Full-resolution MP4 recording: VID_20261118_150957_030.mp4
Scope Note:
ghost-lapis a personal educational project that replays historical Formula 1 qualifying data from the public OpenF1 API (session_key=9554) as a live stream. It is not connected to a live car feed and is not a race-strategy tool.
The Human Question: Where Was Pole Actually Won?
At the 2024 British Grand Prix Qualifying session at Silverstone, George Russell (RUS) took pole position in 85.819 s, edging out Lewis Hamilton (HAM) at 85.990 s by +0.1710 s.
A single finish-line gap tells you who won, not where on the 5.89 km circuit the lap was won or lost. Answering that question requires a continuous delta curve across the entire lap, comparing both cars at every meter of track distance rather than by wall-clock time.
The Engineering Problem: Time Alignment vs. Spatial Alignment
Car telemetry does not arrive indexed by track distance. In the OpenF1 dataset (session_key=9554), car_data and location arrive as asynchronous timestamped streams at roughly 3.8 Hz (about every 260 ms, or roughly 18 m of travel per sample at 70 m/s):
-
RUS(Lap 25):324samples at3.74 Hz(90.0–325.0 km/h) -
HAM(Lap 24):335samples inside[0.0 s, 85.990 s]at3.88 Hz(86.0–324.0 km/h)
If you compare two drivers at identical elapsed timestamps (e.g., t = 42.100 s), you are comparing two cars at different physical coordinates on the track, especially once one driver opens a gap. To compare speed and cumulative time loss accurately, you have to transform both streams from $f(t)$ into $f(S)$, mapping both drivers onto a common 1.0 m spatial distance grid in a streaming pipeline.
Architecture
flowchart LR OpenF1["OpenF1 API\n(Session 9554, Q3)"] -->|export_telemetry.py| Data["data/*.json\n(324 RUS / 338 HAM)"] Data -->|cmd/producer| Redpanda[("Redpanda\ntopic: telemetry-raw")] Redpanda -->|cmd/consumer| Stream["DriverStreamState\nTrim [0, T_lap] & 1.0m Grid"] Data -->|cmd/observer| Metrics["Prometheus /metrics\n& /api/trace (:9090)"] Data -->|cmd/replay-web| Canvas["HTML5 Canvas Replay\n(:8080)"] Replay Producer (
cmd/producer): Reads exported Q3 laps, trims samples strictly to [0.0 s, official_lap_time] via linear boundary interpolation, and publishes chronologically interleaved records to Redpanda (telemetry-raw), partitioned by driver code (RUS,HAM).Stateful Streaming Consumer (
cmd/consumer,pkg/telemetry): Consumestelemetry-rawviafranz-go, maintains per-driver incremental state (DriverStreamState), projects 2D coordinates onto a1.0 marc-length grid, and emits incremental delta points without re-scanning prior frames.Observability and Replay (
cmd/observer,cmd/replay-web): Serves a dependency-free HTML5<canvas>replay view (:8080) and exposes bounded-cardinality Prometheus metrics (/metrics) plus aligned trace arrays (/api/trace) at:9090.
The Two Things That Should Have Stopped Me
In my first pass, two validation results looked too clean:
Zero Finish-Line Error (
0.0000 s): I had calibrated the final grid point to match the official+0.1710 sgap. That guaranteed zero error by construction, grading my own homework with the answer key.Impossible Segment Speeds (
~490 km/h): My per-segment table had one car covering960 min7.053 s. I initially wrote a note blaming pit-exit and cool-down transitions in the data, even though a flying Q3 lap has no pit exit.
When I cut each driver's export strictly to the official lap start and duration, stripped out finish-line calibration, and added a CI check requiring every band over 600 m to average between 35 and 90 m/s, the real behavior surfaced.
Distance Alignment and Resampling Choice (Stages 2 & 3)
Before building the delta curve, I compared two independent ways of computing cumulative track distance $S$ across the trimmed Q3 laps:
| Alignment Method | RUS Computed Distance | HAM Computed Distance | Difference |
|---|---|---|---|
Speed Integration (int v, dt, km/h to m/s) | 5828.112 m | 5825.070 m | 3.042 m |
2D Arc-Length Projection (X, Y decimeters to m) | 5828.725 m | 5823.350 m | 5.376 m |
Speed Integration: Converting
km/htom/sand applying trapezoidal integration across~260 mssample intervals yields5,828.1 m(RUS) and5,825.1 m(HAM), with3.042 minter-driver drift.2D Arc-Length Projection ($\sqrt{\Delta x^2 + \Delta y^2} / 10$): OpenF1 reports track coordinates in decimeters. Summing Euclidean steps in meters yields
5,828.725 m(RUS) and5,823.350 m(HAM), landing about 1% below Silverstone's official5,891 mcenterline, which is plausible for a racing line sampled at3.8 Hz.
To map non-uniform samples onto a uniform 1.0 m grid (5,824 overlapping points from 0 to 5,823 m), I used monotonic linear interpolation rather than cubic splines. Elapsed time RelTimeSec stays strictly monotonic, and binary step channels (throttle snaps and hard braking) never ring or overshoot between samples.
The Honest Number: Explaining the 85 ms Gap (Stage 4)
Without finish-line calibration, computing DeltaSec = time_ham(S) - time_rus(S) directly from each sample's timestamp and arc distance produces:
- Official Ground-Truth Gap:
+0.1710 s(RUS 85.819 svs.HAM 85.990 s) - Computed Finish-Line Delta (Uncalibrated, at
5,823 m):+0.2560 s - Absolute Error vs. Official:
0.0850 s(84.99 ms)
The table below shows where that 85 ms comes from. Both drivers are compared up to 5,823 m, the end of the shorter of the two position traces (HAM's 5,823.35 m trace, where his elapsed time is 85.984 s vs. 85.990 s official, 6 ms short). Because RUS's trace is 5,828.73 m long, comparing him at 5,823 m stops 5.7 m before his finish line. His elapsed time at 5,823 m is 85.728 s against his official 85.819 s (91 ms short). Net difference: 85 ms.
The same distance travelled is not the same place on the circuit. The two position traces differ in length by 5.376 m (a mix of racing-line choice and 3.8 Hz position sampling), and that path-length difference accounts for roughly half of the 0.171 s gap being measured.
Streaming Equivalence and Packet-Loss Stress Test (Stages 5 & 6)
To verify that incremental streaming (DriverStreamState) introduces no drift relative to batch processing, cmd/stream-check compares the Redpanda consumer's incremental output against batch resampling:
- Batch Finish-Line Delta:
+0.2560 s - Incremental Stream Delta:
+0.2560 s - Divergence:
0.000000 s(0.0000 ms)
Next, I dropped intermediate samples at 5%, 10%, and 25% using a single fixed PRNG seed (seed = 42) with start and finish boundary samples anchored:
| Loss Rate | Finish Delta (s) | Error vs. Official (ms) | Max Curve Error vs. Base (ms) | MAE vs. Base (ms) |
|---|---|---|---|---|
0% (Baseline) | +0.2560 s | 84.99 ms | 0.00 ms | 0.00 ms |
5% Loss | +0.2594 s | 88.37 ms | 131.89 ms | 3.87 ms |
10% Loss | +0.2588 s | 87.82 ms | 139.23 ms | 13.29 ms |
25% Loss | +0.3066 s | 135.63 ms | 217.24 ms | 52.41 ms |
Because this run uses a single random draw, it serves as a rough sensitivity check rather than a statistical curve: finish-line error shifts modestly at 5% and 10% loss (88 ms), before jumping to 136 ms at 25% loss as missed samples widen local interpolation gaps (217.24 ms peak curve error).
Distance-Band Breakdown and What Remains Unverified (Stage 7)
Partitioning the 5,823 m overlapping grid into seven contiguous distance bands, and asserting in CI that every band over 600 m averages between 35 and 90 m/s, gives the per-band timing breakdown:
| Distance Band | Length (m) | RUS (s) | HAM (s) | RUS Avg (m/s) | HAM Avg (m/s) | Delta (s) | Ahead |
|---|---|---|---|---|---|---|---|
Band 1 (0-780 m) | 780 | 9.518 | 9.581 | 81.95 | 81.41 | +0.0625 | RUS |
Band 2 (780-1560 m) | 780 | 15.474 | 15.482 | 50.41 | 50.38 | +0.0072 | RUS |
Band 3 (1560-2520 m) | 960 | 15.884 | 15.775 | 60.44 | 60.86 | -0.1095 | HAM |
Band 4 (2520-3180 m) | 660 | 7.972 | 7.979 | 82.79 | 82.72 | +0.0069 | RUS |
Band 5 (3180-3920 m) | 740 | 9.050 | 9.157 | 81.77 | 80.82 | +0.1064 | RUS |
Band 6 (3920-4780 m) | 860 | 10.888 | 11.018 | 78.98 | 78.05 | +0.1297 | RUS |
Band 7 (4780-5823 m) | 1043 | 16.942 | 16.994 | 61.56 | 61.37 | +0.0527 | RUS |
| Total / Sum | 5823 | 85.728 | 85.984 | 67.92 | 67.72 | +0.2560 | RUS |
Two caveats are worth stating plainly:
Conservation is arithmetic, not a physical test: The band deltas sum to
+0.2560 s(0.000000 mserror) because telescoping differences always sum to the endpoint difference.Corner boundaries are unverified: Without aligning braking-point signatures from the brake and speed channels to an official FIA corner map, fixed distance bands cannot be reliably labeled with corner names, so they stay numbered as distance bands.
Bounded-Cardinality Observability
Exposing all 5,824 spatial steps as Prometheus labels would create high-cardinality time series. Instead, cmd/observer restricts Prometheus labels strictly to bounded enums on GET /metrics, recording a 2.64 ms resample duration (ghostlap_resample_duration_seconds_sum 0.002636245), while serving the full 1.0 m trace array over GET /api/trace for Grafana visualization.
What This Project Isn't
It isn't live telemetry, it isn't a strategy tool, and it covers one pair of laps from one qualifying session. What it demonstrates is that a validation check that cannot fail is not a validation, and that when an uncalibrated check shows an 85 ms discrepancy, tracing it to the 5.4 m path-length difference explains the system far better than calibrating the error away.
Repo: github.com/aashiruu/ghost-lap
Author: Okikiola Ashiru, Cloud & DevOps Engineer

