docker-compose.balena.dev.yml.tmpl

# vim: ft=yaml

version: "2"
services:
  anthias-server:
    image: ghcr.io/screenly/anthias-server:${GIT_SHORT_HASH}-${BOARD}
    # A crash must not cost the device its storage. The host's
    # core_pattern is the kernel default `core`, a relative path, so a
    # dump lands in the crashing process's working directory, which is
    # WORKDIR /usr/src/app: inside the container, on no mount. With
    # core_uses_pid=1 every crash adds a fresh `core.<pid>` rather than
    # replacing one, and nothing sweeps them. They are invisible to
    # `docker images`, and `docker image prune` cannot touch a live
    # container's writable layer, so the usual disk investigation comes
    # back clean while the card fills. A device in the field reached
    # 24.2 GB of them in anthias-celery (issue 3417). Sentry already
    # reports crashes, so on a device a dump buys nothing and costs the
    # card. Anchored here and aliased by every service below: the rule
    # is "any of our containers", not "the one that happened to crash".
    ulimits: &no-core-dumps
      core: 0
    build:
      context: .
      dockerfile: ./docker/Dockerfile.server
    ports:
      - 80:8080
    environment:
      - HOME=/data
      - LISTEN=0.0.0.0
      - CELERY_BROKER_URL=redis://redis:6379/0
      - CELERY_RESULT_BACKEND=redis://redis:6379/0
    depends_on:
      - redis
    # No `devices:` here, and none needed. HDMI-CEC is driven by the
    # anthias-viewer container, which is privileged and therefore
    # already sees every /dev/cec* node; this service asks it over the
    # Redis command bus (src/anthias_server/lib/cec_client.py).
    #
    # That indirection is what makes CEC work on balena at all. The
    # compose file here is baked into the release from a workstation,
    # nothing on-device can enumerate the host's adapters, and a
    # statically listed node that turns out to be absent would stop the
    # container from starting. /dev/vchiq used to be listed for libcec,
    # which is gone.
    restart: always
    volumes:
      - resin-data:/data
    labels:
      io.balena.features.supervisor-api: '1'

  anthias-viewer:
    image: ghcr.io/screenly/anthias-viewer:${GIT_SHORT_HASH}-${BOARD}
    ulimits: *no-core-dumps
    build:
      context: .
      dockerfile: ./docker/Dockerfile.viewer
    depends_on:
      - anthias-server
    # Issue #480 — see docker-compose.balena.yml.tmpl. balena Device or
    # Service Variables (e.g. LANG=nl_NL.UTF-8) are injected by the
    # supervisor at runtime; no compose entry needed for AnthiasWebview
    # to pick them up.
    environment:
      - HOME=/data
      - PORT=8080
      - NOREFRESH=1
      - LISTEN=anthias-server
      # See docker-compose.balena.yml.tmpl: lets start_viewer.sh quit the
      # Plymouth boot splash over the host system bus (mounted by
      # io.balena.features.dbus below) so it releases DRM master before
      # cage starts. See release_boot_splash() in bin/start_viewer.sh.
      - DBUS_SYSTEM_BUS_ADDRESS=unix:path=/host/run/dbus/system_bus_socket
    privileged: true
    restart: always
    shm_size: ${SHM_SIZE}
    volumes:
      - resin-data:/data
    labels:
      io.balena.features.supervisor-api: '1'
      io.balena.features.dbus: '1'

  anthias-celery:
    # Runs on the same image as anthias-server with a CMD override.
    # See docker-compose.yml.tmpl for context on the merge.
    image: ghcr.io/screenly/anthias-server:${GIT_SHORT_HASH}-${BOARD}
    ulimits: *no-core-dumps
    command: >
      nice -n 19 ionice -c 3
      celery -A anthias_server.celery_tasks.celery worker -B -n worker@anthias
      --loglevel=info --schedule /tmp/celerybeat-schedule
    depends_on:
      - anthias-server
      - redis
    environment:
      - HOME=/data
      - CELERY_BROKER_URL=redis://redis:6379/0
      - CELERY_RESULT_BACKEND=redis://redis:6379/0
    # No `devices:` here, and none needed. HDMI-CEC is driven by the
    # anthias-viewer container, which is privileged and therefore
    # already sees every /dev/cec* node; this service asks it over the
    # Redis command bus (src/anthias_server/lib/cec_client.py).
    #
    # That indirection is what makes CEC work on balena at all. The
    # compose file here is baked into the release from a workstation,
    # nothing on-device can enumerate the host's adapters, and a
    # statically listed node that turns out to be absent would stop the
    # container from starting. /dev/vchiq used to be listed for libcec,
    # which is gone.
    restart: always
    volumes:
      - resin-data:/data
    labels:
      io.balena.features.supervisor-api: '1'

  # The WiFi setup portal. Started by the supervisor on every boot and
  # every release apply, which is safe because the DAEMON decides whether
  # a portal is needed -- there is no systemd here to hang an
  # ExecCondition on, and nothing in a compose file can gate a service.
  #
  # wifi-connect 2026.10.11 is the release that makes that true: before
  # it, a device online over ethernet alone was not considered online
  # (NetworkManager reports connectivity as "unknown" wherever the check
  # is unavailable, and only an associated radio counted), so every start
  # raised an access point at a perfectly healthy screen. Verified on a
  # wired Pi 5 and Pi 3 B+ against 2026.10.11: exits 0 in ~20 seconds
  # with "already online; the portal is not needed", no access point, no
  # state file. The same test against 2026.10.10 raises one.
  #
  # Do NOT pin this below 2026.10.11. 2026.10.12 matters too:
  # missing branding is a warning there rather than a refusal to
  # start, which is what makes staging it on the shared volume
  # safe, and a guessed `limited` verdict no longer raises an
  # access point at a wired screen on an isolated LAN.
  #
  # Measured on balenaOS 2.83.18 (NetworkManager 1.28.0) booted in QEMU,
  # because the three facts this service depends on there are not
  # obvious from the Debian path:
  #
  #   * ConnectivityCheckAvailable is FALSE, exactly as on Debian and
  #     Raspberry Pi OS -- so the "unknown" fallback above is the code
  #     path that runs here too. (meta-balena's current bbappend does
  #     add `concheck` to PACKAGECONFIG; this released build predates
  #     that, and either way the daemon handles both.) `Connectivity`
  #     itself still answers 4/FULL, which is the trap: it is optimistic
  #     when no probe ran.
  #
  #   * supervisor0 is a BRIDGE that NetworkManager reports as
  #     `connected (externally)`. Upstream's online check counts only
  #     ethernet, wifi and modem devices, which is what stops every
  #     balena device reading as permanently online and never raising
  #     a portal at all.
  #
  #   * The host D-Bus socket reaches the container through the label
  #     below: the pinned image run with exactly this env and mount
  #     exits 3 ("no WiFi device was found" -- the VM has no radio),
  #     not 5, which is what a failure to reach NetworkManager looks
  #     like.
  anthias-wifi-connect:
    image: ghcr.io/screenly-labs/wifi-connect:2026.10.13
    ulimits: *no-core-dumps
    # Serving DHCP, which does not survive Docker's NAT, and binding the
    # access point's own address on the wireless device.
    network_mode: host
    # `always` is the schedule here, because there is nothing else that
    # can be one.
    #
    # `restart: "no"` does NOT mean "start once per boot", on balena or
    # anywhere else. It means nothing will start the container again
    # once it has exited, and that is by design at both layers:
    #
    #   * The supervisor starts a service only if it has never been
    #     started. Its own words, in compose/app.ts: "Only start a
    #     Service if we have never started it before and the service
    #     matches target! This is so the engine can handle the restart
    #     policy configured for the container." An exited service is
    #     neither Installing nor Installed, so it is left alone. The one
    #     exception is a container the Engine failed to start because a
    #     host interface was not ready, matched by its error message.
    #
    #   * balena-engine, like Docker, only starts containers at daemon
    #     start if their policy is `always` or `unless-stopped`.
    #
    # So on Debian `restart: "no"` is right because systemd starts the
    # container; here there is no systemd, and nothing else volunteers.
    # Confirmed on balenaOS 9.0.3 with supervisor 18.2.2: the
    # container's start and finish times were unchanged across a reboot.
    #
    # Reboot-only would not be enough anyway. The common failure is a
    # screen that stays powered on while the WiFi password changes
    # around it, and telling an operator to power-cycle a wall-mounted
    # display is the advice this feature exists to replace.
    #
    # What makes `always` affordable is that the daemon waits before
    # exiting (below). Without that it is a spin: measured on a wired
    # Pi 5, a container start every eighteen seconds, 4,680 a day, with
    # 19,000 log lines to match -- which on balena ship to the cloud.
    restart: always
    environment:
      # NetworkManager lives on the host; io.balena.features.dbus below
      # is what puts its socket in this container.
      - DBUS_SYSTEM_BUS_ADDRESS=unix:path=/host/run/dbus/system_bus_socket
      # A PREFIX: the portal appends four random characters, so five
      # screens in one room do not all broadcast one name.
      - PORTAL_SSID=Anthias Setup
      - WIFI_CONNECT_BRAND_NAME=Anthias
      # Staged on the shared volume by anthias-server's startup
      # script, because ansible does not run here. Missing files
      # are a warning upstream as of 2026.10.12, not a refusal to
      # start, so losing the race on a first boot costs the portal
      # its colours and not its existence.
      - WIFI_CONNECT_BRAND_CSS=/data/hotspot/brand/brand.css
      - WIFI_CONNECT_BRAND_LOGO=/data/hotspot/brand/logo.svg
      - WIFI_CONNECT_DEFAULT_THEME=dark
      # NOT 80: anthias-server publishes 80:8080 on the host.
      - PORTAL_LISTENING_PORT=8000
      - WIFI_CONNECT_QR_OUTPUT=/data/hotspot
      # How anthias-server learns the SSID, passphrase and portal
      # address to put on the screen, and whether a portal is live.
      - WIFI_CONNECT_STATE_FILE=/data/hotspot/state.json
      - WIFI_CONNECT_NO_QR_STDOUT=true
      # Idle, not a session cap, as of 2026.10.9.
      - ACTIVITY_TIMEOUT=900
      # The cadence `restart: always` turns into a schedule, as of
      # 2026.10.13. Five minutes between checks on a screen that is
      # online -- the check is one D-Bus round trip -- and six hours
      # after a setup page nobody used, because that one just held the
      # display for a quarter of an hour. A site with nothing to join
      # therefore spends a minute in twenty-four showing setup instead
      # of content, which is the same budget the systemd timer and its
      # backoff give the Debian path.
      - WIFI_CONNECT_RECHECK_INTERVAL=300
      - WIFI_CONNECT_RETRY_AFTER=21600
    volumes:
      - resin-data:/data
    labels:
      io.balena.features.dbus: '1'

  redis:
    image: ghcr.io/screenly/anthias-redis:${GIT_SHORT_HASH}-${BOARD}
    ulimits: *no-core-dumps
    build:
      context: .
      dockerfile: ./docker/Dockerfile.redis
    ports:
      - 127.0.0.1:6379:6379
    restart: always
    volumes:
      - redis-data:/var/lib/redis

volumes:
  resin-data:
  redis-data: