4.3 KiB
Backend deployment and Gitea Actions
Repository: ssh://git@git.my-rm.com:2222/Robot-Market/mqtt-5-test.git. Deployment branch: main. Target: existing test server 109.122.252.129, separate Compose project mqtt5-test-backend, runtime /opt/test-mqtt-5/backend. Broker remains in /opt/test-mqtt-5/deploy.
The backend uses TLS to the existing broker, accepts raw test messages and sends receipt ACKs. No simulator is deployed. Private credentials and CA are provisioned on the server; never include ESP32 ZIP, private JSON, local .env or certificates/keys in Git. Event history is in memory and resets when the backend restarts.
Access the UI
Run on your local computer and keep it running:
ssh -N -L 8080:127.0.0.1:18080 root@109.122.252.129
Open http://127.0.0.1:8080. If local port 8080 is occupied, use -L 8081:127.0.0.1:18080 and open http://127.0.0.1:8081. The server listens only on 127.0.0.1:18080. No public firewall port is required. Stop a local receipt observer while using this server backend to avoid two independent receipt ACK publishers.
CI/CD
.gitea/workflows/backend.yml runs on main pushes, main pull requests and manual dispatch. The test job builds the Docker test target (unit tests plus vet); remote integration tests are opt-in and skipped without private handoffs. The deploy job runs only on main, never for a PR. It builds an immutable UTC timestamp plus commit tag, then runs server/deploy.sh. No registry or SSH deployment secret is needed: the existing ubuntu-latest runner uses this server's Docker socket.
The runner configuration allows /opt/test-mqtt-5/backend in container.valid_volumes; only the deploy job requests that mount as /runtime. The prior runner configuration is backed up before this addition. Existing frontend job mounts remain unchanged. This workflow assumes that label continues to use the same server Docker daemon; moving the runner requires updating the deployment arrangement.
deploy.sh locks the runtime directory, validates Compose, preserves the previous image-tag file and Compose, deploys only backend, and checks container health for up to 60 seconds. Health requires /healthz to confirm current MQTT readiness, independent of bounded event-history eviction. Failed updates restore the previous release; a failed first deployment stops the backend. Images are retained for rollback. No host-wide image prune is run. Remote device return-path checks should still be performed after significant MQTT changes.
Server operations
Standard Linux commands on the server:
cd /opt/test-mqtt-5/backend
docker compose --env-file .env -f compose.yaml ps
docker compose --env-file .env -f compose.yaml logs -f backend
docker compose --env-file .env -f compose.yaml restart backend
Manual rollback to the preserved release:
cd /opt/test-mqtt-5/backend
cp .env.previous .env
cp compose.previous.yaml compose.yaml
docker compose --env-file .env -f compose.yaml up -d --force-recreate backend
During deployments do not run a simultaneous manual rollback. Before changing broker credentials, update the server's private secrets/backend.json (root:10001, mode 0640, parent directory 0750) and restart backend. Do not rotate the broker's existing credentials as part of ordinary backend deployments.
Deployment verification — 2026-10-10
The deployed observer was healthy and MQTT-ready. Independent device clients passed MQTT3.1.1 and5 over both plaintext and TLS, receiving receipt ACKs for raw text, arbitrary JSON and empty messages. Exact incoming/outgoing evidence was verified in the SERVER UI through the private SSH tunnel. Native response/correlation, controlled wire DUP and auth/TLS rejection checks passed. Duplicate local observer remained stopped. This is independent test-client evidence, not physical ESP32 firmware proof.
For repeatable checks against a tunnel, use:
rtk python3 scripts/check_receipts.py --credentials esp32-test-kit/connection.txt --ca esp32-test-kit/ca.crt --viewer-url http://127.0.0.1:18081
The script allows bounded waiting for UI evidence (default15 seconds), reports transient HTTP failures and still fails if matching evidence never appears. The old tunnel timeout was not reproduced; no historical root cause is claimed. Broker restart/persistence verification remains a separate Sprint5 gate.