Eventual-Inc/Daft - CodSpeed

Daft

Blog Docs Changelog

Performance History

Latest Results

fix(execution): keep intermediate pipelines alive after empty input flush

ConeyLiu:codex/fix-empty-input-flush

6 hours ago

test(udf): verify allowed async batch calls overlap Coordinate two single-row calls with an asyncio Event so serialization fails deterministically instead of satisfying only the upper bound.

ConeyLiu:codex/fix-async-batch-concurrency

7 hours ago

fix(udf): enforce async batch concurrency before spawning tasks Wait for a free execution slot before admitting each RecordBatch task. Add regression coverage for explicit concurrency limits, fragmented inputs, and batch sizes of one and eight.

ConeyLiu:codex/fix-async-batch-concurrency

8 hours ago

Merge branch 'main' into feature/python-io-extensions

rohankumardubey:feature/python-io-extensions

8 hours ago

feat(sql): support ANSI SUBSTRING syntax (#7577) ## Changes Made SUBSTRING only accepted the comma form, and only with all three arguments: daft.sql("SELECT substring(s FROM 2 FOR 3) FROM df") # unsupported daft.sql("SELECT substring(s, 2) FROM df") # unsupported The planner keyed on sqlparser's `special` flag, which only records how the call was written. All shapes now go through one arm: a missing start defaults to 1 (as discussed in the issue) and a missing length reads to the end of the string. `SUBSTRING(s)` with neither is a planner error. Left as is: - empty results come back as NULL rather than `''`, which is the substr kernel and already covered by #7385 / #7503 - `SUBSTR(s, 2, 2)` still treats start as 0-based, unlike `SUBSTRING` AI usage disclosure: Claude Code helped write the tests. I made the planner change myself and verified everything locally before opening this PR. ## Related Issues Closes #7555

main

1 day ago

Merge branch 'main' into feature/python-io-extensions

rohankumardubey:feature/python-io-extensions

1 day ago

Detect all-null nested literals inside the existing pass Track whether any row produced a value while building the tensor, sparse tensor and map arrays instead of re-scanning the results.

chiliec:fix/nested-all-null-udf-concat

2 days ago

build: bump Ray to 2.59 (#7567) ## Changes Made - Allow Ray 2.59.x through the optional dependency range (`>=2.11.0,<2.60.0`) and pin the development dependency/lockfile to 2.59.0. - Preserve `os.environ` in the S3 credentials test and Moto fixture by restoring only the AWS variable they change. Rebinding the environment to a plain dict breaks Python/native environment synchronization, which interferes with Ray 2.59's default token authentication. - Keep Ray authentication enabled; no production-code or CI authentication overrides. Validation: on macOS/Python 3.13, the ordered S3 credential-refresh → Ray dataset tests pass with default authentication (7 passed, 14 existing runner/version skips). Independent cleanup/authentication probes and changed-file lint/format checks pass. Full CI results are pending; see the investigation comment for evidence and limits. ## Related Issues None.

main

2 days ago

Latest Branches

0%

fix(execution): keep intermediate pipelines alive after empty input flush#7589

6 hours ago

e117716

ConeyLiu:codex/fix-empty-input-flush

0%

fix(udf): enforce async batch UDF concurrency limits#7588

8 hours ago

927c48d

ConeyLiu:codex/fix-async-batch-concurrency

0%

feat(io): add async Python IO extensions#7564

9 hours ago

fc4faf4

rohankumardubey:feature/python-io-extensions

© 2026 CodSpeed Technology

Home Terms Privacy Docs