Skip to content

Preserve runtime dataset settings when reloading metadata caches - #3298

Merged
bghira merged 1 commit into
mainfrom
fix/preserve-live-dataset-config
Oct 28, 2026
Merged

bghira merged 1 commit into
mainfrom
fix/preserve-live-dataset-config

Conversation

@bghira

@bghira bghira commented Oct 28, 2026

Copy link
Copy Markdown
Owner

Reloading a metadata cache could replace the active dataset configuration after registration. A requested batch size of 4 could therefore become the cached batch size of 2 before sampler construction; repeats and dataset activation settings were affected too.

Keep cached configuration on the metadata backend for configuration validation, without publishing it into the live registry. Apply this consistently to discovery, Hugging Face, Parquet, Webshart, and caption metadata backends. Regression tests use the real registry and a registered Webshart backend to verify sampler batch sizing.

Validation:

  • .venv/bin/python -m unittest -v -f tests.test_metadata_runtime_config tests.test_metadata_backend tests.test_factory_edge_cases tests.test_factory_caption_backend tests.test_webshart_backend tests.test_huggingface_metadata_backend — 213 tests passed.
  • .venv/bin/python -m unittest discover -v -f tests/ — 7,110 tests completed successfully, with 323 skips.
  • All pre-commit hooks passed.

@bghira
bghira merged commit 0734a7d into main Oct 28, 2026
2 checks passed
@bghira
bghira deleted the fix/preserve-live-dataset-config branch October 28, 2026 13:11
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant