@Byron In EliahKagan#1 (comment) you had suggested that downgrading serde in Cargo.lock might help. This makes sense, since their versions were upgraded in #1536, where the breakage seems to have started. But does the build in the CI fuzzing job actually use locked dependencies? Is it intended to?

• When I first experimented here, a few minutes ago, with adjusting the version, I had failed to change the associated hashes, and the fuzzing build still got further than I would expect. See, for example, 67024c5.
• More decisively, at aafa8b1, where I tried to set the version to 1.0.204, it is nonetheless using 1.0.210, which is currently the latest version. This seems to indicate that Cargo.lock has no effect here.

Thanks for looking into this! Admittedly I completely forget about it already and CI-Fuzzing would be lost in time.

@Byron In EliahKagan#1 (comment) you had suggested that downgrading serde in Cargo.lock might help. This makes sense, since their versions were upgraded in #1536, where the breakage seems to have started. But does the build in the CI fuzzing job actually use locked dependencies? Is it intended to?

It's true, I also have the impression it uses the the latest possible versions of dependencies, so Cargo.lock shouldn't matter for it.

So updating the image to use a more recent Rust version certainly would help. On top of that, I think the minimal rust-version supported by serde should also be adjusted to capture this dependency.

So updating the image to use a more recent Rust version certainly would help.

I've opened google/oss-fuzz#12512 for this.

I assume fuzzing is completely broken and not just when triggered by our CI, but I have only actually verified that it is broken on CI. I'm not sure how to check other runs, or even if I am able to do so, since I have no special access. (I have access to oss-fuzz reports for GitPython but not, as far as I know, for gitoxide.)

If that change is accepted and if it fixes the problem, this PR could then be merged to reenable CI fuzzing of pull requests, but if that is to be done, then the second commit here should either be dropped via force push, or reverted. This is because downgrading serde-related crates in Cargo.lock is ineffective at fixing fuzzing, and probably undesirable otherwise.

To aid with that, I've gone ahead and reverted that commit. But we can still force push back to the first commit, of course.

On top of that, I think the minimal rust-version supported by serde should also be adjusted to capture this dependency.

Based on serde-rs/serde#2770, and specifically serde-rs/serde#2770 (comment), it looks like the issue is known and only occurs when using old nightly compilers, and that serde only attempts to support current nightly compilers. So based on that my guess is that there may not be anything to do in serde for this.

Eventually, I think the default version oss-fuzz uses of the nightly Rust compiler will be upgraded. See google/oss-fuzz#12410. But we shouldn't have to wait for that to happen to fix this. As noted there and in google/oss-fuzz#12512, the latest nightly is being used in a couple of projects to work around this (google/oss-fuzz#12404, google/oss-fuzz#12409).

Thanks a lot for the summary, I am looking forward to upstream merge so CI will be back here.

This is working now, since google/oss-fuzz#12512.

If this pull request is merged, then it might be a good idea, afterwards, to add the CIFuzz job as a required check for automatic merging. Considerations:

• If it is added, then it could fail due to breakages related to using a nightly toolchain, which might not always be relevant to the content of the pull request being fuzzed. (This could happen if the Docker image is rebuilt, but it is more likely to happen the same way as it happened before: a newer version of a dependency does not support a previous nightly toolchain.)
• But if it is not added, then pull requests that fail fuzzing may be automatically merged before fuzzing completes. Since CI fuzzing is not being done on the push trigger, this would then wait until oss-fuzz fuzzing is performed. So to actually get the intended benefit of fuzzing pull requests, the CIFuzz job should be me made "required" so it blocks automatic merging.
• The CIFuzz job does appear to be the longest-running job, though. So making it required may increase the amount of time it takes for PRs to be merged automatically.

Breakages (if they do not reflect a possible bug due to changes a PR proposes) can reasonably be addressed by temporarily disabling CI fuzzing, as was done in #1536. (In unusual cases, they could possibly be addressed by manually merging pull requests for which fuzzing has failed but other checks have passed.)

This pull request does not attempt to introduce changes to cause fuzzing to be effectively required, because I don't know of a way to do that by editing workflows that is clearly better than configuring it manually.

#1536, which (among other changes) removed the CIFuzz job that this PR restores, was merged on August 23, which was before #1551 was opened and merged on August 25. Reviewing #1551 therefore did not involve any considerations of whether fuzzing failures should block PRs. Because the CIFuzz job is in a different GitHub Actions workflow file from the jobs affected by #1551, a cross-workflow dependency would be needed to include it in a similar way. Since this is not necessarily clearer or more elegant than configuring it as required, I have not made that change here.