At least twice folks have run into situations where the CI linter would give an error that they didn't think was related to their changes, wouldn't show up when they ran lintrunner locally, and so they force merged assuming the failure wasn't their fault.

Instance one:
Dev couldn't repo the lint errors on his PR

Probably this was the set of confusing linter errors. If not, it would be something from this list

This seems to have been missed by the local linter since the local one only looks at changes that go one commit back

Instance two:
Don't have the PR for this right now, but Quansight reported that they took a dependency on some Dynamo/Inductor package which introduced a linter error. The error occurred since Inductor/Dynamo packages are currently excluded from some linting/policy enforcement rules but any other package that imports from them has to follow those rules, and those requirements then transversely become imposed on the inductor/dynamo packages.

However, the local lintrunner doesn't apply those transverse errors, making the local lintrunner say things are fine while the CI linter fails.

cc @seemethere @malfet @pytorch/pytorch-dev-infra