Confirmed against the real deployed server: Cloudflare Tunnel's free
tier caps request bodies at ~100MB and returns a plain 413 - a direct
LAN upload of the same 250MB test file succeeds, the identical request
through yaris.ceaf.dev does not. Real dashcam clips average ~320MB, so
no dashcam video could ever complete an upload through the tunnel as
it stood, regardless of the earlier memory-streaming fix.
Server: new /api/v1/dashcam/upload/chunk (accepts one chunk at a time,
keyed by relative_path+sha256+chunk_index/total_chunks, reassembles and
verifies the sha256 once every chunk has arrived) and
/api/v1/dashcam/upload/status (reports which chunk indices a given
upload already has, or "already_uploaded" if it's done) endpoints.
Chunks are staged under STORAGE_DIR/.upload_tmp/<device>/<sha256>/ and
cleaned up after reassembly; a stale partial set left by a differently
-chunked earlier attempt is detected via a small marker file and wiped
before restarting. The original single-shot /upload endpoint is left
in place unchanged.
Client (modules/app_upload.py): _upload_dashcam_files() now calls
_upload_video_chunked(), which checks /status first (so a prior
partial upload resumes instead of restarting from byte 0) and then
POSTs any missing 40MB chunks (DASHCAM_CHUNK_SIZE, comfortably under
Cloudflare's cap) via the same streaming MultipartEncoder approach as
before.
Co-authored-by: Copilot <[email protected]>
The upload loop previously always slept the full 30s UPLOAD_INTERVAL
between batches, even when a full batch was just sent (meaning more
rows were likely still queued). With GPS/OBD rows logged every 5s
locally, an offline/stale device could build up a large backlog that
took a very long time to drain at 100 rows/30s.
- Bump TRACKING_BATCH_SIZE 100 -> 500 (cuts round-trips per backlog).
- Only sleep the full UPLOAD_INTERVAL once actually caught up (a
short batch came back); while a full batch keeps coming back, drain
back-to-back with a short CATCHUP_INTERVAL (1s) instead.
- server/app.py: add a sane MAX_TRACKING_BATCH_POINTS (5000) cap on
the tracking/batch endpoint as a defensive guard against a runaway
client, independent of the Pi's own batch size.
Verified against the real deployed server: a simulated 1250-row
backlog now drains in 3 batches (~0.35s server round-trip time) instead
of 13 batches spread across ~6.5 minutes of idle sleeping.
Co-authored-by: Copilot <[email protected]>
