hot-paths
GitHubAG Charts运行时热路径分析技能，用于识别影响性能的关键文件。通过检测代码变更中的循环、DOM操作等信号，评估对渲染帧率的潜在影响，辅助代码审查和性能优化决策。
Trigger Scenarios
Install
npx skills add ag-grid/ag-charts --skill hot-paths -g -y
SKILL.md
Frontmatter
{
"name": "hot-paths",
"targets": [
"*"
],
"description": "AG Charts runtime hot paths: which files run per datum, per frame or per interaction, the invariants each tier must hold, and how to evidence a performance claim. Use when reviewing a change that touches series, scene, data-model, scale, interaction or DOM code, when asked whether a diff will cost performance, or when deciding which benchmark to run."
}
AG Charts Hot Paths
A chart renders up to a million datums and repaints at 60Hz, so the same edit costs nothing in one file and 80ms of frame time in another. This skill says which files are which, what each tier must hold, and how to evidence a claim rather than assert it.
It is a review and impact-analysis guide. For implementing an optimisation, use
/ag-charts:optimize-series — the full playbook with the BarSeries reference.
Run the detector first
node tools/hot-paths/detect.js --range <base>...<head> # exact revisions, fork point to head
node tools/hot-paths/detect.js --pr <number> --summary # a PR, vs its merge-base
node tools/hot-paths/detect.js --base latest --summary # the working tree, untracked files included
Three dots resolve the fork point of the head from the base — what a branch or PR
actually contains. Two dots compare the two revisions as given, which charges the
change for whatever the base has merged since. Prefer --range with resolved SHAs
when reviewing: --pr re-resolves the head, so a push landing mid-review moves it.
JSON on stdout is the default; --summary swaps it for the human-readable form.
The exit status is 0 whatever the verdict, so read the triggered boolean rather
than testing the status.
Two stages, both mechanical:
- Changed non-test
.tsfiles are matched against the tier globs intools/hot-paths/hot-path-index.json. - Each hit is scored: added lines carrying loops, TypedArrays, per-datum or per-frame entry points, update-type or DOM calls; changes landing within 20 lines of an existing hot-path marker comment; and — the sharpest signal — added lines that execute inside a loop in the head revision.
The JSON output carries the tier, the frequency claim, the invariants for the tiers
hit, and the evidence handles. Read triggered; the invariants come with it, so there
is no need to restate the tier table from here. The shape:
range string the revisions compared
changedSourceFiles number non-test .ts files in the range
maxScore number highest stage-2 score across the hits
stage1, stage2 object { files: number, tiers: number[] }
triggered boolean stage 2 cleared the threshold — the verdict
hits[] object { file, tier, tierId, addedLineCount, score,
signals: { <name>: [{ line, needle, where, text }] },
markers: [...], loops: [...] }
invariants object keyed by tierId: { frequency, invariants[] }
evidence object benchmark trigger, workflow, base, noise filters,
size limits, memory guard, local profiling, playbook
where is added for a line the change introduced and context for one it did
not — a context loop header is what makes an added line per-datum, so quote it
as evidence but never cite it as the changed line.
Calibration at the current threshold, over the last 48 PR merges: stage 1 alone matches 67% of them — the tier globs are deliberately broad — and the stage-2 scoring narrows that to 29%, catching 10 of 13 known performance commits. The scoring is what makes the pass affordable, so re-derive both rates after changing any weight; do not tune the threshold on a single PR.
The four tiers
| Tier | Runs | Where |
|---|---|---|
| 1 | per datum, per SERIES_UPDATE — up to 1e6 |
chart/series/**, enterprise/src/series/**, core utils/aggregation.ts, utils/data/**, utils/geometry/** |
| 2 | per animation frame, per node | scene/**, motion/**, chart/marker/**, series/**/*Node.ts, series/**/*Shape.ts |
| 3 | per pointer move, data transaction or chart update | chart/interaction/**, chart/data/**, chart/axis/**, chart/tooltip/**, scale/**, dom/**, enterprise/src/features/**, chart.ts |
| 4 | once per bundle load | module/**, module-bundles/**, main.ts — a size-limit question, not a benchmark one |
ChartUpdateType (ag-charts-core/src/types/updateType.ts) is the spine to reason
along: FULL → UPDATE_DATA → PROCESS_DATA → PROCESS_DOMAIN → PROCESS_RANGE → PERFORM_LAYOUT → PRE_SERIES_UPDATE → SERIES_UPDATE → PRE_SCENE_RENDER → SCENE_RENDER.
Every stage a change pulls the pipeline back to runs everything downstream of it.
Regressions this repo has actually shipped
The five recurring shapes, each with a real commit — check for these before anything else:
- Full rescan where an incremental path existed.
26d7f84244(epoch-column rescan on incremental updates),f459a39824,1126e7a21c. Look for code that ignoresprocessedData.changeDescriptionor re-derives a column cache from scratch. - Un-gated per-datum work.
7695550125(type-tracker called for every datum in a column whose type was already settled),3842ae946c(tick-skipping now gated on whether the axis can be picked). Look for a call inside a loop whose result the caller may not need. - Too coarse an update type.
3c4bbf73c8(FULL→PERFORM_LAYOUTfor selection state),0fe6d15b8f(redundantcommitPendingTransactionsinperformUpdate). - Allocation reintroduced into a loop a previous commit made allocation-free. The
// scratch object,// avoid allocations,// allocation-freecomments are tripwires — the detector's marker proximity exists for exactly this. - Scale and tick generation cost.
2e6274835e(UnitTimeScaletick generation regression, caught only in release testing).unitTimeScale.tshas nine perf touches.
How to write the finding
State the frequency, then the cost, then the evidence. A hot-path finding that says "this may be slow" is worth less than no finding, because it spends the author's attention without directing it.
P1 — Per-datum allocation in
barSeries.createNodeData()barSeries.ts:412allocates a{ x, y }literal inside the datum loop, which runs once per datum (1e6 onhigh-perf-bar). The surrounding code uses a pre-allocated scratch object for exactly this reason (barSeries.ts:237). Mutate the scratch object instead. Evidence:/benchmarks high-perf-bar high-perf-bar-stacked— Initial Load is the affected measure.
Severity: P1 for a new per-datum or per-frame cost, or an incremental path degraded to a full one. P2 for a tier-3 cost, a widened update type, or a missing-but-available fast path. P3 for tier-4 bundle growth inside budget.
Evidence, and what it is worth
Get the benchmark set from the map rather than guessing — it is derived from the examples on every run, so it cannot go stale:
node tools/hot-paths/benchmark-map.js --for <changed,paths> # → recommended examples + /benchmarks command
node tools/hot-paths/benchmark-map.js --list # every example, with its series/axis types and test cases
It recommends only examples that actually run the changed series, and returns a note instead of a command when none does — most non-cartesian series have no benchmark example. A clean run on a benchmark that never executed the change is worse than no evidence, so profile locally in that case rather than substituting the nearest example.
/benchmarks <example...>as a PR comment runs the browser suite. It compares the PR head against the merge-base, not the base-branch tip, so it measures the PR's own diff.compare-browser-results.jsonly calls a change a regression once it clears--min-pct,--min-abs-msandNOISE_BAND_FACTOR ×the measured cross-run CV — so do not read a sub-noise-band percentage as a result in either direction..size-limit.jsin each package holds the src and dist budgets. This is the instrument for tier 4; benchmarks say nothing about it.chartHeapMemory.test.tsis the existing memory guard, retried 5× because heap comparisons are flaky./ag-charts:benchmark-profilecaptures a local before/after profile;/ag-eng:profile-analysereads it without opening DevTools. Use these when the benchmark moves but the cause is not obvious from the diff.
What not to flag
- Cost outside tiers 1–3 — a slow path that runs once per chart creation is not a finding, and calling it one trains authors to ignore the pass.
- Micro-optimisations with no measurement behind them (
forover.mapin code that runs twice,+=over template strings outside a loop). - Allocation in code guarded by a flag that is off in the measured configuration — check the guard before raising it.
- A benchmark delta inside the noise band. Say "no measurable change" rather than reporting the number.
- The absence of a benchmark run, on a PR the detector scores at the bottom of the range. Recommend the run; do not gate on it.
Version History
- ab3e30b Current 2026-08-27 19:30
