Bench results

These numbers anchor every claim on the homepage. Generated by running the same dataset through apps/bench — pretable against three real third-party grids: AG Grid Community, TanStack Table v8 with TanStack Virtual, and MUI X DataGrid Community. We measure scroll frame timings, blank gaps under scroll, row-height fidelity for wrapped text, and anchor stability across rebuilds. All raw data is in the repo; this page renders one milestone JSON file.

Runset 2026-05-09t00-49-34-903z · Chromium · S2 (3,000 rows, wrapped multilingual messages) · hypothesis scale · 3 repeats per adapter. Pretable / MUI scroll p95 also re-measured at 20 repeats — parity confirmed (see below).

H1 — wrapped-text scroll

Scenario S2(3,000 rows, wrapped multilingual messages, scroll script). Frame p95 measured via Performance Observer; row-height error measured by comparing each adapter’s rendered DOM row height against the engine’s planned height. Lower is better for every column.

On this dataset, pretable and MUI X DataGrid Community sit at parity on scroll frame p95: a 20-repeat re-measurement gives pretable 9.07ms ± 0.20 and MUI 9.13ms ± 0.19. The single-repeat snapshot above (pretable 9.7ms vs 8.7ms) read as a small MUI lead; under tighter sampling that gap is well inside the 2σ noise floor (≈ 0.40ms). Both adapters clear the single 60Hz frame budget with zero blank gaps and ≤ 1px row-height drift.

AG Grid Community and TanStack Table land at roughly twice that frame p95 (~16.7ms) and both drop a blank gap during the scripted scroll; AG Grid additionally drifts 2px on row height, a sign that wrapped-cell layout doesn’t round-trip through its line-height pipeline as cleanly as pretable’s text-core does.

The honest read: pretable’s wedge on this script isn’t a raw frame-speed lead over the best comparator — it’s parity on raw frame timing, with the combination of zero blank gaps, zero anchor shift, and ≤ 1px row-height fidelity at full-grid feature weight. MUI matches pretable on quality but ships a fundamentally different feature surface (no headless engine, no streaming primitives, no theming-as-data). The H1 evaluator marks this run satisfied after the high-repeat correction; the original failing verdict from the n=3 snapshot was a low-sample artifact, not a real regression.

Interactions (sort, filter)

Scenario S2 (3,000 rows, wrapped multilingual messages). Sort applies a column-state change; filter-metadata applies an equals filter on a metadata column; filter-text applies a contains filter on the wrapped-text primary column. Latency measured from trigger to first changed frame; lower is better.

Pretable sorts and filters 3,000 wrapped-text rows in 17–18 ms across all three scripts — about a millisecond over the single 60Hz frame budget on every interaction script, but 2–3.5× faster than every measured comparator. AG Grid Community runs sort and filter 3–3.5× slower; MUI X DataGrid Community lands at roughly 2× slower across all three scripts. TanStack Table v8 + TanStack Virtual runs ~2× slower on average, with high variance on filter-metadata (samples span 8–42 ms at n=8).

High-repeat (n=20) follow-up confirms pretable is reliably 1–2 ms over budget on all three scripts (sort 17.10 ± 1.83 ms; filter-metadata 17.51 ± 2.44 ms; filter-text 16.79 ± 0.31 ms). The comparative wedge is the story, not absolute single-frame compliance — pretable’s wrapped-text filter pipeline is on the roadmap for a perf-fix pass. Like the scroll story, the H6/H7/H8 evaluators check pretable’s absolute thresholds (≤ 32 ms interaction p95) rather than comparator parity; all three hypotheses stay satisfied.

Streaming

H13–H15 (streaming update frame budget, operating envelope, and row stability under stream) require S5 and rate-tagged update runs; this matrix run is S2 only, so those hypotheses are reported as insufficient rather than re-evaluated. The last full streaming evidence remains in the milestone archive.

Methodology

• Same dataset, same script. Each adapter runs the canonical S2 scenario from @pretable-internal/scenario-data through identical interaction plans. The harness is in @pretable-internal/bench-runner.
• Repeated-run medians. 3 samples per (adapter, scenario, script) tuple in this runset. Reported value is the median; full per-sample data is in the runset JSON.
• Idiomatic out-of-the-box config. AG Grid uses themeQuartz with Community modules registered, sortable + filterable + resizable defaults, and applyTransaction updates. TanStack uses useReactTable with getCoreRowModel/getSortedRowModel/getFilteredRowModel plus TanStack Virtual for row virtualization. MUI X uses DataGrid from @mui/x-data-grid with default sort/filter/resizable enabled. Pretable uses RAF-batched grid.applyTransaction via @pretable/stream-adapter. None of the adapters carry per-bench tuning — what we measure is what each library ships out of the box.
• Wedge integrity: the pretable adapter renders the same <PretableSurface> the website’s homepage hero uses. What we measure is what you ship.

Run it yourself

The bench is part of the open-source repo. Clone, install, run:

git clone https://github.com/cacheplane/pretable.git
cd pretable
pnpm install
pnpm bench:matrix \
  --project=chromium \
  --adapters=pretable,ag-grid,tanstack,mui \
  --scenarios=S2 --scripts=initial,scroll \
  --scale=hypothesis --repeats=3
# Output: status/runsets/<timestamp>.hypotheses.json

For the full contributor walkthrough, see apps/bench/README.md.

Raw data

• status/milestones/2026-05-08-b2-scroll-summary.json — per-adapter scroll medians (this page’s table)
• status/milestones/2026-05-08-b2-comparative-bench.hypotheses.json — full H1–H21 evaluator report for this runset
• status/milestones/ — milestone archive (H1 baseline, S3 column virtualization, streaming, selection/nav)