feat: Use true division for the `/` operator in Polars SQL by Kevin-Patyk · Pull Request #27391 · pola-rs/polars

Kevin-Patyk · 2026-04-23T14:36:32Z

Resolves #25396.

I saw you self-assigned @alexander-beedie, but figured I would give it a stab since it's been awhile. I hope that's okay 😃!

🤖 Claude Sonnet 4.6 for navigating and explaining existing code.

codecov · 2026-04-24T09:32:13Z

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 81.46%. Comparing base (1e3a7b4) to head (51a2aea).
⚠️ Report is 5 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main   #27391      +/-   ##
==========================================
+ Coverage   79.88%   81.46%   +1.58%     
==========================================
  Files        1840     1840              
  Lines      256051   256054       +3     
  Branches     3180     3180              
==========================================
+ Hits       204546   208595    +4049     
+ Misses      50680    46634    -4046     
  Partials      825      825

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

Kevin-Patyk · 2026-04-27T11:44:11Z

As a random thought, @alexander-beedie do you think maybe we should add a warning that this behaviour changed for the end user?

alexander-beedie · 2026-05-14T12:53:01Z

As a random thought, @alexander-beedie do you think maybe we should add a warning that this behaviour changed for the end user?

SQL support is still marked as being technically "unstable" so it's not required to do so, but is probably good to note it with a .. versionchanged:: in the associated SQL operator docs. I'll try to take a look at this at the weekend and get it merged if everything looks good 👍

Kevin-Patyk · 2026-05-14T15:12:25Z

@alexander-beedie Okay so I updated the documentation is py-polars/docs/source/reference/sql/operators/index.rst. I put the version as 1.41.0 for the change. Thanks! 😄

Reclassify 873 baseline entries from sqlite/select{1,2,3}.slt that fail only due to integer division divergence. Polars SQL deliberately uses true (float) division for `/` (PR #27391), while SQLite truncates int/int toward zero. These records are now marked with `skipif polars` instead of being tracked in the expected_failures baseline. Conversions: - select1.slt: 168 converted (98 baseline → 95 after renumbering) - select2.slt: 151 converted (84 baseline → 83 after renumbering) - select3.slt: 554 converted (416 baseline → 416 after renumbering) Baseline: 1610 entries → 733 entries Total test records: 10091 (unchanged) Pass rate: 92.0% All conversions verified with cargo run -p polars-sqllogictest (exit 0, zero unexpected FAILs/PASSes). Updated UPSTREAM documentation with new skipif cluster counts. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Vendors a curated subset of the public-domain SQLite sqllogictest corpus (gregrahn/sqllogictest mirror, pinned commit c67f97bf3ca7e590d12e073408bcacaf2ff0f3a0, fetched 2026-07-21) into crates/polars-sqllogictest/slt/sqlite/: select1-4.test, evidence/{slt_lang_aggfunc,slt_lang_createview,slt_lang_dropview,in1,in2}.test, and 4 truncated random/{select,aggregates}/slt_good_{0,50}.test files (150 records each after setup). select5.test is dropped: every one of its 732 records is a 4-to-64-way implicit join stress test ("join-N-M"), and without join reordering polars-sql attempts a literal N-way cross product that is combinatorially intractable past ~N=20 - bisection confirmed the harness process gets OOM-killed inside a `join-24-4` record. A hang has no representation in the expected_failures.txt ratchet, so the file can't be used as conformance signal; see UPSTREAM for the full drop list (triggers, REPLACE/UPDATE, index existence-tracking cases) and the other vendoring-time transformations (inline-comment stripping so the sqllogictest-rs 0.29.1 parser accepts onlyif/skipif lines, hash-threshold backfill, and reflowing one-value-per-line result blocks to one-row-per-line to match this crate's default validator). Shim extensions (crates/polars-sqllogictest/src/{setup,engine}.rs), minimal, harness-only: - untyped CREATE TABLE columns (sqlparser DataType::Unspecified) default to Int64 - CREATE INDEX / DROP INDEX are unconditional no-ops - CREATE VIEW registers the view's SELECT as a LazyFrame via SQLContext, with existence tracking so duplicate-create and drop-of-nonexistent-view error correctly; DROP VIEW unregisters it - a single "statement ok" record containing multiple setup statements (occurs once in the corpus) now runs each statement in sequence instead of bailing 160 records get `skipif polars` (intentional divergences, each noted in UPSTREAM): 135 for Polars' true (float) division for `/` vs SQLite/C-style truncating int division (pre-existing decision, #27391), 17 for Postgres-aligned native booleans vs SQLite's integer-as-boolean convention, 8 for rejecting empty `IN ()` lists (SQL92/Postgres reject them; SQLite permits them as an extension). Verification: `cargo run -p polars-sqllogictest --release` exits 0 with 4829 passed / 5200 expected-fail / 10029 total (48.2% pass rate); slt/polars/ alone is 96/96 (100%); slt/sqlite/ alone is 4733/9933 (47.6%). Removing one baseline entry was spot-checked to turn the run red. cargo fmt --check clean. Ranked triage (failure records -> Phase 3 mapping): 1484 scalar/EXISTS subquery used as a comparison operand -> 3a subqueries 1309 self-join / multi-table FROM: duplicate column names -> new (context.rs across joined tables collide in projections/with_columns join handling, adjacent to 3a) 878 3+ chained/mixed UNION/EXCEPT/INTERSECT in one query -> 3e set-ops (process_except_intersect) 709 EXISTS subquery outside simple WHERE position -> 3a subqueries 655 subquery comparison operators (>, <, = ...) rejected -> 3a subqueries (sql_expr.rs:567-577) 135 integer division semantics -> skipif (intentional) 74 SubPlan not allowed outside equality context -> 3a subqueries 25 DISTINCT inside SUM/MIN/MAX/AVG -> 3c function library 17 boolean vs 0/1 representation -> skipif (intentional) 11 TOTAL() aggregate function missing -> 3c function library 8 empty IN() list -> skipif (intentional) 8 IN-list element is an arbitrary expression, not literal -> 3e small semantics 8 unary negation of COUNT() wraps via unsigned overflow -> new (correctness bug, not dialect-specific) 7 shim: INSERT ... SELECT not supported (VALUES only) -> harness limitation 7 IN/NOT IN with NULL: should return NULL, not false/true -> 3e small semantics 6 SQLite "NOT INDEXED" query hint (proprietary syntax) -> no fix planned 6 SQLite dynamic type-affinity edge cases (aggfunc.slt) -> not fixable (static typing, like Postgres) 3 SUM(<literal>) doesn't multiply by row count -> new (correctness bug) 3 SUM/MIN over all-NULL or empty group returns 0, not NULL -> 3c / new 2 hex string literal (x'...') -> 3c function library 2 query expected to fail but our engine succeeds -> baseline (low value) 2 shim: bit-shift literal in INSERT VALUES unsupported -> harness limitation 1 misc one-off join constraint error -> baseline Two items above ("self-join duplicate column names", "COUNT negation wraparound") are not on the pre-existing Phase 3 list and are flagged here as new, high-value findings - the duplicate-column one alone accounts for more baseline records than any single subquery cluster. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Reclassify 873 baseline entries from sqlite/select{1,2,3}.slt that fail only due to integer division divergence. Polars SQL deliberately uses true (float) division for `/` (PR #27391), while SQLite truncates int/int toward zero. These records are now marked with `skipif polars` instead of being tracked in the expected_failures baseline. Conversions: - select1.slt: 168 converted (98 baseline → 95 after renumbering) - select2.slt: 151 converted (84 baseline → 83 after renumbering) - select3.slt: 554 converted (416 baseline → 416 after renumbering) Baseline: 1610 entries → 733 entries Total test records: 10091 (unchanged) Pass rate: 92.0% All conversions verified with cargo run -p polars-sqllogictest (exit 0, zero unexpected FAILs/PASSes). Updated UPSTREAM documentation with new skipif cluster counts. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

feat: Use true division for /

dcb9e1d

Kevin-Patyk requested review from alexander-beedie, c-peters, orlp and ritchie46 as code owners April 23, 2026 14:36

update: Add Python files

245854e

Kevin-Patyk requested a review from MarcoGorelli as a code owner April 23, 2026 14:37

Kevin-Patyk changed the title ~~feat: Use true division for / operator in SQL~~ feat: Use true division for the / operator in SQL Apr 23, 2026

Kevin-Patyk changed the title ~~feat: Use true division for the / operator in SQL~~ feat: Use true division for the / operator in Polars SQL Apr 23, 2026

github-actions Bot added A-sql Area: Polars SQL functionality enhancement New feature or an improvement of an existing feature python Related to Python Polars rust Related to Rust Polars labels Apr 23, 2026

Kevin-Patyk and others added 3 commits April 23, 2026 17:00

update: Fix doc test

0bae193

update: Fix Rust test

2d2cec7

Merge branch 'main' into refactor/use_true_division_sql

d45b8b4

Kevin-Patyk added 6 commits April 28, 2026 11:10

Merge branch 'main' into refactor/use_true_division_sql

09e6696

Merge branch 'main' into refactor/use_true_division_sql

b7ccec2

Merge branch 'main' into refactor/use_true_division_sql

31ceaf3

Merge branch 'main' into refactor/use_true_division_sql

27116fd

Merge branch 'main' into refactor/use_true_division_sql

0fe1546

Merge branch 'main' into refactor/use_true_division_sql

2f75c18

alexander-beedie and others added 2 commits May 14, 2026 16:53

Merge branch 'main' into refactor/use_true_division_sql

765ed50

update: Adjust docs and example

dfc12a6

Merge branch 'main' into refactor/use_true_division_sql

51a2aea

nameexhaustion approved these changes May 18, 2026

View reviewed changes

nameexhaustion merged commit 18e3422 into pola-rs:main May 18, 2026
33 checks passed

Navigation Menu

Uh oh!

feat: Use true division for the `/` operator in Polars SQL - #27391

feat: Use true division for the `/` operator in Polars SQL#27391
nameexhaustion merged 14 commits into
pola-rs:mainfrom
Kevin-Patyk:refactor/use_true_division_sql

Kevin-Patyk commented Apr 23, 2026

Uh oh!

codecov Bot commented Apr 24, 2026 •

edited

Loading

Uh oh!

Kevin-Patyk commented Apr 27, 2026

Uh oh!

alexander-beedie commented May 14, 2026 •

edited

Loading

Uh oh!

Kevin-Patyk commented May 14, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Uh oh!

Conversation

Kevin-Patyk commented Apr 23, 2026

Uh oh!

codecov Bot commented Apr 24, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

Kevin-Patyk commented Apr 27, 2026

Uh oh!

alexander-beedie commented May 14, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Kevin-Patyk commented May 14, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

codecov Bot commented Apr 24, 2026 •

edited

Loading

alexander-beedie commented May 14, 2026 •

edited

Loading