Files
cuframes-composer/docs/en/operations.md
T
gx e76360dbc4 docs: руководства пользователя / разработчика / operations (RU+EN)
Полный комплект документации к Phase 11b:

  docs/ru/user.md        — для админа инсталляции (motion-mode, PTZ,
                            templates.json, mqtt_overlays.json, ZMQ verbs)
  docs/ru/developer.md   — архитектура (Cell / Layout / Decoration),
                            как добавить новый Cell/Decoration, ABI shim,
                            algorithms (best-fit + asymmetric hysteresis)
  docs/ru/operations.md  — build (host + jammy + incremental bake),
                            deploy, logs/telemetry, troubleshooting
                            (broken pipe, MQTT-overlay, motion-mode)
  docs/en/*.md           — английская версия всех трёх
  README.md              — переписан с overview + ссылками на docs/

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
2026-06-04 22:02:47 +01:00

13 KiB

cfc-grid — operations / deploy / troubleshooting

Audience: who builds, deploys, monitors cfc-grid in production.

If you're a user — see user.md. For developer documentation — see developer.md.

1. Production setup (R9-88.23)

1.1 Stack

docker compose -f docker-compose.yml \
               -f cuda-grid/docker-compose.override.yml \
               -f cuframes-composer/docker-compose.override.yml \
               -f onvif/docker-compose.override.yml \
               up -d

Files — in localhost-infra/hosts/R9-88.23/docker/cctv/.

Service Image Purpose
cuframes-ipc-anchor gx/cuframes:0.4 Shared VMM IPC anchor for cuframes
cuframes-pub-* (parking/back_yard/front_yard/gate_lpr) gx/cuframes:0.4 RTSP → cuframes per-camera publishers
cuda-grid-mediamtx bluenviron/mediamtx RTSP/HLS/WebRTC gateway
cctv-mosquitto eclipse-mosquitto MQTT broker (+bridge to 192.168.88.4)
cfc-grid gx/cuframes-composer:0.11b-step1 Composer (main service)
cfc-grid-ffmpeg ffmpeg-vf-cuda-grid:phase4b-final H.264 pipe → RTSP push
cfc-grid-watchdog gx/cuda-grid-watchdog:0.4 Restart cfc-grid on stuck inboundBytes
cctv-onvif gx/cctv-onvif:0.6 ONVIF discovery + PTZ → ZMQ
cctv-frigate ghcr.io/blakeblackshear/frigate Object detection → MQTT events

1.2 Frame flow

cuframes-pub-X ──VMM──┐
cuframes-pub-Y ──VMM──┼──→ cfc-grid (composer)
cuframes-pub-Z ──VMM──┘         │
                                 │ H.264 NVENC
                                 ↓ named pipe /tmp/cfc-pipe-dir/grid.h264
                            cfc-grid-ffmpeg (re-mux)
                                 │ RTSP push
                                 ↓
                            cuda-grid-mediamtx
                          rtsp://*/cfc-grid (TCP/UDP)
                          http://*:8888/cfc-grid (HLS)
                          http://*:8889/cfc-grid (WebRTC)

1.3 Networks

  • Internal docker network: cctv
  • External ports on R9-88.23:
    • 554/tcp — RTSP (mediamtx)
    • 8888/tcp — HLS (mediamtx)
    • 8889/tcp — WebRTC (mediamtx)
    • 5599/tcp — ZMQ composer control plane
    • 8085/tcp — ONVIF SOAP (cctv-onvif)
    • 3702/udp — WS-Discovery multicast (cctv-onvif)

2. Build

2.1 Local host build (Ubuntu 24.04, dev machine)

cd /home/claude/projects/cuframes-composer
cmake -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build -j$(nproc)

Artifacts in build/src/libcuframes_composer.so and build/examples/grid_record.

IMPORTANT: host binary (Ubuntu 24.04, glibc 2.39, libavformat60) is incompatible with runtime container (Ubuntu 22.04 jammy, glibc 2.35, libavformat58). See memory incremental-ffmpeg-rebuild.

2.2 Jammy build (for production image)

Uses cached builder container cuframes-composer-builder:cached (Ubuntu 22.04 + nvidia/cuda:12.4.1-devel + apt-deps):

cd /home/claude/projects/cuframes-composer

# If builder not yet cached:
docker image inspect cuframes-composer-builder:cached >/dev/null 2>&1 || {
  docker run -d --name cfc-builder-tmp \
    nvidia/cuda:12.4.1-devel-ubuntu22.04 sleep 3600
  docker exec cfc-builder-tmp bash -c '
    apt-get update -qq && apt-get install -y -qq --no-install-recommends \
      build-essential cmake git pkg-config \
      libpng-dev libfreetype-dev \
      libzmq3-dev libjson-c-dev libmosquitto-dev \
      libavformat-dev libavcodec-dev libavutil-dev'
  docker commit cfc-builder-tmp cuframes-composer-builder:cached
  docker rm -f cfc-builder-tmp
}

# Actual build:
docker run --rm --gpus all -v "$PWD":/src -w /src/build-jammy \
  cuframes-composer-builder:cached \
  bash -c 'cmake -DCMAKE_BUILD_TYPE=Release .. && make -j$(nproc)'

Artifacts in build-jammy/.

2.3 Bake image (incremental — without docker build)

We don't use docker build (4GB CUDA pull on cache miss). Instead:

docker rmi gx/cuframes-composer:0.11b-step1 -f 2>/dev/null
CID=$(docker create gx/cuframes-composer:0.10)
docker cp build-jammy/examples/grid_record "$CID":/usr/local/bin/grid_record
docker cp build-jammy/src/libcuframes_composer.so.0.1.0 \
          "$CID":/usr/lib/x86_64-linux-gnu/libcuframes_composer.so.0
docker cp docker/templates.json "$CID":/opt/templates.json
docker cp docker/mqtt_overlays.json "$CID":/opt/mqtt_overlays.json
docker commit \
  --change 'ENTRYPOINT ["/usr/local/bin/grid_record"]' \
  --change 'CMD ["--help"]' \
  --change 'ENV NVIDIA_DRIVER_CAPABILITIES=compute,utility,video' \
  "$CID" gx/cuframes-composer:0.11b-step1
docker rm "$CID"

Uses gx/cuframes-composer:0.10 as base (with runtime deps already installed) + overlays fresh artifacts. Faster and no network traffic.

2.4 Build ONVIF image

cd hosts/R9-88.23/docker/cctv/onvif
docker build -t gx/cctv-onvif:0.6 -f Dockerfile .

Python image, lightweight. If you change server.py — rebuild image (bump tag) + update docker-compose.override.yml.

3. Deploy

3.1 Production (R9-88.23)

cd /home/claude/projects/localhost-infra/hosts/R9-88.23/docker/cctv
docker compose -f docker-compose.yml \
               -f cuda-grid/docker-compose.override.yml \
               -f cuframes-composer/docker-compose.override.yml \
               -f onvif/docker-compose.override.yml \
               up -d cfc-grid

Compose auto-recreates container if image tag changed in docker-compose.override.yml.

3.2 Verify post-deploy

docker logs --tail 30 cfc-grid 2>&1 | grep -iE "loaded|template|pool|motion"

# Expect something like:
#   [cfc/loader] /opt/templates.json: loaded 9 templates
#   [cfc/composer] templates loaded: 9 (path='/opt/templates.json')
#   [cfc/composer] pool+ 'cam-parking' (frigate=parking_overview prio=100)
#   [cfc/composer] motion_mode=1 ttl=45000ms pool=4
#   [cfc/composer] grow → template='tpl_3' active=3

3.3 Rollback

sed -i 's|gx/cuframes-composer:0.11b-step1|gx/cuframes-composer:0.10|' \
  hosts/R9-88.23/docker/cctv/cuframes-composer/docker-compose.override.yml
docker compose ... up -d cfc-grid

4. Logs

4.1 Live tail

docker logs -f --tail 50 cfc-grid
docker logs -f --tail 30 cfc-grid-ffmpeg
docker logs -f --tail 30 cuda-grid-mediamtx
docker logs -f --tail 30 cctv-onvif
docker logs -f --tail 30 cctv-frigate

4.2 Telemetry patterns

Marker Meaning
[grid_record] N kadrov, M IDR, X MB ... Composer successfully encoding every ~50 frames
[cfc/composer] grow → template='X' New template applied (growth, immediate)
[cfc/composer] shrink → template='X' New template applied after hysteresis (shrink)
[cfc/composer] manual override 'X' до +60000ms PTZ via ONVIF
[cfc/composer] manual override expired, возврат в motion-mode Auto-return after 60s
[cfc/mqtt-overlay/<id>] '<text>' MQTT-overlay received/rendered new text
[cfc/frigate] connected, subscribe 'frigate/events' Frigate subscriber connected

4.3 When something breaks

Symptom Where to look
src active=0 stale=0 dead=4 cuframes-pub-* containers; check docker ps and network access to cameras
overlay 0 draw failed cfc_overlay_text_rebuild_atlas — usually invalid font or text
RTSP stream not delivering cfc-grid-ffmpeg logs; see §6.1
TV/ONVIF can't find cctv-onvif logs; check multicast WS-Discovery in LAN

5. Monitoring

5.1 MQTT health

cfc-grid publishes health to cuda_grid/health/composer/cfc-grid every ~10 seconds:

{
  "uptime_s": 3600,
  "frames_encoded": 90000,
  "fps_actual": 25.0,
  "bitrate_kbps": 6000,
  "src_active": 4,
  "src_stale": 0,
  "src_dead": 0,
  "idr_count": 1
}
PW=$(grep '^COMPOSER_MQTT_PASSWORD=' \
  /home/claude/projects/localhost-infra/hosts/R9-88.23/docker/cctv/.env | cut -d= -f2)
mosquitto_sub -h 192.168.88.23 -u composer -P "$PW" \
  -t 'cuda_grid/health/composer/cfc-grid' -v

5.2 Watchdog

cfc-grid-watchdog is a separate service, monitors mediamtx inboundBytes for cfc-grid path. If 30 seconds of silencedocker restart cfc-grid.

Watchdog logs:

docker logs --tail 30 cfc-grid-watchdog

On trigger — publishes to cuda_grid/health/watchdog/cfc-grid.

6. Troubleshooting

6.1 RTSP not delivering / cfc-grid-ffmpeg "Broken pipe"

Symptom: docker logs cfc-grid-ffmpeg shows [out#0/rtsp] Task finished with error code: -32 (Broken pipe).

Cause: --intra-refresh in composer (no IDR bursts), mediamtx tears RTSP publisher when it can't deliver start-frame to a new client.

Treatment:

  • Full pipeline restart:
    docker compose ... restart cfc-grid-ffmpeg cfc-grid cuda-grid-mediamtx
    
  • If recurring — disable --intra-refresh in compose override (cost: IDR bursts in bitrate, but more stable for downstream clients with frequent disconnect/reconnect)

6.2 ffmpeg doesn't receive frames from RTSP

Symptom: ffmpeg -i rtsp://192.168.88.23:554/cfc-grid -frames:v 1 out.jpg hangs for 30+ seconds.

Cause: Composer writes H.264 without regular IDR (intra-refresh). A new RTSP client waits for a keyframe to start decoding. ffmpeg in default config doesn't wait long enough.

Workaround:

ffmpeg -rtsp_transport tcp \
       -analyzeduration 10000000 -probesize 10000000 \
       -i rtsp://192.168.88.23:554/cfc-grid \
       -frames:v 1 -y out.jpg

Or use HLS:

ffmpeg -i http://192.168.88.23:8888/cfc-grid/index.m3u8 \
       -frames:v 1 -y out.jpg

6.3 MQTT-overlay not updating

Checklist:

  1. Bridge to HA broker (192.168.88.4) working?

    docker logs cctv-mosquitto 2>&1 | grep -i 'bridge'
    

    Look for Connecting bridge ha-bridge and connect confirmation.

  2. Required topic in bridge config?

    docker exec cctv-mosquitto grep 'topic.*in 0' /mosquitto/config/mosquitto.conf
    

    If new prefix — add topic XXX/# in 0 and restart mosquitto.

  3. Subscriber connected?

    docker logs cfc-grid 2>&1 | grep 'mqtt-overlay/<id>.*connected'
    
  4. Test publish:

    mosquitto_pub -h 192.168.88.4 -t '<your topic>' -m 'test' -r
    

    In composer logs, should appear [cfc/mqtt-overlay/<id>] 'test'.

6.4 Motion-mode not switching layouts

Checklist:

  1. Frigate sending events?

    mosquitto_sub -h 192.168.88.23 -u composer -P "$PW" \
      -t 'frigate/events' -C 3
    
  2. Composer receiving events?

    docker logs cfc-grid 2>&1 | grep 'frigate.*started\|grow\|shrink'
    
  3. Camera-name matches? frigate=<name> in --source must match event.after.camera.

  4. Zone-filter not blocking? If zones=A:B:C in --source — check Frigate event current_zones. If empty or doesn't intersect — pulse is discarded.

  5. TTL not expired? Logs motion_ttl=45000 (45 sec) — if events come less frequently — camera drops from active.

6.5 ONVIF PTZ presets empty in TV

Cause: TV cached old GetPresets response (Phase 9 names).

Treatment: delete and re-add camera in TV client.

6.6 Templates loaded but motion-mode doesn't use new

Composer reads global registry cfc::current_templates() on every frame — changes via cfc_layout_load_file (ZMQ or CLI) should be picked up immediately. If not — check:

echo '{"cmd":"list_layouts"}' | python3 -c "
import zmq,json,sys
s = zmq.Context().socket(zmq.REQ)
s.connect('tcp://192.168.88.23:5599')
s.send_json({'cmd':'list_layouts'})
print(json.dumps(s.recv_json(), indent=2, ensure_ascii=False))"

source field shows currently loaded path. If built-in (only tpl_1 + tpl_4) — JSON didn't load (syntax error, wrong path).

7. Configs in repo

What Where
templates.json cuframes-composer/docker/templates.json
mqtt_overlays.json cuframes-composer/docker/mqtt_overlays.json
compose override localhost-infra/hosts/R9-88.23/docker/cctv/cuframes-composer/docker-compose.override.yml
ONVIF config localhost-infra/.../onvif/onvif.yaml
ONVIF server localhost-infra/.../onvif/server.py
Mosquitto config localhost-infra/.../cctv/mosquitto/config/mosquitto.conf
.env (passwords) localhost-infra/.../cctv/.env (gitignored)

After changing compose override — docker compose ... up -d cfc-grid auto-recreates.

8. Known limitations / TODO

  • --intra-refresh ↔ RTSP clients: trade-off bitrate vs latency (see §6.1)
  • Watchdog only for cfc-grid: cfc-grid-ffmpeg in zombie state not detected directly; only full restart helps
  • Hot-reload of mqtt_overlays.json: no ZMQ verb
  • Per-overlay MQTT broker config: all via single broker; for foreign broker — need to extend MqttBrokerCfg per-item

9. See also

  • user.md — composer configuration
  • developer.md — internals, adding modules
  • memory/host-and-project.md — general R9-88.23 infra
  • memory/project_cfc-grid-deployed.md — first prod deploy
  • memory/project_cfc-grid-cpp-refactor.md — Phase 11b refactor
  • memory/incremental-ffmpeg-rebuild.md — incremental docker recipe