Benchmarks

Find measured runtimes for your operation below. The tables summarize full Polars queries on synthetic inputs. Open a report for input patterns, comparisons, and limitations.

Interval queries and native Polars comparisons

Speedups compare the fastest equivalent native Polars query tested on the same input. Several reports show native Polars matching or beating the package on many tiny groups, such as ten or fewer rows per group.

Operation	Purpose	Measured headline
Overlap counting	Count overlaps for each row	At least 1.3× faster than Polars · under 250 ms for 3M rows
Coverage depth and resource demand	Measure active intervals or their total demand	At least 1.2× faster than Polars · under 96 ms for 1M rows
Per-query coverage statistics	Count source records and measure covered length for every query	At least 8.1× faster than Polars · under 480 ms for 1M queries against 1M sources
Clustering, union, and gaps	Label connected intervals, merge ranges or find gaps	At least 1.1× faster than Polars · under 16 ms for 100k rows
Subtraction and intersection	Remove or intersect covered ranges across two collections	Under 41 ms for 1M total input rows · slower than Polars on empty-only grouped inputs
Containment counting	Count containments for each row	At least 1.2× faster than Polars · under 1.1 s for 3M rows
Nesting depth	Measure the longest containment chain above each row	Under 210 ms for 3M rows (four Polars threads) · native comparison unmeasured

Optimizers without a built-in Polars solver

These reports show optimizer runtimes. Lane balancing is a heuristic for balance; the other operations return an exact optimum for their stated objective.

Operation	Purpose	Measured headline
Lane assignment	Assign the fewest lanes without overlaps	Exact · under 180 ms for 1M rows
Lane balancing	Improve row-count balance using the fewest lanes	Improved balance on 108/132 datasets · under 25 ms for 100k rows (one Polars thread)
Weighted scheduling	Select nonoverlapping intervals with maximum total weight	Exact · under 270 ms for 1M rows
Maximum-weight clique	Select mutually overlapping intervals with maximum total weight	Exact · under 150 ms for 1M rows (one Polars thread)
Capacity scheduling	Maximize weight under an overlap limit	Exact · under 120 ms for 1M rows in small independent overlap groups
Scheduling with changing capacity	Maximize weight under an overlap limit that changes over time	Exact · under 170 ms for 1M rows in small independent overlap groups
Minimum covering	Cover a target with the fewest intervals	Exact · under 70 ms for 1M rows
Minimum-cost covering	Cover a target at minimum total cost	Exact · under 270 ms for 1M rows
Minimum-cost dominating set	Select cheapest representatives so every interval is selected or overlaps one	Exact · under 18 ms for 100k rows (one Polars thread)
Minimum stabbing points	Find the fewest points that hit every interval	Exact · under 140 ms for 3M rows
Maximum covered length	Maximize covered length with a limited selection	Exact · under 690 ms for 1M rows, selecting up to 64 intervals

Hardware

Published runs use a fixed machine: AMD Ryzen 9 3900X, 12 physical cores / 24 logical processors, 32 GiB RAM, running Windows 11 x86-64. The original hardware records remain in each run's metadata for provenance.

Benchmark setup and reporting describes how to reproduce or contribute results.