpful tool that offers an simple way to handle huge repositories with ease. By contributing it to the core Git project, we are able to share it with more users, especially some users who do not want to install the microsoft/git fork. We intend to include Scalar as a component in git-for-windows/git, but are contributing it here first. Further, we think there is benefit to the Git developer community as this presents an example of how to set certain defaults that work for large repositories. * Does this integrate with the built-in FSMonitor yet? No, not yet. I do have a couple of add-on patch series lined up, one of them being the integration with the built-in FSMonitor, which obviously has to wait until the FSMonitor patch series advances further. Derrick Stolee (4): scalar: 'register' sets recommended config and starts maintenance scalar: 'unregister' stops background maintenance scalar: implement 'scalar list' scalar: implement the `run` command Johannes Schindelin (10): scalar: create a rudimentary executable scalar: start documenting the command scalar: create test infrastructure scalar: let 'unregister' handle a deleted enlistment directory gracefully scalar: implement the `clone` subcommand scalar: teach 'clone' to support the --single-branch option scalar: allow reconfiguring an existing enlistment scalar: teach 'reconfigure' to optionally handle all registered enlistments scalar: implement the `version` command scalar: accept -C and -c options before the subcommand Matthew John Cheetham (1): scalar: implement the `delete` command Makefile | 8 + contrib/scalar/.gitignore | 5 + contrib/scalar/Makefile | 57 +++ contrib/scalar/scalar.c | 844 +++++++++++++++++++++++++++++++ contrib/scalar/scalar.txt | 152 ++++++ contrib/scalar/t/Makefile | 78 +++ contrib/scalar/t/t9099-scalar.sh | 88 ++++ 7 files changed, 1232 insertions(+) create mode 100644 contrib/scalar/.gitignore create mode 100644 contrib/scalar/Makefile create mode 100644 contrib/scalar/scalar.c create mode 100644 contrib/scalar/scalar.txt create mode 100644 contrib/scalar/t/Makefile create mode 100755 contrib/scalar/t/t9099-scalar.sh base-commit: ebf3c04b262aa27fbb97f8a0156c2347fecafafb Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-1005%2Fdscho%2Fscalar-the-beginning-v1 Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-1005/dscho/scalar-the-beginning-v1 Pull-Request: https://github.com/gitgitgadget/git/pull/1005