Benchmarks

qsv's benchmark suite runs 292 benchmarks on a 1,000,000-row × 41-column, 520 MB sample of NYC's 311 data, timed with hyperfine (2 warmups + 3 timed runs each). See scripts/results/README.md for the methodology and the raw CSVs.

[!NOTE] 25.0.0 — PGO is back. The Apple Silicon prebuilt is profile-guided again (#4757), and the median benchmark held level with 24.0.0. sortcheck_unsorted is back to ~11 ms, and split_chunks_index_j1 and stats_index are back in their 23.0.1 range. extdedup and snappy_compress did not recover. They are still 19-31% slower than on 23.0.1. The big wins: writestat is 35-50x faster (#4747, which takes column types from the stats cache instead of Polars inference). Polars 2.0 (#4760) makes pivotp ~5x, joinp ~2.6x and plain sqlp ~2x faster. CSV → GeoJSON geoconvert is ~3x faster now that it streams (#4800), and template is 21% faster on MiniJinja 3.0 (#4785). On the other side, count and flatten slowed 13-17%. The 19 benchmarks new in suite 9.9.0 have no earlier baseline, so they don't appear in the change charts.

The sqlp_format_*, sqlp_lowmemory, sqlp_nooptimizations and sqlp_tryparsedates* benchmarks have no 25.0.0 numbers: a schema cache leaked from another benchmark, so their runs timed a cached-schema read (#4812). Those rows were removed and will be back on the next run. Historical rows too fast to be real were removed for good: a full scan of the 520 MB file in under 0.1 s for sqlp and frequency, or geocoding 1M rows in under 0.2 s (#4823).

Looking for the full per-command timing tables? See the classic tabular benchmarks at qsv.dathere.com. This page is the interactive, visual companion to it.

📊 Interactive Data Schematic

➡️ Open the interactive benchmark Data Schematic

The Data Schematic is rendered by qsv's own viz command and is fully interactive (hover for values, zoom, download PNGs). It covers the with/without-index advantage, the sqlp schema-cache knob, throughput across every release, deep-dives into four flagship commands (stats, frequency, validate and moarstats) showing they grew features without losing speed, and this release's biggest speedups.

Note: cross-version numbers are not strictly apples-to-apples — commands gain features over time. Treat the historical trend as a broad trajectory, not an exact speedup.

Last generated for qsv 25.0.0 (aarch64-apple-darwin). Regenerate with python3 scripts/gen_benchmark_viz.py and copy this page's link — the Data Schematic itself redeploys to GitHub Pages automatically on push to master.