---
type: report
date: 2026-10-09
---
# Recorder bench test

Replaying the first in-car capture (`z27ag-20261009-163736-ign.log`, 10 s,
8925 frames) onto a `vcan0` with its original timing, against a fake gpsd on the
workstation and the real gpsd on the car Orin.

## Scenarios

| # | Scenario | Result |
|---|---|---|
| 1 | Full replay, `mark` at t≈3 s | auto session; 8925 frames; ID#DATA sequence **byte-identical** to the source; 3 chunks at `--chunk-secs 4`, all compressed |
| 2 | 1 s replay | discarded at close (under `--min-secs`, no marks) — after fixing the span to first→last frame; the first version counted the idle wait too and kept it |
| 3 | `ctl start bench-test`, no CAN, `mark`, `stop` | manual session with label, mark and GNSS; no CAN chunk left behind |
| 4 | `vcan0` deleted mid-session, recreated 1 s later | one session; `Network is down` → `No such device` → `socket open` recorded as `can_link` events |
| 5 | SIGTERM mid-session | session closed, reason `shutdown`, exit 0 |
| 6 | SIGKILL mid-session (3 s replay), restart | all 2676 frames on disk; marked `truncated` and compressed at restart |
| 7 | `--cap-gb 0` | every older closed session deleted, the active one kept; sequence numbers continue from `.last-seq` |

On the car Orin (aarch64 cross-build, highest GLIBC_2.34 against its 2.35): the
same full replay recorded 8925 frames and 55 KB of zstd; real gpsd reports (no
fix indoors) recorded with the last TPV before start stamped `before_start`.
Resident memory at idle: 5 MB.

## Not tested

- Real PCAN-USB hot-plug in the car (the `vcan` deletion stands in for it).
- Sustained throughput beyond 890 frames/s, and the queue-overflow path.
- A clock step during a session (the detection is in place; nothing stepped).

## Installed

`yukart-recorder.service` on `mikkeli-orin-nano`, enabled, running as `mikkeli`,
waiting for `can_pcan`. Recordings: `/var/lib/yukart/recordings/`.