[FabianHofmann]

Note

The following content was generated by AI.

CSRStore.aggregate (the kernel behind sparse groupby().sum(), DenseStore.aggregate and, on #1122, LinearExpression.from_coo) spent most of its unsorted-rows time in scipy's CSR row gather source[order]. When every source row holds exactly one term — the common case from from_labels / 1.0 * var — the stable cell order of the rows is also the order of their entries, so data/indices are now gathered with plain numpy fancy indexing and the existing indptr is reused. Multi-term rows and sorted rows keep their existing paths. Output is identical (arrays and dtypes) to before.

Related: #1124 (unifying the two grouped-sum kernels is out of scope here), #1122.

case	before	after
2M single-term rows → 200k cells, unsorted	101 ms	63 ms (1.6x)
500k rows × ~4 terms, unsorted	26 ms	23 ms (unchanged path)
2M single-term rows, sorted	15.6 ms	15.0 ms (unchanged path)
sparse (1.0*flow).groupby(bus).sum(), 10 × 200k → 20k	94 ms	61 ms (1.55x)

Approaches measured and rejected

• Entry-level counting sort (np.repeat(rows, row_nnz) → _grouped/coo_to_csr): 1.36x on single-term rows but 2.1x slower on multi-term rows (grouping over nnz instead of rows).
• numpy arange+repeat entry gather for multi-term rows: slower than scipy's row gather.
• np.argsort(kind="stable") instead of _grouped: ~6x slower.

Correctness: randomized old-vs-new equivalence check (800 cases; int32/int64 index dtype, empty rows/grid, absent cells, NaN consts, explicit zeros, duplicates) — all identical; new test for stable term order with single- and multi-term unsorted rows.

---

Generated by Claude Code

[codspeed]

Merging this PR will not alter performance

✅ 181 untouched benchmarks
⏩ 181 skipped benchmarks¹

---

_{Comparing perf/aggregate-unsorted (11cbdea) with master (22a1882)}

Footnotes

1. 181 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩