Eventual-Inc/Daft - CodSpeed

Daft

Blog Docs Changelog

Performance History

Latest Results

style(core): apply rustfmt to decimal inference div/mul arms The shorter bound_decimal128 arm body makes the surrounding match fully formattable, so rustfmt now breaks the `.or_else` chain and wraps the sibling Decimal128 arms that previously exceeded the width limit. This matches what CI's fmt hook produces.

hello-peter-tang:fix/7601-decimal-precision-overflow

6 hours ago

style(parquet): apply cargo fmt to count-only positional-deletes test Rustfmt breaks the read_parquet(...).await.unwrap() chain (63 chars, over chain_width 60) and collapses the Int32Array::from vec literal. Fixes the failing style (fmt) CI check on #7591.

hello-peter-tang:fix/7560-count-only-positional-deletes

9 hours ago

fix(core): bound decimal arithmetic precision at 38 instead of failing to plan Decimal add/sub/mul/div inference hard-errored whenever the ideal result precision exceeded 38, so ordinary chained decimal arithmetic failed to plan. This breaks TPC-H Q1 (`price * (1 - discount) * (1 + tax)`, precision 61) and Q9 (subtraction, precision 39): Cannot infer supertypes for multiply on types: Decimal[38,4], Decimal[23,2] result precision: 61 exceed bounds of [1, 38] Decimal128 is backed by an i128, so 38 digits is a hard physical limit; the ideal type simply is not representable. Instead of erroring, cap the precision at 38 and shrink the scale only as far as needed to stay valid (precision >= scale >= 0, precision >= 1) via a shared bound_decimal128 helper. This mirrors Daft's own decimal aggregation supertypes (try_sum_supertype / try_product_supertype use Decimal128(_, s) => Decimal128(38, s)) and the behaviour of Spark and DuckDB. Division also preserves the quotient's integer digits when bounding: its max(6, p1 + s2 + 1) term inflates the scale far beyond what the operands need, so keeping that scale under a 38 cap would leave zero integer digits and null out every whole-number result. Results that already fit within 38 digits are unchanged. No kernel change is required because evaluation casts both operands to the inferred output type, so the Decimal128 kernel inherits the bounded dtype. Fixes #7601

hello-peter-tang:fix/7601-decimal-precision-overflow

9 hours ago

docs(sql): remove CROSS JOIN example from SELECT page

easyrev:feat/sql-cross-join

2 days ago

Merge branch 'main' into perf/dense-integer-argsort

GENG-CHOVYYYY:perf/dense-integer-argsort

2 days ago

refactor(io): reuse Rust Hive helpers for ORC partitions Use private Python bindings to infer Hive partition schemas and convert directory values with the existing Rust helpers. Keep ORC missing-key nulls and directory-value precedence, and retain typed constants through pruning, serialization, and batch injection. Signed-off-by: jiangxt2 <jiangxt2@vip.qq.com>

jiangxt2:orc-hive-partitioning

3 days ago

refactor(io): unify ORC schemas and constant projection Pass Daft schemas from ORC sources to tasks and readers, and use a local fill map for path literals in the existing batch projection. Extend coverage for same-named physical fields, task projection and serialization, and remote filtering and row counts. Signed-off-by: jiangxt2 <jiangxt2@vip.qq.com>

jiangxt2:orc-file-path-column

3 days ago

test(io): make ORC subprocess check explicit Set check=False while retaining explicit exit-code and output assertions. Signed-off-by: jiangxt2 <jiangxt2@vip.qq.com>

jiangxt2:orc-ignore-corrupt-files

3 days ago

Latest Branches

0%

fix(core): bound decimal arithmetic precision at 38 instead of failing to plan#7603

6 hours ago

5d63784

hello-peter-tang:fix/7601-decimal-precision-overflow

0%

fix(parquet): apply positional deletes in count-only fast path#7591

9 hours ago

8dd350e

hello-peter-tang:fix/7560-count-only-positional-deletes

0%

feat(sql): support CROSS JOIN syntax#7573

2 days ago

409e560

easyrev:feat/sql-cross-join

© 2026 CodSpeed Technology

Home Terms Privacy Docs