Skip to content

[Perf] motion(): every update resolves each element's timing context with DOM queries and scans of all points #271

Open

Description

@kyleromines

Version: @tanstack/charts 1.5.0 (b236dae).

Under motion(), an update's cost grows faster than its element count. One keyed rect mark, only y values changing, median synchronous host.update in headless Chromium 153:

Bars motion() Static SVG renderer
100 3.5 ms 0.9 ms
200 7.6 ms 1.3 ms
400 19.4 ms 2.1 ms
800 60.7 ms 4.0 ms

Compiling the scene is about 1 ms at 800 bars. A CPU profile at 400 bars puts 68% of host.update inside elementTimingContext, 26% in motionPointForKey alone.

Why

addUpdateTrack resolves a timing context for every retained element (motion.ts:2198), and createPresentationTracks again for every point that moved (motion.ts:3368-3371). Each one starts from the leaf:

A mark with n elements does about n² string comparisons for each update.

Repro

Bundle for a browser and run await bench().

import { defineChart, rect } from '@tanstack/charts'
import { motion } from '@tanstack/charts/motion'
import { mountChartRenderer } from '@tanstack/charts/renderer'
import { scaleLinear } from '@tanstack/charts/scales/linear'
import { createSvgChartRenderer } from '@tanstack/charts/svg/renderer'

const definition = (count: number, step: number) =>
  defineChart({
    marks: [
      rect(
        Array.from({ length: count }, (_, k) => ({ key: `r${k}`, x1: k - 0.4, x2: k + 0.4, y2: 40 + 30 * Math.sin(step + k) })),
        { x1: 'x1', x2: 'x2', y1: () => 0, y2: 'y2', key: 'key' },
      ),
    ],
    scales: { x: { scale: scaleLinear().domain([-0.5, count - 0.5]) }, y: { scale: scaleLinear().domain([0, 80]) } },
    pointer: false,
  })

async function measure(count: number, animated: boolean) {
  const renderer = animated ? motion({ initial: false }) : createSvgChartRenderer()
  const options = (step: number) => ({ definition: definition(count, step), renderer, width: 900, height: 300, ariaLabel: 'cost' })
  const host = mountChartRenderer(document.body.appendChild(document.createElement('div')), options(0))
  const times: number[] = []
  for (let step = 1; step <= 60; step++) {
    await new Promise(requestAnimationFrame)
    const next = options(step)
    const started = performance.now()
    host.update(next)
    times.push(performance.now() - started)
  }
  host.destroy()
  return times.sort((a, b) => a - b)[30]?.toFixed(1)
}

window.bench = async () => {
  await measure(100, true) // warm up
  for (const count of [100, 200, 400, 800])
    console.log(`${count} bars: motion() ${await measure(count, true)} ms, static SVG ${await measure(count, false)} ms`)
}

Proposal

Build a key-to-point map and a point count for each markId once for a scene, and find the owning group, role and peers on the way down the reconcile walk, not from each leaf.

Related: #186, #188, #201, PR #268.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions