[Bug] macOS CI builds intermittently fail with hdiutil resource busy errors

[oznogon]

CI builds disabled by daid as a result in commit 662354d

Known issue closed as not planned/working as designed upstream in actions/runner-images#7522 (comment)

This is a long-standing and well-known issue around hdiutil on macOS.

The problem is confirmed and is caused by a combination of two factors:

• the high sensitivity of hdiutil to disk I/O state and concurrent filesystem activity;
• macOS system behavior itself (background security scanning, metadata/indexing, delayed I/O flushes), which can temporarily hold disk resources and result in Resource busy failures.

...

From the runner-images perspective, a definitive fix is not possible. All involved components (hdiutil, filesystem behavior, and macOS system services) are immutable upstream dependencies and cannot be preconfigured in a way that would permanently resolve this issue for all projects.

Practical and proven mitigation strategies on the consumer side include adding retries with backoff, introducing delays between build steps, explicitly running filesystem synchronization (sync), and in some cases disabling Spotlight/ indexing on build or staging directories.

Closing this issue as working as designed from the image perspective. 🤷‍♂️

[oznogon]

CMake committed a workaround mitigation 6 months ago to the DragNDrop generator used to build disk images: https://gitlab.kitware.com/cmake/cmake/-/commit/ef739edd2078be23761f6e9b2c9dff7a20493cc7

This MR has the 4.2.0 milestone. It will be included in the 4.2 release

cf. https://gitlab.kitware.com/cmake/cmake/-/issues/27136

The current versions of the upstream images all include cmake 4.2.3.

[oznogon]

When @Faileb restored macOS builds in #2547 f86e43f, they added a workaround for actions/runner-images#12912 to uninstall the image's cmake and install homebrew's.

#12912 was closed as fixed after this change, and this workaround might no longer be necessary. If the issue is ultimately related to cmake, the workaround might now be contributing to the issue.