Skip to content

CI: cache KGs/LPs/NCESData downloads with actions/cache #637

Closed

Description

@Demirrr

Suggested by @TheSheep-stack in #622: CI downloads KGs.zip, LPs.zip, NCESData.zip and CLIPData.zip from files.dice-research.org on every run (see .github/workflows/test.yml), and download speed is occasionally very slow.

Caching these with actions/cache (keyed on the archive URL/version) would cut CI time and remove the dependency on the file server being responsive for every run.

Not a bug and not blocking; split out of #622 once its root cause (DataLoader worker-pool deadlock, #614/#627/#628) was fixed.

Activity

  1. TheSheep-stack commented on Oct 16, 2026

    @TheSheep-stack
    Collaborator
    1. With python and external download caches, the CI times are only cut by around 2.5 minutes.
    2. Apparently caches are invalidated every 7 days automatically.
    3. Creating a coverage report only on merge requests, instead of each push may cut some more time.
    4. What about test parallelization ? It may need a bit more effort because of Fuseki, but should be possible.

    What do you think @Demirrr

  2. Demirrr commented on Oct 16, 2026

    @Demirrr
    MemberAuthor

    Thanks @TheSheep-stack, my thoughts on your points:

    1. Caching downloads (actions/cache): yes, let's do it. The gain is small (~2.5 min of a 35-45 min run), but the main benefit is that CI no longer depends on files.dice-research.org being fast on every run. I'd key the cache on a manual version string (e.g. ontolearn-data-v1) rather than the URL, since the archives change rarely and without notice.

    2. 7-day eviction: acceptable. On a miss we just re-download. Note that caches are scoped per branch with fallback to the default branch, so develop needs to populate it for PRs to get hits (it does, since the workflow runs on every push).

    3. Coverage only on PRs: I'd skip this. Coverage overhead is small compared to the learner tests, and the coverage report output isn't uploaded anywhere. The bigger free win is the trigger: on: [push, pull_request] runs the full suite twice for every PR branch push. Restricting push to develop/main and keeping pull_request should save more than anything else here.

    4. Test parallelization: the only option that would materially cut the runtime, but also the riskiest:

    A safer alternative is splitting test files across matrix jobs (learners / neural / triplestore). That needs no test changes, but each job restores the data cache, which is another reason to do (1) first.

    Proposed order: (1) add the cache step, closing this issue; (2) fix the duplicate push/PR triggers; (3) open a separate issue for parallelization, trying the matrix split before xdist.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

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