Skip to content

Bug: Evaluation.yml CI workflow jobs are not robust against cache evictions #1765

@alexkeizer

Description

@alexkeizer

The jobs in the evaluation.yml CI workflow are not robust against cache eviction: none of the jobs besides the initial "Build Lean-MLIR" install elan explicitly, instead relying on the .elan folder being cached.

There is, however, no guarantee that the cache will remain available in between runs, see here for example a run where the cache seems to have gotten evicted in between the "build" job and subsequent jobs, which then caused all downstream jobs to fail since they didn't have lake installed.

NOTE: the specific error that happens on a cache eviction is lake: command not found.

WORKAROUND: when this happens, re-running the "Build Lean-MLIR" job usually repopulates the cache properly and fixes the error. Do note that often the cache seems to get evicted after the "Build Lean-MLIR" job succeeds, causing only the downstream jobs to fail. Thus, you have to re-run all jobs to get the cache re-populated, rather than just the failed jobs.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions