v0.1.21

Features

• Signal computations are now cached. If a signal fails half-way through, it will be resumed.
• Source loading is much faster, up to 40x faster for some sources (e.g. HuggingFace)
• Map dtype is now supported for parquet sources.

Details

• Add jsonl intermediate caching to signals. Introduce a central spot for this cache abstraction. by @nsthorat in #858
• Rename fast_process to load_to_parquet by @brilee in #862
• Implement fast_process for parquet sources by @brilee in #860
• Implement CSV direct to parquet by @brilee in #863
• Implement fast json source by @brilee in #865
• Add map<key, value> dtype. No support in the UI yet. by @dsmilkov in #870
• Implement fast processing for huggingface datasets by @brilee in #869

Bug Fixes & Other Changes

• add development docs on profiling by @brilee in #861
• Add docs for settings and compare mode by @dsmilkov in #859
• Add a nest_under field to dataset.map(). by @nsthorat in #866
• Avoid computing stats for every single field on page load by @dsmilkov in #873
• Fix a sample_size yaml bug by @dsmilkov in #874
• UI fixes for expanding long rows. by @nsthorat in #875
• Fix small bug with compute signal / concepts and filtering by valid dtypes. by @nsthorat in #877
• Add support for map field in the schema and UI by @dsmilkov in #878
• Fix a bug with previewing and comparing on repeated values. by @nsthorat in #879
• Allow custom signals to work with dask processes. by @nsthorat in #880

Full Changelog: v0.1.20...v0.1.21