perf: reduce remaining Cmd+K input latency and worst typing tail #1228

Open
opened 2026-10-06 20:03:55 +00:00 by kayg · 0 comments
Owner

Remaining Cmd+K typing tail exceeds the instant-interaction budget

Follow-up from #1124 round 2. Five interleaved production A/B pairs, 45 real browser input events per condition, Chromium 4× CPU/macOS, perf-test under the shared lock, HDD 8 ms /200 IOPS /150 MB/s. Real 5k Files, 240 indexed API Logs and 121 CalDAV Events. Type file-0049 on the Month route. Clock is the input event to two requestAnimationFrame callbacks; it does not await the debounced network search or a spring.

Intrinsic height invalidations now share one queued frame and do not observe their own animated surface/results viewport. Input p50 177.2 →145.1 ms (18.12% gain), p95 297 →296 ms. Maximum 314.6 →354.8 ms. The p95 retention rule passes, but this is still far above DESIGN §59's frame budget, and the maximum rises 12.78%.

Separate marked typing trace: Layout 281.875 →241.839 ms, Paint 317.990 →231.075 ms. These are diagnostic costs, not latency samples. Investigate result projection/selection, intrinsic-height layout reads and shared material paint under dense real results; retain all nine inputs, real empty/error/offline states, keyboard and touch behavior. No new exception or relaxed ratchet.

Raw evidence: docs/perf/runs/2026-10-06-1124-round2-calendar-search-ab.json and before/batched phase summaries. This is a remaining performance path, not a correctness blocker.

# Remaining Cmd+K typing tail exceeds the instant-interaction budget Follow-up from #1124 round 2. Five interleaved production A/B pairs, 45 real browser input events per condition, Chromium 4× CPU/macOS, perf-test under the shared lock, HDD 8 ms /200 IOPS /150 MB/s. Real 5k Files, 240 indexed API Logs and 121 CalDAV Events. Type file-0049 on the Month route. Clock is the input event to two requestAnimationFrame callbacks; it does not await the debounced network search or a spring. Intrinsic height invalidations now share one queued frame and do not observe their own animated surface/results viewport. Input p50 177.2 →145.1 ms (18.12% gain), p95 297 →296 ms. Maximum 314.6 →354.8 ms. The p95 retention rule passes, but this is still far above DESIGN §59's frame budget, and the maximum rises 12.78%. Separate marked typing trace: Layout 281.875 →241.839 ms, Paint 317.990 →231.075 ms. These are diagnostic costs, not latency samples. Investigate result projection/selection, intrinsic-height layout reads and shared material paint under dense real results; retain all nine inputs, real empty/error/offline states, keyboard and touch behavior. No new exception or relaxed ratchet. Raw evidence: docs/perf/runs/2026-10-06-1124-round2-calendar-search-ab.json and before/batched phase summaries. This is a remaining performance path, not a correctness blocker.
Commenting is not possible because the repository is archived.
No labels
No milestone
No project
No assignees
1 participant
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
kayg/calternal-v0#1228
No description provided.