Skip to content

Benchmark Results

Reference: WolfXL v1.7.0 historical baseline. Status as of: 2026-04-28.

This page is historical context, not the final WolfXL 2.0 release proof. Use Public Evidence Status for the current claim policy and release-readiness checklist.

Current claim boundary: the current supported-scope SOTA gate is green in the trust reports for the measured clean-source and clean-wheel OpenPyXL and selected-surface Rust/Rust-backed lanes. The registered cross-platform artifact matrix still has stale lanes, and the all-future-surface SOTA claim is still not ready. Do not use this historical v1.7 page as proof that WolfXL is better than every current or future Excel library in every workflow.

The benchmark numbers below summarise the v1.7 posture vs openpyxl 3.1.x. For the current harness and timestamped snapshots see ExcelBench.

Hardware context

Numbers below were collected on:

  • Apple MacBook Pro M4 Pro
  • 24 GB RAM
  • Python 3.13
  • openpyxl 3.1.5 (comparison baseline)
  • WolfXL 1.7.0

Always attach runtime details (OS, Python, package versions) when sharing new runs.

Headline numbers (v1.7 baseline)

Median of 5 runs per cell. Lower is better. Wall-clock seconds.

Read

Workload openpyxl 3.1.5 WolfXL 1.7 Speedup
1k rows × 10 cols, plain values 0.18 0.04 4.5×
10k rows × 10 cols, plain values 1.12 0.18 6.2×
100k rows × 10 cols, plain values 11.4 1.05 10.9×
100k rows, read_only=True (streaming) 9.8 0.55 17.8×
10k rows × 20 cols with styled cells 5.2 0.62 8.4×
.xlsb, 50k rows × 8 cols n/a (no openpyxl support) 0.41 —

Write

Workload openpyxl 3.1.5 WolfXL 1.7 Speedup
1k rows × 10 cols, fresh workbook 0.21 0.06 3.5×
10k rows × 10 cols, fresh workbook 1.85 0.22 8.4×
100k rows × 10 cols, fresh workbook 18.1 1.78 10.2×
10k rows × 10 cols, with styles 3.6 0.45 8.0×
1 chart × 4 series × 1k points 0.31 0.09 3.4×

Modify mode (the v1.0 / v1.1 differentiator)

For modify-mode workloads WolfXL surgically rewrites the changed parts of the ZIP and copies everything else verbatim — in contrast to openpyxl which loads the entire workbook into a Python DOM, mutates, and re-serialises. The speedup on small-edit workloads is where WolfXL is most differentiated.

Workload openpyxl 3.1.5 WolfXL 1.7 Speedup
Touch 1 cell, save (1k-row workbook) 0.42 0.03 14×
Touch 1 cell, save (100k-row workbook) 22.4 0.18 124×
Add 1 hyperlink, save (10k-row workbook) 4.1 0.09 45×
copy_worksheet (10k-row, with table + DV + CF) 3.9 0.21 18×
insert_rows(idx=2, amount=100) (10k-row sheet with formulas) 2.8 0.28 10×

Methodology

See methodology.md for the full reproduction recipe. Quick summary:

  1. Each test case is a Python script that uses time.perf_counter to bracket the operation under measurement.
  2. Numbers are the median of 5 runs, with one warm-up run discarded.
  3. Fixtures live under fixtures/excel/ in the ExcelBench repo and are regenerated deterministically.
  4. Memory is sampled via tracemalloc and reported separately (omitted from the headline tables to keep them readable).

Regenerate the reports

From the ExcelBench repo:

uv run excelbench benchmark --tests fixtures/excel --output results
uv run excelbench perf --tests fixtures/excel --output results
uv run excelbench report --input results/xlsx/results.json --output results/xlsx
uv run excelbench heatmap
uv run excelbench html

The HTML output ships to the dashboard.

What to publish with each release

  1. Fidelity summary table — every fixture round-trips bit-identical or with documented ratchet entries.
  2. Performance summary table (this page).
  3. Hardware / runtime metadata.
  4. Link to raw JSON artifacts in ExcelBench.
  5. Notable regressions and fixes since the prior release.

Where WolfXL is not yet faster

  • Tiny-workbook overhead. For a workbook with < 100 cells the Python ↔ Rust FFI cost dominates; openpyxl can be on par or marginally faster. Most relevant for unit-test fixtures, not production workloads.
  • Pure-Python computation paths (e.g. iterating in Python over every cell). The Rust backend speeds up the I/O and serialisation layers; once you're in a Python loop, you pay Python loop costs. Use iter_rows(values_only=True) (which streams in Rust) over per-cell Python iteration when the workload allows.
  • Formula recompute. WolfXL's formula engine handles 67 functions; for workbooks that depend on the long tail of spreadsheet formulas, defer recompute to Excel-on-open by preserving formulas verbatim (load_workbook(..., data_only=False)).

Run benchmarks on your own files

See run-on-your-files.md for a copy-pasteable harness covering read / write / modify mode.

Track perf in CI

WolfXL's own CI runs the tests/test_streaming_perf.py slow-marked test (100k-row streaming benchmark). Add WOLFXL_RUN_PERF=1 to your CI environment to enable it; the test asserts a wall-clock budget so regressions surface immediately.