mqtt-5-learning/lessons/broker-operations.md

Broker operations — maintain and deploy your lab

Use this alongside Lesson 1, then revisit it during Section 9. All commands run from /home/ryan/Media/Code/Code/robot-market/mqtt-5-learning and target the mqtt5-learning project. These are instructions to perform in your lab; writing this guide has not changed or started a broker.

Lifecycle: know which state you are changing

Operation	Example	Effect
Stop	docker compose -p mqtt5-learning stop broker	Gracefully stops the container; retains it and its mounts
Start	docker compose -p mqtt5-learning start broker	Starts that existing container; does not apply edited Compose settings
Restart	docker compose -p mqtt5-learning restart broker	Stops and starts the existing container; interrupts clients and rereads broker startup configuration
Apply Compose changes	docker compose -p mqtt5-learning up -d broker	Reconciles the service definition and recreates the container when needed; keeps the named volume
Force recreation	docker compose -p mqtt5-learning up -d --force-recreate broker	Creates a fresh container even if Compose detects no change; useful after replacing a bind-mounted file
Remove containers/network	docker compose -p mqtt5-learning down	Stops and removes project containers/network; the declared named data volume remains
Delete data too	docker compose -p mqtt5-learning down -v	Also removes declared named volumes; this erases lab broker persistence

-p mqtt5-learning selects the Compose project name, grouping this lab's service, network, and volume resources. It does not select a port. Keep it identical across all commands. Your existing compose.yml is valid; the lesson's compose.yaml is an alternative filename, not an additional required file.

After a password-file repair, explicitly restart an existing container and inspect logs --since=1m broker so historical failures are distinguished from the new attempt. up -d may leave an unchanged running container alone. Read-only entrypoint ownership warnings do not establish failure; verify the broker's own listener/error messages and authenticated messaging.

Use restart after editing broker configuration that requires a restart. Use up -d after editing ports, image tags, mounts, or container environment in Compose. Docker's restart documentation explains that restart does not apply Compose changes. For single-file bind mounts, an editor that replaces a file rather than writing in place can leave an existing container referring to the old file; recreation establishes fresh mounts.

After every change, inspect ps, inspect logs --tail=50 broker, and repeat an authenticated publish/subscribe check. These check container state, broker startup, and actual messaging respectively. logs -f broker follows new logs until Ctrl+C; Ctrl+C stops log following, not the detached broker.

Choose the active credential store before maintenance

The basic lesson uses a host bind mount. This workspace's current compose.yml uses the external mqtt5-learning_broker-secrets volume because host permission changes did not stick. Inspect the service mount first. For this workspace, replace -v "$PWD/lab/secrets:/work" in the rotation commands below with -v mqtt5-learning_broker-secrets:/work. Update and back up the active volume; modifying the preserved host copy has no effect on the broker. See Lesson 1's filesystem fallback for provisioning and access checks. External secret volumes survive Compose down -v, but declared nonexternal broker-data volumes do not.

Rotate a lab credential

Use the Lesson 1 utility command without -c to update the existing backend user:

docker run --rm -it --user 0 -v "$PWD/lab/secrets:/work" eclipse-mosquitto:2.0.22 mosquitto_passwd /work/passwords backend
docker run --rm --user 0 -v "$PWD/lab/secrets:/work" eclipse-mosquitto:2.0.22 chown mosquitto:mosquitto /work/passwords
docker run --rm --user 0 -v "$PWD/lab/secrets:/work" eclipse-mosquitto:2.0.22 chmod 600 /work/passwords
docker compose -p mqtt5-learning restart broker

The first command prompts for a new password and replaces only this user's hash. The next commands restore broker ownership and owner-only access. Restart reloads the file and disconnects existing clients. Update the backend terminal's password environment variable, reconnect, and test that the new password succeeds and the old one fails on a fresh connection. Simply editing a password file does not establish that already-connected clients have been revoked.

Mosquitto also supports reloading password-file configuration with SIGHUP; see its authentication guide. Restart is the simpler explicit lab operation. For deployed systems, plan client credential rollout and reconnection before rotation to avoid unexpectedly stranding devices. Adding/removing a user also requires reviewing that user's ACL entries. mosquitto_passwd -D /work/passwords username removes a user; -D means delete, so use it only for an account you intend to remove.

Backup and restore: what actually needs saving?

Save three things: the Compose/broker/ACL configuration, the protected password file, and the broker data volume. The Go application's in-memory records are a fourth, separate concern and are not covered by this backup. Password hashes and any TLS private keys remain sensitive backup material.

For a simple consistent lab backup, stop the broker gracefully first so it can write persistence state. Confirm ps reports it stopped. Use docker volume ls to list volumes and identify the project's broker-data volume; ls lists names, not contents. Verify it with docker volume inspect <actual-volume-name> before using it. Replace placeholders with the observed name; do not infer a production volume from an example name.

Create a private backup directory outside Git, archive the stopped volume with a temporary utility container mounting it read-only, and save configuration plus secrets with their permissions preserved. A typical volume-copy utility needs a source mount and a backup-output mount; run it with the permissions required to read the broker's files. The data volume can contain mosquitto.db, while credentials live in the separate host secrets directory. Copying only one of them is incomplete.

Restore into a separate empty lab volume while the target broker is stopped. Restore ownership/permissions, connect that volume to an isolated broker using the recorded image/configuration, and verify authentication, ACL denial, retained state, and session behavior. Do not extract an archive over an active broker data directory. A backup is useful only after a restore exercise succeeds; record the commands, actual volume name, date, image version, and observed recovered state in your journal.

Upgrade with an observable rollback

Record the current image version and back up configuration, secrets, and stopped broker state. Read the target release's migration notes. Test the new pinned image against a restored copy in a separate lab project with a different loopback port; repeat authentication, ACL, reconnect/session, and retained-state checks before changing the primary lab.

docker compose -p mqtt5-learning pull broker downloads the image selected by the current Compose file; it does not replace the running container. After selecting a tested new image tag, up -d broker applies it. Keep the previous tag and backup: restoring an older image over data rewritten by a newer version is not a guaranteed rollback. For reproducibility beyond a version tag, record the resolved image digest as well.

Preparing an actual remote deployment

The initial lesson exposes only loopback plaintext, so remote devices cannot connect. Do not merely change the bind address to make it reachable. Work through Section 11 before remote use: create a server certificate with the correct DNS/IP identity, configure a TLS listener, mount certificate/key files with broker-readable permissions, and make clients validate the CA and server identity. A password hash on disk does not encrypt credentials traveling over plaintext MQTT.

Choose the device-facing hostname, TLS port, firewall policy, per-device credentials/ACLs, persistent storage, backup/restore process, log retention, monitoring, and credential renewal responsibilities. Add a deliberate container restart policy such as restart: unless-stopped to the service if the deployment should recover after daemon/host restarts; the initial lab omits it to keep stop/start experiments explicit. A restart policy helps process availability but does not prove messaging readiness or provide high availability.

Container logs need a bounded retention/rotation policy so a busy broker cannot fill the host disk. Monitor connection failures, permission failures, disk use, and application heartbeat/command evidence. Test broker readiness through an authorized synthetic message round trip; a TCP port check alone cannot prove MQTT authentication or subscriptions work.

Keep provisioning repeatable with documented setup steps or a reviewed script, use the platform's secret provisioning mechanism, and plan certificate/password rotation and client rollout. Do not bake credentials or private keys into the image. Deployment-specific values belong in that environment's configuration, not in the learning project's committed secrets.

Advanced Docker operations: predict the MQTT consequence

Inspect the configuration without starting anything

docker compose -p mqtt5-learning config -q
docker compose -p mqtt5-learning ps -a
docker compose -p mqtt5-learning logs --since=2m --tail=100 broker
docker volume inspect mqtt5-learning_broker-data

config -q validates the resolved Compose model without printing it or contacting MQTT. ps -a includes stopped containers, helping distinguish "never created" from "exited on startup." The log filters restrict evidence to recent lines; they do not clear old logs. volume inspect shows volume metadata and its Docker host mountpoint, not the broker's session contents. The volume name above is the expected default for this project: if inspection fails, use docker volume ls and identify the actual lab volume rather than creating one to hide the mismatch.

Ask three separate questions: is the container running, did Mosquitto load the intended files, and can the authorized clients exchange packets? These commands answer only parts of that chain. Finish with the authenticated CLI round trip from Lesson 1.

Understand execution context

docker compose -p mqtt5-learning exec -T broker mosquitto_pub --help

This runs the image's CLI in the existing broker container. -T disables the pseudo-terminal, useful for scripted output. --help is passed to mosquitto_pub; it performs no publish. When adapting the actual CLI test to this form, use container port 1883, not host port 18883. A new utility container started with docker run has a different loopback interface and does not reach the broker merely by using 127.0.0.1.

Isolate the session-persistence comparison

For Lesson 9's nonpersistent comparison, create compose.session-test.yaml:

services:
  broker:
    volumes:
      - broker-data-session-test:/mosquitto/data
volumes:
  broker-data-session-test:

First stop the original lab. Temporarily set persistence false in its broker configuration, then inspect and apply the override:

docker compose -p mqtt5-learning stop broker
docker compose -p mqtt5-learning -f compose.yml -f compose.session-test.yaml config
docker compose -p mqtt5-learning -f compose.yml -f compose.session-test.yaml up -d --force-recreate broker

Use the actual primary filename if yours is compose.yaml. The later -f file overrides the data mount by its container destination /mosquitto/data, while keeping configuration and credential mounts. Verify that merged result before up. --force-recreate replaces the container so the changed file/mount is actually used; it does not delete the original data volume. Use an unused comparison volume name on repeat experiments so previous state cannot contaminate the result.

After measuring, stop the comparison with the same two -f files, restore persistence true, and run docker compose -p mqtt5-learning -f compose.yml up -d --force-recreate broker. This restores the original data mount. Inspect logs and verify authentication plus messaging again. The comparison volume remains for inspection; no volume deletion is required.

Crash is different from graceful stop

For a deliberate broker crash experiment only:

docker compose -p mqtt5-learning kill -s SIGKILL broker
docker compose -p mqtt5-learning start broker
docker compose -p mqtt5-learning logs --since=2m broker

SIGKILL bypasses the broker's graceful persistence flush and interrupts every lab client. start runs the same container with the same mounted data; inspect what was actually recovered. Compare against stop, which permits graceful shutdown before Docker's stop timeout. A named volume survives both, but only state already written to it can be recovered after the crash. Do not infer crash durability from the existence of the volume.

These commands target the broker, so they test broker recovery. To test a device Will, terminate the simulator instead and keep the broker alive, as explained in Lesson 10.

Practice and explain

1. Stop and start the lab; explain which resources survived.
2. Rotate one lab password and test both passwords on fresh connections.
3. Recreate the container and confirm the data volume remains mounted.
4. Perform an isolated backup/restore exercise and record what was recovered.
5. Explain what would change for TLS, remote access, automatic restart, and log retention.

Read the official Compose up, Compose down, and Mosquitto configuration documentation when planning changes. Their latest pages may describe features beyond this lab's pinned Mosquitto 2.0.22; check version compatibility before adopting them.