The following page may contain information related to upcoming products, features and functionality. It is important to note that the information presented is for informational purposes only, so please do not rely on the information for purchasing or planning purposes. Just like with all projects, the items mentioned on the page are subject to change or delay, and the development, release, and timing of any products, features or functionality remain at the sole discretion of GitLab Inc.
Git lies at the core of what Gitaly does and is of critical importance to GitLab. We are aiming to increase up our investment into the Git community to ensure that it stays healthy in the long term.
Direction for the GitLab contributions to the Git project is tracked separately from the Gitaly direction, even though they are closely tied to each other. This is done because not only is Git a standalone project, but more importantly it is a community-led open source project, and thus not under direct control of the Gitaly team.
While we can influence the direction of Git somewhat by being regular contributors to the project and building trust and credit with the community, it will never be (and should not be!) under our full control. This is only natural in an open source project and also by design, as the success of Git relies on a healthy and diverse community with interests that may at times be conflicting with each other. We want to foster that community and collaborate with it, including with our direct competitors.
There's no question, repository data sets are growing rapidly. With the prevalence of machine learning and artificial intelligence adoptions, we're seeing a steady increase in repository size as well as the number of files needing to be stored. The Git team needs to ensure that repository storage scales with these expanding data storage needs.
Monorepo performance is a complex topic that involves many different parts of Git.
As projects grow older and grow more commits, the repository's object graph naturally expands. This creates scalability issues over time regardless of the respective object sizes. We want to ensure that the object database of Git repositories can scale well when facing repositories with dozens of millions of objects.
Our goals here are:
Pluggable object databases: The object storage format and the transport format of Git are tightly coupled. Furthermore, access to the object format happens rather ad-hoc without a lot of abstraction. It is clear though that the current design of the object database has limited scalability, and the more features we pile on top of it the harder it becomes to continue scaling.
We need to get back to the drawing board and explore whether it is feasible to make the object database pluggable. This is a huge effort and will likely span over multiple years, but it is needed to ensure long-term scalability of Git.
Better support for bundle URIs: We have started to use bundle URIs to help reduce the load on Gitaly servers. This is especially important in the context of CI systems. The feature is still relatively young though, and there are issues with both the implementation and the UX. We need to invest more time to help polish this feature.
Despite issues with large object graphs, Git is also known to not scale well with large objects like binary files. These are becoming increasingly more important with the advent of AI, but are also critical to open up to additional markets that are heaviliy dependent on binary assets like game studios.
Our goals are:
Offloading blobs through promisor remotes: We need server side changes to enable a triangular workflow between Git server, client, and a promisor remote which will host blobs that we offload from the git server.
Allow easy access to promisor remotes: Make it easier for Git to access promisor remotes. If we have a lot of objects offloaded to object storage because they are rarely used, Git should be able to access the objects and metadata easily without having to pull the objects back into the repository.
Storing large objects more efficiently: Investigate whether we can store large objects more efficiently and allow for deduplication when such large objects are changed. One example would be to use rolling hash functions to split up large objects better, which would potentially allow us to store changes to such large objects incrementally. Furthermore, if we are able to expose those chunks via the transport layer, then this would potentially be able to provide a native replacement for Git LFS when combined with partial clones.
Sending large objects more efficiently: Next to the problem of storing large objects efficiently, we also need to investigate how to send incremental changes to large objects over the wire efficiently. This may require us to expose the above representation with rolling hash functions to clients to efficiently compute deltas. An alternative would be to introduce a new object type that allows us to serve large binaries more efficiently.
We have seen repeatedly that Git and Gitaly start to struggle in repositories that have millions of references. While one approach should be to reduce the number of references that we need to store, we also need to acknowledge that our customers have valid usecases for repositories with this many references. Adapting Git so that it can handle repositories with many references is thus important.
Our goals here are:
Implement reftable support in GitLab: Reftables are a new storage data structure within Git that we have the privilege of contributing to upstream Git. This back end for Git references will improve the performance of repositories with many references, as well as eliminate the current race condition around updating reference. This will translate into better support for monorepos and will help ensure continued stability for our GitLab.com offering.
Performance improvements for the reftable backend: While the reftable has been upstreamed and generally outperforms the files backend in most scenarios, there still are lots of performance optimizations feasible. To reap the full benefits of this new backend we should invest more time into this.
Partial-update transactions: For the reftable backend to operate efficiently, updates of multiple references should ideally happen in a single transaction. Tools like git-fetch(1) may want to update only a subset of those refs though where the expected state matches the actual state of a reference. This is done via separate transactions. With the introduction of partial-update transactions, we allow evicting references from such a transaction, while still continuing with remaining references.
Many commands in Git scale with the size of the file tree, and large monorepos tend to have file trees that are both deep and broad. This creates bottlenecks in commands that scale with tree size. Adapting those commands to handle such large trees better is thus another angle for ensuring future scalability.
Our goals here are:
Better UI for sparse checkouts (client side): Sparse checkouts help users scale by not requiring a fully checked out worktree. For one, this saves disk space. Second, this allows users to skip fetching many blobs for partial clones. The sparse checkout tool in Git is hardly used because of usability challenges. We can vastly improve this command to make it more easily usable for monorepo cases.
git-replay(1): Improve git-replay(1) to support additional usecases for cherry-picks, rebases, rebasing merges and reverts.
git-blame-tree(1): Upstream git-blame-tree(1) to improve performance when
computing the last commit entries in a specific tree have been changed. This
should speed up RPCs like ListLastCommitsForTree() significantly.
Many of our customers have expressed a big interest in cloud-native deployments of Gitaly. Our official stance is that this is not supported, where a huge contributor to this stance is that Git does not handle constrained environments well.
Our goals here are:
Git can be an intimidating tool with very complex syntax. This poses a direct threat because it allows a contender in the same space to take away market share if they provide a significantly better user experience. This has been showing with the recent advent of Jujutsu, which is claimed by many to provide a superior user experience compared to Git.
Our goals here are:
More streamlined interfaces: The user interfaces provided by Git have grown organically over time and are often redundant, inconsistent, confusing and/or hard to discover. We want to invest the time to make the interfaces more consistent and thus easier to use. We want to help the community to adopt a set of guidelines to help adapt existing and design future user interfaces.
Better user journey with large repositories: Provide a better user journey for interacting with very large repositories (in collaboration with the Source Code group).
Observe challengers like Jujutsu: We want to observe development of challengers like Jujutsu (JJ), a new version control system funded by Google for improved user experience with monorepos. We have noticed that many users praise its usability, especially compared to Git. As JJ can use Git repositories as backend, we may want to investigate whether it may be a good addition to the toolchain of users.
Furthermore, we may want to help Google with the libification of Git in this context. If the Git backend used by JJ plays well with Git due to this effort, then we ensure that JJ will stay compatible with repositories hosted on GitLab.
By staying close to developments in JJ, we position ourselves such that we can start to pivot to JJ if this ever becomes necessary.
As we have started to invest more into the Git community, we need to treat it with the same focus on maintainability as we treat projects that are owned by GitLab. We want to ensure that the codebase remains accessible to existing and new contributors and readily extensible. As we entrust Git with handling our customer's most important data, we also need to ensure that the Git project is safe.
Our goals here are:
Integer overflows: During the audit of X41 D-Sec we found many integer overflows in Git that can lead to exploitable out-of-bounds writes. The worst issues were able to trigger remote code execution bugs in Gitaly. While known instances have been fixed, Git is full of implicit integer conversions that can lead to this kind of bug. We should invest the time to harden Git against this kind of vulnerability and fix this whole class of bugs. This requires lots of fixes throughout the codebase so that we can enable relevant compiler warnings.
Memory leaks: While many parts of Git are leak free, others aren't. This is reflected in Git's test suite, where only a subset of tests run with the leak sanitizer enabled. This makes it easy to introduce new memory leaks that go unnoticed. The effect is that we are working against a moving target when trying to reduce memory usage of Git overall. We want to help the Git project such that the whole test suite is free of memory leaks.
Libification: The Git project provides the command line tools as the only
interface to Git repositories, which makes it hard to embed Git-related
functionality into applications directly. There are ongoing efforts to also
provide a linkable library that exposes parts of Git's functionality to such
applications. This requires a large effort to make Git play nice for such a
usecases: removal of global state, removing calls to exit() and abort() and
general refactorings of the to-be-exposed interfaces. We want to help this
effort to eventually have a proper Git library, which would also open up the
ability for ourselves to for example provide long-running Git daemons.
Structured error handling: Improve error handling such that Git can return information about the root cause of failures in a structured way. We currently have to rely on parsing error messages to figure out what exactly has gone wrong, which is fragile and often leads to bugs.
Improved tracking of direction While core developers of the Git community are typically aware of ongoing projects, this is less true for the wider community. This makes it hard to pick up work as a new contributor, requires the community to often re-hash similar discussions, and doesn't allow outside parties to learn about the direction. We want to help the Git project to track direction better to improve the flow of information.
Partial Clone is built-in to Git and available in GitLab 13.0 or newer. Scalar is compatible with partial clone, and Microsoft is contributing to its improvement based on their learnings from the VFS for Git protocol.
Divergent solution for CDN Offloading
While we recognize that a lot of good work has gone into independent solutions, we are committed to work with the Git community on a CDN approach. We intend to support, implement, and contribute to this solution as it be comes available. This is currently being explored in our Support Git CDN offloading epic.