Working with diffs
We rely on different sources to present diffs. These include:
- Gitaly service
- Database (through
- Redis (cached highlighted diffs)
In January 2019, Oswaldo Ferreira hosted a Deep Dive (GitLab team members only:
https://gitlab.com/gitlab-org/create-stage/issues/1) on GitLab Diffs and Commenting on Diffs
functionality to share domain-specific knowledge with anyone who may work in this part of the
codebase in the future:
- Recording on YouTube
- Slides on Google Slides
- PDF slides
Everything covered in this deep dive was accurate as of GitLab 11.7, and while specific details may have changed since then, it should still serve as a good introduction.
Merge request diffs
When refreshing a merge request (pushing to a source branch, force-pushing to target branch, or if the target branch now contains any commits from the MR)
we fetch the comparison information using
Gitlab::Git::Compare, which fetches
head data using Gitaly and diff between them through
The diffs fetching process limits single file diff sizes and the overall size of the whole diff through a series of constant values. Raw diff files are
then persisted on
Even though diffs larger than 10% of the value of
ApplicationSettings#diff_max_patch_bytes are collapsed,
we still keep them on PostgreSQL. However, diff files larger than defined safety limits
(see the Diff limits section) are not persisted in the database.
In order to present diffs information on the merge request diffs page, we:
- Fetch all diff files from database
- Fetch the old and new file blobs in batch to:
- Highlight old and new file content
- Know which viewer it should use for each file (text, image, deleted, etc)
- Know if the file content changed
- Know if it was stored externally
- Know if it had storage errors
- If the diff file is cacheable (text-based), it's cached on Redis
When commenting on a diff (any comparison), we persist a truncated diff version
NoteDiffFile (which is associated with the actual
DiffNote). So instead
of hitting the repository every time we need the diff of the file, we:
- Check whether we have the
NoteDiffFile#diffpersisted and use it
- Otherwise, if it's a current MR revision, use the persisted
- In the last scenario, go the repository and fetch the diff
As explained above, we limit single diff files and the size of the whole diff. There are scenarios where we collapse the diff file, and cases where the diff file is not presented at all, and the user is guided to the Blob view.
Diff collection limits
Limits that act onto all diff files collection. Files number, lines number and files size are considered.
Gitlab::Git::DiffCollection.collection_limits[:safe_max_files] = Gitlab::Git::DiffCollection::DEFAULT_LIMITS[:max_files] = 100
File diffs are collapsed (but are expandable) if 100 files have already been rendered.
Gitlab::Git::DiffCollection.collection_limits[:safe_max_lines] = Gitlab::Git::DiffCollection::DEFAULT_LIMITS[:max_lines] = 5000
File diffs are collapsed (but be expandable) if 5000 lines have already been rendered.
Gitlab::Git::DiffCollection.collection_limits[:safe_max_bytes] = Gitlab::Git::DiffCollection.collection_limits[:safe_max_files] * 5.kilobytes = 500.kilobytes
File diffs are collapsed (but be expandable) if 500 kilobytes have already been rendered.
Gitlab::Git::DiffCollection.collection_limits[:max_files] = Commit::DIFF_HARD_LIMIT_FILES = 1000
No more files are rendered at all if 1000 files have already been rendered.
Gitlab::Git::DiffCollection.collection_limits[:max_lines] = Commit::DIFF_HARD_LIMIT_LINES = 50000
No more files are rendered at all if 50,000 lines have already been rendered.
Gitlab::Git::DiffCollection.collection_limits[:max_bytes] = Gitlab::Git::DiffCollection.collection_limits[:max_files] * 5.kilobytes = 5000.kilobytes
No more files are rendered at all if 5 megabytes have already been rendered.
All collection limit parameters are sent and applied on Gitaly. That is, after the limit is surpassed,
Gitaly only returns the safe amount of data to be persisted on
Individual diff file limits
Limits that act onto each diff file of a collection. Files number, lines number and files size are considered.
Expandable patches (collapsed)
Diff patches are collapsed when surpassing 10% of the value set in
That is, it's equivalent to 10kb if the maximum allowed value is 100kb.
The diff is persisted and expandable if the patch size doesn't
Although this nomenclature (Collapsing) is also used on Gitaly, this limit is only used on GitLab (hardcoded - not sent to Gitaly).
Gitaly only returns
Diff.Collapsed (RPC) when surpassing collection limits.
Not expandable patches (too large)
The patch not be rendered if it's larger than
Users see a
Changes are too large to be shown. message and a button to view only that file in that commit.
Commit::DIFF_SAFE_LINES = Gitlab::Git::DiffCollection::DEFAULT_LIMITS[:max_lines] = 5000
File diff is suppressed (technically different from collapsed, but behaves the same, and is expandable) if it has more than 5000 lines.
This limit is hardcoded and only applied on GitLab.
Diff Viewers, which can be found on
models/diff_viewer/* are classes used to map metadata about each type of Diff File. It has information
whether it's a binary, which partial should be used to render it or which File extensions this class accounts for.
DiffViewer::Base validates blobs (old and new versions) content, extension and file type to check if it can be rendered.
Merge request diffs against the
HEAD of the target branch
Historically, merge request diffs have been calculated by
git diff target...source which compares the
HEAD of the source branch with the merge base (or a common ancestor) of the target branch and the source's.
This solution works well until the target branch starts containing some of the
changes introduced by the source branch: Consider the following case, in which the source branch
feature_a and the target is
- Checkout a new branch
- Add a commit that removes
The merge request diff still contains the
file_a removal while the actual diff compared to
HEAD has only the
file_b removal. The diff with such redundant
changes is harder to review.
In order to display an up-to-date diff, in GitLab 12.9 we
introduced merge request
diffs compared against
HEAD of the target branch: the
target branch is artificially merged into the source branch, then the resulting
merge ref is compared to the source branch to calculate an accurate
Until we complete the epics "use merge refs for diffs"
and "merge conflicts in diffs",
main (base) and
main (HEAD) are available to be displayed in merge requests:
main (HEAD) option is meant to replace
main (base) in the future.
In order to support comments for both options, diff note positions are stored for
main (base) and
main (HEAD) versions (introduced in 12.10).
The position for
main (base) version is stored in
Note#original_position columns, for
main (HEAD) version
has been introduced.
One of the key challenges to deal with when working on merge ref diffs are merge conflicts. If the target and source branch contains a merge conflict, the branches cannot be automatically merged. The recording on YouTube is a quick introduction to the problem and the motivation behind the epic.
In 13.5 a solution for both-modified merge conflict has been introduced. However, there are more classes of merge conflicts that are to be addressed in the future.