perf(react): grouped streaming stops paying a full-set height-index pass per commit

[blove]

The diagnosis (this was a diagnosis-first investigation)

Two CDP traces of S5 group-updates (20k rows, 1000 upd/s) had shown the react layout layer's per-commit row-sequence re-ingest at ~40% of the streaming window's second half — #stepIngest balanced-sequence frames and row-model group-index reads among the top frames. #522 fixed the grouping-APPLY replacement waste and deliberately left this steady-state seam alone. This branch answers what the steady-state cost actually is.

Ingest proportionality answer: per grouped streaming commit, the row-layout controller was doing full-set work where zero extra work is achievable — and the trigger was columns, not rows:

1. CompiledQuery.get query() mints a fresh defensive copy on every read (packages/row-model/src/compiled-query.ts:1756 → snapshotQuery). Deliberate and test-pinned: an exposed Date operand a consumer mutates must not poison later reads (compiled-query.test.ts "detaches Date operands…"). So observedQuery in usePretableModelInternal changes identity on every controller publish.
2. That defeats the columns memo (packages/react/src/pretable-model.ts:835), so resolveEffectiveColumns (packages/react/src/pretable-surface.tsx) re-runs per commit — and minted the synthetic group column with a fresh value: () => "" closure each time.
3. controller.setColumns's deep-equality check (packages/renderer-dom/src/row-layout-controller.ts:2277) failed on [__pretable_group__].value alone — instrumented directly: 30 streaming commits → 30 diffs, every one only that field.
4. Each failure took the columns-reset path: RowHeightIndex.clearEstimates() (packages/layout-core/src/row-height-index.ts:1625) — a synchronous full-set in-order walk + balanced rebuild, O(rows) per commit. Before perf(renderer-dom): grouping apply pays one height-index replacement, not three #522 the same trigger started a full cooperative replacement per commit — which is exactly the #stepIngest signature the original traces caught.

The journal path itself is healthy: with a bare controller (no react layer), 60 grouped streaming commits produced zero replacements and zero fallbacks — the per-commit ingest is batch-proportional. The waste was entirely this columns seam.

The fix

Hoist the synthetic group column's value accessor to a module constant (SYNTHETIC_EMPTY_VALUE) so its identity is stable across renders. One field, and the deep-equality check now short-circuits every steady-state commit.

Not touched, and documented at the site: the row-select synth keeps its inline closure — hoisting it un-bails the react-hooks compiler analysis and surfaces a pre-existing react-hooks/refs finding at the surfaceGrid useMemo (the indexed facade carries ref-reading closures). A grid with rowSelectionColumn enabled still pays the per-commit columns reset; follow-up filed.

Numbers

Pre-fix (traced, S5 group-updates target, streaming window 3050ms):

• clearEstimates was the top react-layout frame in the window: 18.9ms self, ~4.2% of non-idle time, called once per commit via setColumns (caller chain verified through the trace's node parents: clearEstimates ← setColumns ← usePretableModelInternal's layout effect).
• jsdom pin: 20 streaming commits → columnsResetPathCount +20.

Post-fix (same recipe, same machine):

• clearEstimates absent from the streaming window's top frames; window idle 85.4% → 86.8%.

• jsdom pin: 20 streaming commits → columnsResetPathCount +0, replacementStartCount +0, all fallback counters 0, and the streamed "Bank" group sum visibly lands in the DOM (the cheap path did not buy its zeros by dropping commits).

• Bench matrix (5 repeats, S5 target, 1000 upd/s; updates is the like-for-like control, same machine, same evening):

  script	scroll_frame_p95_ms pre-fix	post-fix	frame_max_ms pre → post
  group-updates	9.90 10.10 9.90 10.00 10.00	9.90 10.00 9.90 9.90 9.90	10.3–11.1 → 10.3–10.4
  updates (control)	9.90 10.00 10.00 9.90 9.90	9.90 9.90 10.00 10.00 9.80	10.3–10.5 → 10.3–10.9

  No regression; the control's spread brackets both sides, so the p95 stays ~10.0ms (this metric sits on the frame cadence, which is why the win shows up in the trace's busy-time share rather than p95 — the fix is headroom, exactly as scoped).

Scale note: clearEstimates is O(rows) per commit, so pre-fix the seam grew linearly toward the frame budget with row count (100k = 5× the walk, plus the GC pressure of rebuilding a 100k-node balanced tree per commit). Post-fix the steady-state layout cost is journal-path only, i.e. batch-proportional.

Tests

• New end-to-end pin: packages/react/src/__tests__/grouped-streaming-layout-cost.test.tsx — mutation-verified (reverting the one-line fix turns it red with columnsResetPathCount 20 ≠ 0).
• Full suites green: react 111 files / 1568 tests, renderer-dom 168, layout-core 157. Lint, typecheck, prettier clean.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
pretable	Ignored		Aug 30, 2026 3:12am

[github-actions]

Vercel preview ready

Preview: https://pretable-e4flte5nu-cacheplane.vercel.app
Commit: 3dbb6e01d7b3dbae6ce03f23e0904c99f3638514

_{Updated automatically by the deploy-preview job.}