Eventual-Inc/Daft - CodSpeed

Daft

Blog Docs Changelog

Performance History

Latest Results

feat(io): deprecate built-in OpenDAL backends in favor of the opendal Python package OSS, COS, OBS, TOS, GooseFS and GitHub are no longer compiled by default; the same services are now served through the `opendal` Python package via 'pip install "daft[extra-fs]"', which Daft drives as a runtime fallback for any scheme not compiled in. Failures report their precise cause (missing package, missing config, or a service the wheel does not ship). Native support remains available behind per-service features. Resolves #7309

Lucas61000:issue-7309

2 hours ago

fix(core): keep Image(None) type-level mode for mode-agnostic image casts (#7433) Refining the type-level mode to Image(Some(mode)) for uniform columns violates the plan-time/runtime dtype equality invariant: the cast is planned as Image(None), so expression evaluation panics on the exact dtype assertion when the runtime result is narrowed. Drop the type-level refinement and keep the cast output at Image(None), matching the planned dtype. The per-image mode is still derived from the tensor's inner dtype instead of a hardcoded UInt8. Adds regression tests at both the Series and DataFrame level.

Lucas61000:issue-7433

5 hours ago

fix(distributed): preserve cleanup on background failure

zhouwenjiajia:bugfix/joinable-stream-background-error

20 hours ago

fix(dashboard): require a full scheme prefix when normalizing the Ray UI URL Found while dogfooding against a live Ray cluster: `startswith("http")` treats a scheme-less hostname that happens to begin with "http" as already-qualified, so `DAFT_RAY_DASHBOARD_URL=http-proxy.example.com` was recorded as `http-proxy.example.com/#/jobs/<id>`. Check for `http://`/`https://` instead, which is what the dashboard frontend does when it renders the link.

claude/ray-ui-url-investigation-Q9HBW

1 day ago

fix(sql): use the default io config in SQL readers read_parquet, read_csv, read_json and read_deltalake built their io_config from the named argument only, so an absent argument became None and the context default_io_config was dropped. The Python readers fall back to it. read_iceberg keeps None on purpose so the scan can fall back to the table's own FileIO properties. Closes #7448

jackylee-ch:sql-reader-default-io-config

1 day ago

test(iceberg): read the timestamptz table back through pyiceberg

jackylee-ch:iceberg-partition-repr-parity

1 day ago

fix(parquet): error instead of nulling columns when field ids are absent apply_field_ids_to_arrowrs_parquet_metadata drops fields whose id is missing from the mapping. A file with no field ids at all lost every column, so the reader returned all-null data with no warning. Raise MissingParquetFieldIds instead; partially annotated files keep the existing behavior. Closes #7445

jackylee-ch:parquet-field-id-mapping-error

1 day ago

fix(parquet): account for deleted rows in limit pushdown prune_row_groups subtracted each row group's pre-delete num_rows() from the limit budget, so a heavily-deleted row group exhausted it and later row groups with matching rows were never decoded. Subtract the deletes falling in each row group's visible range instead. Adds a test module to helpers.rs, which had none.

jackylee-ch:parquet-limit-deleted-rows

1 day ago

Latest Branches

0%

feat(io): serve additional OpenDAL filesystems through the opendal Python package#7452

13 hours ago

c892482

Lucas61000:issue-7309

0%

feat: infer image mode for casts to mode-unknown Image type#7435

3 days ago

73e3a19

Lucas61000:issue-7433

0%

fix(distributed): surface background errors while input is pending#7396

1 day ago

72b5de5

zhouwenjiajia:bugfix/joinable-stream-background-error

© 2026 CodSpeed Technology

Home Terms Privacy Docs