Blackbox Lab

After a crash, every FPV pilot asks the same question: was that me, or was that the quad? The flight controller records the answer — several hundred samples a second of gyro, motor output, RC input and battery voltage — but the existing desktop tools are slow, and none of them run on a phone at the field.

Blackbox Lab parses those logs in the browser. Nothing is uploaded; the file never leaves the device.

Why this was harder than it looks#

A five-minute flight is roughly [[120 MB]] of binary log and about [[600,000]] samples across [[40]] fields. Parsing that in JavaScript on the main thread freezes the tab for [[20 seconds]] or more.

The fix was to move the decoder into Rust compiled to WebAssembly, run it inside a Web Worker, and stream results back in chunks so the first chart paints while the rest of the file is still decoding.

// The worker streams decoded frames back in chunks rather than resolving once
// at the end, so the first chart can paint in well under a second on a file
// that takes several seconds to decode completely.
export async function decodeLog(
  file: File,
  onChunk: (frames: LogFrame[]) => void,
): Promise<LogSummary> {
  const worker = new Worker(new URL("./decoder.worker.ts", import.meta.url), {
    type: "module",
  });
 
  return new Promise((resolve, reject) => {
    worker.onmessage = ({ data }: MessageEvent<DecoderMessage>) => {
      if (data.kind === "chunk") onChunk(data.frames);
      if (data.kind === "done") {
        worker.terminate();
        resolve(data.summary);
      }
    };
    worker.onerror = reject;
    worker.postMessage({ file }, [file as unknown as Transferable]);
  });
}

The measurements that mattered#

Rewriting the hot path was only worth it because I measured first. The decoder was never the bottleneck I assumed it was — the original version spent most of its time allocating objects per sample.

Charts are drawn straight to Canvas rather than through a charting library. At [[600,000]] points, every library I tried either dropped frames while panning or refused to render at all.

What I learned#

• Profile before rewriting. My first instinct was the parser. The profiler said allocation. I would have spent a week on the wrong problem.
• WebAssembly is not automatically fast. The naive port was slower than the JavaScript it replaced until I removed the per-sample copies across the boundary.
• Open source brings real review. A pull request from a stranger caught a sign error in my gyro scaling that I had been flying with for months.

Status#

Actively maintained and open source. [[Roughly 400]] pilots use it monthly. Next up is support for the newer log format and a side-by-side mode for comparing two tunes on the same axes.