Plot Internals Skill

Chart.js Worker Architecture

ChartRenderer.worker.ts

// Worker-side: owns the Chart.js instance
Chart.register(LineElement, PointElement, LineController, ...);

Comlink.expose({
  async init(args: InitArgs): Promise<ChartRenderer> {
    await fontLoaded;  // Wait for fonts before first render
    return new ChartRenderer(args);
  }
});

ChartRenderer Class

• Owns a single Chart instance (Chart.js)
• Receives UpdateAction messages from main thread
• Returns Bounds (data range) and HoverElement[] (interaction)
• Manages scales, datasets, annotations all within the Worker

Dataset Building

TimestampDatasetsBuilderImpl (Worker)

class TimestampDatasetsBuilderImpl {
  #seriesMap = new Map<string, SeriesData>();  // path → accumulated points

  // Called for each batch of messages
  handleMessages(messages: MessageEvent[]): void {
    // Extract numeric values using message-path logic
    // Append to per-series arrays
    // Enforce the 50,000-datum accumulation cap (MAX_CURRENT_DATUMS_PER_SERIES)
  }

  // Returns current datasets for rendering
  buildActions(): UpdateAction[] {
    return this.#seriesMap.values().map(seriesData => ({
      datasets: [{ data: seriesData.points }],
    }));
  }
}

Two Distinct Point Caps

There are two separate constants — don't conflate them:

• When current data exceeds MAX_CURRENT_DATUMS_PER_SERIES, the builder culls the oldest data via splice, dropping the overflow plus an extra MAX_CURRENT_DATUMS_PER_SERIES * 0.25 so culling isn't triggered every single append.
• The per-series render budget is MAX_POINTS / numSeries — the 5,000 is shared across all signals.

Downsampling Algorithm (NOT LTTB)

Downsampling (components/TimeBasedChart/downsample.ts) is a custom stateful pixel-interval min/max bucketing, not LTTB:

1. The viewport is divided into intervals sized by MINIMUM_PIXEL_DISTANCE = 3 pixels
2. Each interval emits up to POINTS_PER_INTERVAL = 4 points: first, last, min, and max (intFirst, intLast, intMin, intMax)
3. State (DownsampleState) is carried across batches via a cursor so streaming data downsamples incrementally without reprocessing consumed points
4. Re-triggered when the viewport/zoom changes (interval sizing depends on pixel bounds)

⚠️ Do not describe this as LTTB / "Largest-Triangle-Three-Buckets". It is interval min/max decimation keyed on pixel distance.

Range Source Flag

#hasRangeSource: boolean;
// true = preloaded data available (full time range in blocks)
// false = only current-frame data (live or non-preloading mode)

When #hasRangeSource is true:

• Builder receives preloaded block data (historical)
• Must merge block data with current-frame messages
• Datasets represent full recording, not just visible window

Scale Management

Time Axis (X)

• TimeScale from Chart.js chartjs-adapter-luxon
• Handles nanosecond timestamps (converted to milliseconds for Chart.js)
• Zoom/pan interactions update visible time range

Value Axis (Y)

• Auto-scaling based on visible data range
• Per-series Y-axis support (left/right axis)
• Manual bounds via panel settings

Interaction Handling

Hover/Tooltip

// Main thread asks Worker for elements at mouse position
const elements = await renderer.getElementsAtPixel({ x, y });
// Returns: series index, data index, value — used for tooltip rendering

Zoom/Pan

• Handled via Chart.js annotation plugin
• Zoom: scroll wheel → update time range → PlotCoordinator notifies builder
• Pan: drag → update time range → same flow

Font Loading

// Worker must load fonts before rendering (no DOM access)
const fontLoaded = loadDefaultFont();  // Fetches font file via fetch()
// Every new ChartRenderer awaits this before creating Chart instance

Performance Considerations

1. Worker isolation: Chart.js runs entirely in Worker — no main thread blocking
2. OffscreenCanvas: Canvas operations don't trigger main-thread compositing
3. Batch updates: Multiple dataset changes coalesced into single chart.update()
4. Lazy rendering: Chart only re-renders when data or options actually change
5. Transfer optimization: Dataset arrays transferred (not copied) when possible
6. Memory: 50k cap × ~32 bytes/point × N series = bounded memory usage