@pixagram/virtualized — optimized

@pixagram/virtualized (a react-virtualized fork) with a set of behavior-preserving performance optimizations applied to the core windowing engine, the Collection spatial index, and (on by default) the Grid scroll/render path — plus a reproducible verification suite.

source/        The library source, with optimizations applied (ships the
               *.jest.js test suite too).
               - CHANGES.md ............ full changelog
               - NOTES_FURTHER_OPTIMIZATIONS.md ... remaining proposals
verification/  Differential harnesses + benchmarks proving the optimizations
               are bit-identical / DOM-identical to the original behavior.
               - README.md ............. how to run, results
               - baseline/ ............. pristine files the harnesses diff against
jest.config.js / babel.config.js   Run the project's own *.jest.js suite
                                   against this optimized tree.

Optimizations applied

Pure-logic layer (verified bit-identical to the original by randomized differential harnesses — 685,064 assertions, 0 mismatches):

1. Fixed-size fast path (CellSizeAndPositionManager) — constant integer rowHeight/columnWidth ⇒ O(1) arithmetic geometry, no arrays/search. ~15× faster visible-range queries; ~33.5 MB → 0 backing memory on 1M cells.
2. Lazy Fenwick delta allocation — point-update arrays allocated only on first resizeCell (another ~16.8 MB saved on 1M cells).
3. Non-allocating scalar accessors + renderer hoist — remove the rows×columns per-render style-input allocations.
4. CellMeasurerCache capacity seeding — optional columnCount/rowCount hints avoid column-growth relayout.
5. Count-based rescan mitigation (CellMeasurerCache) — a shrink skips the O(n) max-rescan unless the last max-holder is removed.
6. Finger (galloping) search (CellSizeAndPositionManager) — ~1.8× faster smooth scroll on a 500k variable-height list.
7. Collection de-dup without string coercion (SectionManager/Section) — getCellIndices uses a Set<number> + ascending sort instead of an object-as-set + Object.keys()/map(). ~2.45× faster on a 20k-cell collection.

React component layer (verified by render-differential and the project's own test suite):

8. Grid scroll range-gating (Grid, on by default; disable with reuseChildrenOnScroll={false}) — while actively scrolling, when the rendered cell set and positions are unchanged, reuse the previously rendered children and skip the cell-range renderer. ~38% fewer renderer calls during scroll, with byte-identical DOM. Disable it only if your cellRenderer / cellRangeRenderer have per-call side effects that must run on every scroll frame (with gating on they are not invoked on unchanged-range frames).

Third pass (10.1.0): stability + React-layer efficiency

Full details in source/CHANGES.md → "Third pass"; harnesses in verification/README.md → section 4.

Stability:

• Range-gate soundness — a custom cellRangeRenderer that positions from scrollLeft/scrollTop (documented params) was frozen during sub-cell scrolling; the gate now compares the offsets for non-default renderers.
• React 19 — no hard dependency on ReactDOM.findDOMNode anywhere (WindowScroller, Table, CellMeasurer); every component runs on React 19.3.
• ResizeObserver-first resize detection (AutoSizer, WindowScroller) — no DOM injected into your parent element, no forced position: relative, no injected stylesheet, no capture-phase scroll listener on the parent. Legacy detector kept as fallback.
• WindowScroller re-derives the child offset when the page reflows (content above the list changing size no longer misaligns it).
• Cross-frame-safe element checks; null-safe Grid.componentDidUpdate; no work after unmount in AutoSizer / WindowScroller / InfiniteLoader.

Efficiency (DOM byte-identical, verified against the pristine tree):

• CellMeasurer batches no longer wipe Grid's caches — only the entries a batch can have changed are dropped — and style identity is preserved across the whole scroll burst: PureComponent row renders 207 → 124 in a real List + CellMeasurer scenario.
• Collection per-cell style cache: cell renders 246 → 132.
• Table Children.toArray memoized (was per row); no redundant setState per update. Cell-cache hits short-circuit the range renderer.
• The "batched DOM measurement" proposal is withdrawn: a first-time batch already costs one layout, not N (see NOTES_FURTHER_OPTIMIZATIONS.md).

Tooling (fresh clone on Node 22 / npm 10): npm install (no flags) → npm test → npm run build → npm run lint all succeed. The unused extract-text-webpack-plugin (webpack 3 peer) and react-codemirror (React <16 peer) that broke npm install on npm ≥ 7 are gone, scripts use npm run, and source/styles.css — silently dropped by an unanchored .gitignore rule when the fork was created, which made npm run build / prepublishOnly fail — is restored. Details in source/CHANGES.md → "Tooling".

Verification at a glance

• 685,064 logic-layer differential assertions, 0 mismatches.
• Grid range-gating: render-differential shows enabling the flag leaves the DOM byte-identical (150 steps across fixed/variable/scaled sizes, scroll/settle, prop changes, recompute mid-scroll, real CellMeasurer cache), with the gate confirmed firing.
• The project's own *.jest.js suite passes 568/568 on this optimized tree (and identically on the original), across every component — zero regressions — on React 16 and React 18; real React 19.3 smoke passes; under preact/compat the tree passes/fails exactly the same tests as the pristine 10.0.4 tree.
• Third-pass render-differentials (pristine vs optimized): 0 DOM mismatches across all steps.

See verification/README.md to reproduce everything.

Running the project's test suite

npm i -D jest@29 jest-environment-jsdom@29 jest-jasmine2@29 babel-jest@29 \
         @babel/core @babel/preset-env @babel/preset-react @babel/preset-flow \
         @babel/plugin-proposal-class-properties \
         react@16 react-dom@16 react-test-renderer@16 clsx dom-helpers@5 immutable jsdom
npx jest                 # whole suite
npx jest Grid/Grid.jest  # just Grid (126 tests)

The project targets the jasmine2 runner (its TestUtils registers an afterEach lazily on first render, which jest's default circus runner rejects); jest.config.js sets testRunner: 'jest-jasmine2' accordingly.

Not applied (proposal)

One further optimization (opt-in transform-based positioning) remains a documented proposal in source/NOTES_FURTHER_OPTIMIZATIONS.md; it depends on real browser compositing, which the available (jsdom) test environment does not implement, so it can't be proven correct the way everything else here was.