support for shallow clones and fetches with `gitoxide` #11840

Byron · 2023-03-13T10:58:29Z

This PR makes it possible to enable shallow clones and fetches for git dependencies and crate indices independently with the -Zgitoxide=fetch,shallow_deps and -Zgitoxide=fetch,shallow_index respectively.

Tasks

Review Notes

there is a chance of 'split brain' in git-dependencies as the logic for determining whether the clone/fetch is shallow is repeated in two places. This isn't the case for registries though.

Notes

I am highlighting that this is the gitoxide version of shallow clones as the git2 version might soon be available as well. Having that would be good as it would ensure interoperability remains intact.
Maybe for when git2 has been phased out, i.e. everything else is working, I think (unscientifically) there might be benefits in using worktrees for checkouts. Admittedly I don't know the history of why they weren't used in the first place. Also: gitoxide doesn't yet support local clones and might not have to if worktrees were used instead.

rustbot · 2023-03-13T10:58:35Z

r? @epage

(rustbot has picked a reviewer for you, use r? to override)

Eh2406 · 2023-03-13T20:24:07Z

tests/testsuite/git.rs

+            .all()?
+            .count(),
+        2,
+        "it fetched only the new portion of the history, even though it was unaware of the shallow boundary"


This is truly amazing. Do we know that this will always work correctly?

I'm worried about the case where a repository returns to an earlier version of the file. Say we have commit 1 with "a", 2 with "b", and 3 with "a". We do a shallow clone of 2. Then we do a libgit2 full update from 2->3, the server tells us "the contents of the file is the same as it was in commit 1 which you already have" but we don't have it.

I am concerned that we always correctly handle any errors from libgit2 and that we never deepen a shallow clone.

Right, a valid point! But is it that exhaustive?

If that was true, it wouldn't only have to align the commits to send along with the changes within them, but also back-track through the entire history to do some sort of tree optimization. Not having implemented that part yet definitely leaves my knowledge spotty there, but having implemented pack generation I know that when building the smallest possible pack of a set of commits, I do diff each of them to only send changes. Each of these object candidates to send would now have to pass through a filter that answers the question of "does this object exist already in the portion of the commit graph that the client has? If yes, don't send it." However, what speaks for it being able to do that is its ability to base a delta entry in the back off an object that it knows the other side has (so it doesn't have to send it).

I am aware that the server side has caches which facilitate all kinds of operations (which are mostly non-functional when doing a shallow clone), but that level of sophistication still baffles me :D.

Excuse my ramblings, I think the answer here is that it's still early and I am writing tests to see what happens, and more testing has to be done to see if it's truly safe. If there is any doubt, and I will see if I can get clarification, it would be better to nuke the repository and re-clone it with git2.

Instead of researching this further, I thought it's safest to reinitialize repo instead so using git2 on top of a shallow repo will unshallow it effectively and in a controlled fashion. I am also watching the 'shallow' PR over at the libgit2 repo so this can be adjusted once it becomes available there.

There will always be older cargoes that do not know that shallow clones are a thing. We need a mechanism to ensure that if we shallow clone a repository older cargoes will ether "work entirely correctly" or "not see the checkout".

There are two big things that would not qualify as "work entirely correctly". The obvious one, is crashes with an error. The less obvious one is unshallow the checkout in one of the ways that is extremely expensive for GitHub.

A undesirable scenario that I think still qualifies as "work entirely correctly" is, do a Delta fetch (which as you demonstrate almost always works) and if there's an error blow away the repository and do a full fetch. If we can determine what errors are likely from libgit2 and we have a re-fetch fallback for those errors, I will feel significantly better.

If for any of these reasons we need to make sure that old cargoes do not see shallow clones we will have to make sure that they are put in a different location. Either by adding it to the hash in the repo name or by /git/shallow-db.

All of this applies just as well whether were adding shallowness from git2 or from gix.

Thanks for bringing this to the forefront of my mind.

Having seen what happens, I think the default would be that git2 can operate on shallow repos and receive only an incremental pack. This might not contain all objects they need due to the scenario you mentioned if git really works that well (which it probably does), resulting in some probability of hard errors. With that being a possibility (and even if it is not right now it might be in the future), I feel like it would be best to not let git2 touch shallow repos at all by placing them in their own directory.

Something I also thought about is if it's an option to 'cut' the feature down to an MVP level that yields value while not being used by everyone. Specifically, what if those shallow clones of the index would only be 'for CI' which makes them one-time-use only. Then there is no worry of older cargo's seeing them.

Thinking about it more, git-dependencies might be major benefactors after all, and having them on CI only is certainly a start, but ultimately people probably also want them locally which makes me arrive with the original problem that seemed like changing their hash (to make them invisible) is the safest bet and easiest to implement.

After having made sure that git-dependencies with revisions also work as they should by disabling shallow clones or unshallowing existing shallow clones it became clear that these situations can't be handled correctly by older cargos - instead they will find the revision they try to checkout missing and fail with an error.

The only way around this that I see would be to adjust the place where shallow clones are stored or make them unusable in other ways, I don't know yet.

A problem I see with the current implementation is that it itself makes a decision what to do about the shallowness of a clone, at a stage where the destination repository is already set and with it its location. I will see if it's feasible to move the whole clone into another location then once it's observable whether or not a clone is still shallow.

Running all tests against the 'make all clone and fetches shallow' has proven itself to be very useful to see more special cases and make them work, so when changing the location of clones I will use that as a measure of feasibility as well even though as of now there is no automated test that runs like this.

Byron

I think I will add one more test today, but besides that I believe shallow clones and fetches are working :). I will also run the entire test suite with shallow clones to see if there is unusual failures.

Please feel free to try it yourself.

Byron · 2023-03-14T13:12:13Z

src/cargo/sources/git/oxide.rs

+                &msg,
+            )?;
+        } else if let Some((action, remote)) =
+            find_in(&tasks, |t| progress_by_id(remote_progress, t))


This is specifically for avoiding pauses when shallow cloning the index, which is when the remote does far more work.

src/cargo/sources/git/utils.rs

Byron · 2023-03-15T18:51:56Z

A couple of additional tests have been added to particularly validate that locked revisions of git dependencies work as they should, and furthermore all shallow clones are now placed under a name with the -shallow suffix (I wasn't particularly creative). That way, these shallow repositories aren't going to be visible to git2 or to older cargo's which should make shallow repos backwards compatible. The changes to clone locations have been placed into their own commits to make clearer how tests had to adjusted to accommodate.

All in all, this seems to address all issues that were brought up so far and I am looking forward to receiving more feedback.

bors · 2023-04-01T00:33:23Z

☔ The latest upstream changes (presumably #11881) made this pull request unmergeable. Please resolve the merge conflicts.

bors · 2023-04-03T15:10:50Z

☔ The latest upstream changes (presumably #11928) made this pull request unmergeable. Please resolve the merge conflicts.

weihanglo

Thanks for the great work! It is so close to complete that only tiny suggestions can I give 😆.

As of dealing with the blocker, we're getting closer of making cargo a workspace I believe.

src/cargo/sources/git/source.rs

src/cargo/sources/git/utils.rs

weihanglo · 2023-04-14T08:58:11Z

src/cargo/sources/git/mod.rs

+                if matches!(self, RemoteKind::GitDependencyForbidShallow) {
+                    return History::Unshallow;
+                } else {
+                    gix::remote::fetch::Shallow::NoChange


Would it put pressure on server side? According to this post from GitHub, they suggest avoiding subsequent fetches on a shallow cloned repo. Is this a thing we need to consider?

I think the gist of this advice is that unshallowing shallow clones intentionally or accidentally incurs high cost on the server as it bypasses caches.

Unintentionally unshallowing is something that depends on branch configuration and refspecs (assuming the client code supports shallow), and the fact that cargo limits refspecs for fetches means that unintentional unshallowing would be the exception, not the rule. That could happen nonetheless when the history is rewritten past the shallow boundary for example.

For the crates index, I think this is really something to think about as users would run into this very situation every couple of months. But when they do, the server would traverse only a small history, so the cost incurred is not on the worst end of the spectrum.

Now that we have an implementation to show for, we might even pass it maybe along with an analysis to GitHub to be sure, and I think it's worth seeing this on a timescale as well. We are talking about maybe a year from now when this code might become available by default once git2 is phased out. And by then, the HTTP index is probably already cemented as default so the amount of clones we't actually see like this would be greatly reduced.

Just for the record. Here are some concern of shallow clones from GitHub Infra Team in the past.

Issues Cloning Spec repo - GitHub taking a very long time to download changes to the Specs Repo CocoaPods/CocoaPods#4989 (comment)

update.sh: refuse to update shallow homebrew-core/cask clones. Homebrew/brew#9383

We will continue the conversation on Zulip and post a summary here when we make progess.

src/cargo/sources/git/oxide.rs

tests/testsuite/git.rs

src/cargo/sources/git/utils.rs

rust-lang#11840 (review)

Byron · 2023-04-17T19:13:51Z

Thanks a lot for the review, it's much appreciated.

In todays session I didn't address everything yet, but replied wherever I did. Tomorrow (or in the next days), I'd expect to address everything.

This recreates the previously lost change from rust-lang#12005.

weihanglo · 2023-05-02T15:33:38Z

tests/testsuite/git_shallow.rs

+    Ok(())
+}
+
+fn find_lexicographically_first_bar_checkout() -> std::path::PathBuf {


Just one not important style suggestion: move these helper functions either to the top or the bottom.

Thanks for calling it out - fixed in c207407551ca7888bff0ab06ea86ea879997f81a.

weihanglo · 2023-05-02T15:52:47Z

tests/testsuite/git_shallow.rs

+}
+
+#[cargo_test]
+fn gitoxide_unshallows_git_dependencies_that_may_not_be_shallow_anymore() -> anyhow::Result<()> {


No more unshallow, so title needs a chance?

A good catch, done in in c207407551ca7888bff0ab06ea86ea879997f81a.

weihanglo · 2023-05-02T16:16:06Z

src/cargo/sources/git/utils.rs

+                    // There is a specific commit to fetch and we will just do so in shallow-mode only
+                    // to not disturb the previous logic. Note that with typical settings for shallowing,
+                    // we will just fetch a specific slice of the history.
+                    refspecs.push(format!("+{0}:refs/remotes/origin/HEAD", rev));


Looks good! We still have room to optimize with fallbacks and retries. I am not too concerned with that at this moment, but perhaps we mark it as unresolved in the tracking issue.

weihanglo

Some highlights

Two unstable options for -Zgitoxide
- -Zgitoxide=fetch,shallow_deps for fetching git dependencies
- -Zgitoxide=fetch,shallow_index for fetching registry index
Shallow-cloned and shallow-checked-out git repositories reside at their own -shallow suffixed directories, i.e,
- ~/.cargo/registry/index/*-shallow
- ~/.cargo/git/db/*-shallow
- ~/.cargo/git/checkouts/*-shallow
When the unstable feature is on, fetching/cloning a git repository is always a shallow fetch. This roughly equals to git fetch --depth 1 everywhere.
Even with the presence of Cargo.lock or specifying a commit { rev = "…" }, gitoxide is still smart enough to shallow fetch without unshallowing the existing repository.

I am going to merge this sooner or later.

weihanglo · 2023-05-02T16:21:24Z

tests/testsuite/git_shallow.rs

+    Complete,
+}
+
+#[cargo_test]


Not really your fault. I found git-related tests are all a bit hard to follow, because you need to understand git2 and each helper functions beforehand. I wonder if there is a way to make it simple to learn. Maybe add some equivalent git command in comments to explain what it tries to do?

Not a blocker. I don't expect we to do more sizable updates in this PR 😆.

Not a blocker. I don't expect we to do more sizable updates in this PR 😆

Maybe no change is done in this PR, but I'd love to do better. My only means of describing what's going on right now is a wordy title for the #[test] function, and messages as last argument in assertions. Maybe a leading paragraph could describe the journey from start to end?

Byron · 2023-05-02T17:04:54Z

Some highlights

Two unstable options for -Zgitoxide

-Zgitoxide=fetch,shallow_deps for fetching git dependencies

-Zgitoxide=fetch,shallow_index for fetching registry index

Shallow-cloned and shallow-checked-out git repositories reside at their own -shallow suffixed directories, i.e,

~/.cargo/registry/index/*-shallow

~/.cargo/git/db/*-shallow

~/.cargo/git/checkouts/*-shallow

When the unstable feature is on, fetching/cloning a git repository is always a shallow fetch. This roughly equals to git fetch --depth 1 everywhere.

Even with the presence of Cargo.lock or specifying a commit { rev = "…" }, gitoxide is still smart enough to shallow fetch without unshallowing the existing repository.

I am going to merge this sooner or later.

I wish this summary would be retained in a more prominent spot as it is so very useful :). The tracking issue might be a spot, but maybe there are other options as well.

In any case, I am excited to see this one merged :).

arlosi · 2023-05-02T18:08:00Z

I wish this summary would be retained in a more prominent spot as it is so very useful

Could we add the summary to the existing notes in unstable.md? At, minimum "(planned)" should be removed from shallow-index and shallow-deps as it's being implemented here :)

Co-authored-by: Weihang Lo <[email protected]>

Byron · 2023-05-03T06:56:19Z

@arlosi I think that's just the right place :) - please take a look at the updated unstable.md, which now also contains a modified version of @weihanglo 's original comment.

weihanglo · 2023-05-03T07:12:08Z

Seems great enough for an unstable feature. We can always tweak it if needed afterward. Thank you!

@bors r+

bors · 2023-05-03T07:12:14Z

📌 Commit d2734d3 has been approved by weihanglo

It is now in the queue for this repository.

bors · 2023-05-03T07:12:23Z

⌛ Testing commit d2734d3 with merge 2f06c80...

bors · 2023-05-03T08:08:36Z

☀️ Test successful - checks-actions
Approved by: weihanglo
Pushing 2f06c80 to master...

Update lock to normalize `home` dep I'm not sure what happened, but in #11840 `Cargo.lock` was updated in such a way that the `home` dependency had a qualified source, even though it was equivalent to crates.io. This should only happen if there is some ambiguity (like if something had a path source to the in-tree home). This causes the `Cargo.lock` file to be modified whenever running cargo commands locally. This doesn't fail `--locked` because the resolve is equivalent, pointing to the the same source. Closes #12082

Update cargo 10 commits in ac84010322a31f4a581dafe26258aa4ac8dea9cd..569b648b5831ae8a515e90c80843a5287c3304ef 2023-05-02 13:41:16 +0000 to 2023-05-05 15:49:44 +0000 - xtask-unpublished: output a markdown table (rust-lang/cargo#12085) - fix: hack around `libsysroot` instead of `libtest` (rust-lang/cargo#12088) - Optimize usage under rustup. (rust-lang/cargo#11917) - Update lock to normalize `home` dep (rust-lang/cargo#12084) - fix: doc-test failures (rust-lang/cargo#12055) - feat(cargo-metadata): add `workspace_default_members` (rust-lang/cargo#11978) - doc: clarify implications of `cargo-yank` (rust-lang/cargo#11862) - chore: Use `[workspace.dependencies]` (rust-lang/cargo#12057) - support for shallow clones and fetches with `gitoxide` (rust-lang/cargo#11840) - Build by PackageIdSpec, not name, to avoid ambiguity (rust-lang/cargo#12015) r? `@ghost`

rustbot assigned epage Mar 13, 2023

rustbot added A-git Area: anything dealing with git A-registries Area: registries S-waiting-on-review Status: Awaiting review from the assignee but also interested parties. labels Mar 13, 2023

Byron marked this pull request as draft March 13, 2023 10:58

Byron changed the title ~~[DRAFT] support for shallow clones and fetches with gitoxide~~ support for shallow clones and fetches with gitoxide Mar 13, 2023

Byron force-pushed the shallow-support branch 2 times, most recently from 4459329 to 367074c Compare March 13, 2023 17:16

Eh2406 reviewed Mar 13, 2023

View reviewed changes

Byron force-pushed the shallow-support branch 2 times, most recently from 45f874a to e81797d Compare March 14, 2023 13:07

Byron commented Mar 14, 2023

View reviewed changes

Byron force-pushed the shallow-support branch 5 times, most recently from 0ee9c5d to 44924a7 Compare March 15, 2023 18:46

Byron marked this pull request as ready for review March 15, 2023 18:52

Byron mentioned this pull request Mar 15, 2023

[integration] Shallow clones for cargo GitoxideLabs/gitoxide#449

Open

9 tasks

Byron force-pushed the shallow-support branch from 44924a7 to 518b909 Compare April 1, 2023 12:28

Byron force-pushed the shallow-support branch from 518b909 to 29f0760 Compare April 4, 2023 11:42

weihanglo reviewed Apr 14, 2023

View reviewed changes

Byron added a commit to Byron/cargo that referenced this pull request Apr 17, 2023

Various improvements to address the PR review (part 1).

5716fe5

rust-lang#11840 (review)

Byron force-pushed the shallow-support branch 2 times, most recently from 439c4f2 to c700901 Compare April 18, 2023 06:11

Byron added 2 commits April 29, 2023 14:20

set openssl-src to specific version

99316c8

This recreates the previously lost change from rust-lang#12005.

Only fetch single object in shallow mode for compatibility

2aaebcb

weihanglo reviewed May 2, 2023

View reviewed changes

weihanglo approved these changes May 2, 2023

View reviewed changes

Move helper functions to the bottom for consistency; fix name

3bd08d9

Update unstable.md with new gitoxide options for shallow clones

d2734d3

Co-authored-by: Weihang Lo <[email protected]>

rustbot added the A-documenting-cargo-itself Area: Cargo's documentation label May 3, 2023

bors added S-waiting-on-bors Status: Waiting on bors to run and complete tests. Bors will change the label on completion. and removed S-waiting-on-review Status: Awaiting review from the assignee but also interested parties. labels May 3, 2023

bors merged commit 2f06c80 into rust-lang:master May 3, 2023

Byron deleted the shallow-support branch May 3, 2023 08:15

This was referenced May 3, 2023

feat(cargo-metadata): add workspace_default_members #11978

Merged

chore: Use [workspace.dependencies] #12057

Merged

ehuss mentioned this pull request May 4, 2023

Update lock to normalize home dep #12084

Merged

weihanglo mentioned this pull request May 5, 2023

Update cargo rust-lang/rust#111258

Merged

ehuss added this to the 1.71.0 milestone May 5, 2023

rfcbot added finished-final-comment-period FCP complete to-announce and removed final-comment-period FCP — a period for last comments before action is taken labels May 7, 2023

weihanglo mentioned this pull request May 9, 2023

Update libc to 0.2.144 #12098

Merged

weihanglo mentioned this pull request Dec 5, 2023

gitoxide auth tests sporadically failing #11821

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

support for shallow clones and fetches with `gitoxide` #11840

support for shallow clones and fetches with `gitoxide` #11840

Byron commented Mar 13, 2023 •

edited

Loading

rustbot commented Mar 13, 2023

Eh2406 Mar 13, 2023

Byron Mar 14, 2023 •

edited

Loading

Byron Mar 14, 2023

Eh2406 Mar 14, 2023

Byron Mar 14, 2023

Byron Mar 15, 2023

Byron left a comment

Byron Mar 14, 2023

Byron commented Mar 15, 2023

bors commented Apr 1, 2023

bors commented Apr 3, 2023

weihanglo left a comment

weihanglo Apr 14, 2023

Byron Apr 17, 2023 •

edited

Loading

weihanglo Apr 21, 2023

Byron commented Apr 17, 2023

weihanglo May 2, 2023

Byron May 2, 2023

weihanglo May 2, 2023

Byron May 2, 2023

weihanglo May 2, 2023

weihanglo left a comment

weihanglo May 2, 2023

Byron May 2, 2023

Byron commented May 2, 2023

arlosi commented May 2, 2023

Byron commented May 3, 2023

weihanglo commented May 3, 2023

bors commented May 3, 2023

bors commented May 3, 2023

bors commented May 3, 2023

support for shallow clones and fetches with gitoxide #11840

support for shallow clones and fetches with gitoxide #11840

Conversation

Byron commented Mar 13, 2023 • edited Loading

Tasks

Review Notes

Notes

rustbot commented Mar 13, 2023

Choose a reason for hiding this comment

Byron Mar 14, 2023 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Byron left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Byron commented Mar 15, 2023

bors commented Apr 1, 2023

bors commented Apr 3, 2023

weihanglo left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Byron Apr 17, 2023 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Byron commented Apr 17, 2023

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

weihanglo left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Byron commented May 2, 2023

arlosi commented May 2, 2023

Byron commented May 3, 2023

weihanglo commented May 3, 2023

bors commented May 3, 2023

bors commented May 3, 2023

bors commented May 3, 2023

support for shallow clones and fetches with `gitoxide` #11840

support for shallow clones and fetches with `gitoxide` #11840

Byron commented Mar 13, 2023 •

edited

Loading

Byron Mar 14, 2023 •

edited

Loading

Byron Apr 17, 2023 •

edited

Loading