move_detect/edge/daemon.py

import os


# Rate-limit for the swallowed-exception log: a camera failing every frame must be VISIBLE
# (not a silent green unit) but must not spam the journal. Log the 1st of a run of consecutive
# failures, then every LOG_EVERY-th, and again after any recovery.
_LOG_EVERY = 300


def _default_log(msg, exc=None):  # pragma: no cover (thin stderr wrapper)
    import sys
    import traceback
    print(msg, file=sys.stderr)
    if exc is not None:
        traceback.print_exception(type(exc), exc, exc.__traceback__, file=sys.stderr)


def run_loop(runtime, client, queue, *, clock, sleep, iterations=None,
             sync_every=150, heartbeat_every=900, flush_every=300,
             frame_interval=0.0, zones_for=None, log=None,
             heartbeat_interval_s=None, health_fn=None, configured_cameras=None,
             notify=None):
    # `runtime` may be a single PipelineRuntime or a list of them (multi-camera). All
    # cameras share this loop, one shared clock read per tick, and the periodic
    # sync/heartbeat/flush. A per-camera step exception is counted and rate-limit-logged,
    # never fatal — but it is no longer SILENT (a permanently broken pipeline used to look
    # perfectly green).
    runtimes = list(runtime) if isinstance(runtime, (list, tuple)) else [runtime]
    if zones_for is None:
        from edge.zones import zones_from_config
        zones_for = lambda cfg, rt: zones_from_config(cfg)
    if log is None:
        log = _default_log
    if health_fn is None:
        from edge.health import read_health
        health_fn = read_health
    if notify is None:
        from edge.watchdog import ping as notify
    # The ping happens at the TOP of each tick, before any camera work. That is the point: it
    # proves the loop is turning. Pinging after the work would still ping while a camera step
    # blocks forever, which is the exact failure the watchdog exists to catch.
    counters = {"frames": 0, "syncs": 0, "beats": 0, "flushes": 0, "errors": 0}
    streak = {}  # id(rt)/"periodic" -> consecutive-error count, for rate-limited logging

    def note(key, label, exc=None):
        counters["errors"] += 1
        n = streak.get(key, 0) + 1
        streak[key] = n
        if n == 1 or n % _LOG_EVERY == 0:
            if exc is not None:
                log(f"{label}: {type(exc).__name__}: {exc} (consecutive #{n})", exc)
            else:
                log(f"{label} (consecutive #{n})", None)

    beat_anchor = None  # (wall_time, frames) at the previous heartbeat, for measured fps
    last_beat = None    # wall time of the last heartbeat, for the time-based cadence
    reported_version = None  # last config version the server told us; echoed in heartbeats
    i = 0
    while iterations is None or i < iterations:
        i += 1
        try:
            notify()
        except Exception as e:
            note("watchdog", "watchdog ping failed", e)
        now = clock()  # one timestamp per tick, shared across cameras
        if beat_anchor is None:
            beat_anchor = (now, 0)
        if last_beat is None:
            last_beat = now
        for rt in runtimes:
            try:
                produced = rt.step(now)
            except Exception as e:
                cam = getattr(rt, "camera_id", "?")
                note(id(rt), f"camera {cam} step failed", e)
                continue
            if produced is None:
                # step() returned no frame (unplugged/browned-out camera). Count it toward the
                # erroring streak instead of as a healthy frame — otherwise a dead camera ticks
                # frames++ forever and reports fps~=target with cameras_erroring=0 (a silent
                # monitoring outage that looks green). A single dropped frame self-clears on the
                # next good frame, so only SUSTAINED starvation shows as erroring.
                cam = getattr(rt, "camera_id", "?")
                note(id(rt), f"camera {cam} produced no frame")
            else:
                counters["frames"] += 1
                streak.pop(id(rt), None)  # recovered — reset the streak
        try:
            if sync_every and i % sync_every == 0:
                cfg = client.fetch_config()
                if cfg is not None:
                    for rt in runtimes:
                        rt.apply_zones(zones_for(cfg, rt))
                    # Remember the config version so heartbeats can echo it back — the server's
                    # config-drift indicator is meaningless until the device reports what it runs.
                    if isinstance(cfg, dict) and cfg.get("version") is not None:
                        reported_version = int(cfg["version"])
                    counters["syncs"] += 1
            # Heartbeat cadence: time-based when heartbeat_interval_s is set (production), so the
            # interval is predictable regardless of per-tick cost (CPU MoveNet, multi-cam, thermal
            # throttle can push a tick past 100ms) — an iteration count cannot guarantee a wall-clock
            # cadence, and the server marks a device offline on ELAPSED time. Falls back to the
            # iteration count when unset (used by the deterministic unit tests).
            if heartbeat_interval_s:
                do_beat = (now - last_beat) >= heartbeat_interval_s
            else:
                do_beat = bool(heartbeat_every) and (i % heartbeat_every == 0)
            if do_beat:
                # Measured fps over the window since the last beat, and per-camera health, so
                # the server can tell a live device from a degraded/dead one (was hardcoded 0.0
                # + a frames counter that ticked even when no frame was read).
                t0, f0 = beat_anchor
                elapsed = now - t0
                fps = (counters["frames"] - f0) / elapsed if elapsed > 0 else 0.0
                # cameras_total is the CONFIGURED count, not just the started one: a camera that
                # failed to construct at boot (wrong /dev index, unseated ribbon) is skipped by
                # build_runtimes and would otherwise vanish from the heartbeat entirely — the
                # server could never see that a configured camera is missing. Count the not-started
                # cameras as erroring so a 4-cam device that lost one reports 4 total / 1 erroring.
                total = configured_cameras if configured_cameras is not None else len(runtimes)
                not_started = max(0, total - len(runtimes))
                erroring = not_started + sum(1 for rt in runtimes if id(rt) in streak)
                metrics = {"fps": round(fps, 2), "frames": counters["frames"],
                           "cameras_total": total, "cameras_erroring": erroring,
                           "errors": counters["errors"], "queue_pending": queue.pending()}
                # Pi thermal telemetry (None off-Pi or on failure → field omitted, never a fake 0).
                health = health_fn()
                if health.get("temp") is not None:
                    metrics["temp"] = health["temp"]
                if health.get("throttle") is not None:
                    metrics["throttle"] = health["throttle"]
                if reported_version is not None:
                    metrics["config_version"] = reported_version
                if client.send_heartbeat(metrics):
                    counters["beats"] += 1
                beat_anchor = (now, counters["frames"])
                last_beat = now
            if flush_every and i % flush_every == 0:
                queue.flush()
                counters["flushes"] += 1
            streak.pop("periodic", None)
        except Exception as e:
            note("periodic", "periodic op (sync/heartbeat/flush) failed", e)
        if frame_interval:
            sleep(frame_interval)
    return counters


def parse_cameras(env, default_camera_id):
    """CAMERAS='backend:device:camera_id,...' -> [(backend, device, camera_id), ...].
    When CAMERAS is unset, fall back to a single camera from
    CAMERA_BACKEND + {UVC,CSI}_DEVICE (using default_camera_id)."""
    raw = (env.get("CAMERAS") or "").strip()
    if raw:
        cams = []
        for spec in raw.split(","):
            spec = spec.strip()
            if not spec:
                continue
            parts = [p.strip() for p in spec.split(":")]
            if len(parts) != 3:
                raise ValueError(f"bad CAMERAS entry {spec!r} (want backend:device:camera_id)")
            cams.append((parts[0].lower(), parts[1], parts[2]))
        if not cams:
            raise ValueError("CAMERAS is set but has no valid entries")  # never monitor zero cameras
        return cams
    backend = (env.get("CAMERA_BACKEND") or "uvc").lower()
    device = env.get("CSI_DEVICE" if backend == "csi" else "UVC_DEVICE") or "0"
    return [(backend, device, default_camera_id)]


def build_runtimes(cameras, make_source, make_runtime, log=None):
    """Build one (source -> runtime) per camera. A camera that fails to CONSTRUCT
    (e.g. an unseated/absent CSI camera whose Picamera2(...) raises) is logged and
    SKIPPED so it never crash-loops the whole daemon and the other cameras still run.
    Raises if no camera comes up (never run monitoring nothing)."""
    if log is None:
        import sys
        log = lambda m: print(m, file=sys.stderr)
    runtimes = []
    for backend, device, cam in cameras:
        try:
            runtimes.append(make_runtime(cam, make_source(backend, device)))
        except Exception as e:
            log(f"camera {cam} ({backend}:{device}) failed to start, skipping: {e}")
    if not runtimes:
        raise RuntimeError("no cameras could be started")
    return runtimes


def main() -> None:  # pragma: no cover (integration entrypoint, exercised on device)
    import time
    from dataclasses import replace
    from edge.config import (frame_interval_from_env, load_config, queue_db_path,
                             require_upload_config)
    from edge.serverclient import ServerClient
    from edge.alert import PostClient
    from edge.queue import OfflineQueue
    from edge.runtime import PipelineRuntime
    from edge.zones import zones_from_config

    def build_source(backend, device):
        if backend == "csi":
            from edge.adapters.picamera_source import PiCameraSource
            return PiCameraSource(int(device))
        from edge.adapters.uvc_source import UvcSource
        return UvcSource(int(device))

    def build_pose():
        # POSE_BACKEND: "hailo" (Pi5 + Hailo-8, needs HAILO_HEF) or CPU "movenet"
        # (needs MOVENET_ONNX). One shared adapter serves every camera (the Hailo
        # device / onnxruntime session is reused per frame, round-robin across cams).
        if os.environ.get("POSE_BACKEND", "movenet").lower() == "hailo":
            from edge.adapters.hailo_pose import HailoPoseAdapter
            return HailoPoseAdapter(os.environ["HAILO_HEF"])
        from edge.adapters.movenet_pose import MoveNetPoseAdapter
        return MoveNetPoseAdapter(os.environ["MOVENET_ONNX"])

    cfg = load_config(dict(os.environ))
    require_upload_config(cfg)  # fail loud now, not silently upload nothing forever
    nonce = lambda: os.urandom(8).hex()
    # Two clocks, deliberately: the loop/detector timers run on a MONOTONIC clock so an NTP step
    # (a Pi 5 has no RTC battery; it boots on stale fake-hwclock time until chrony corrects it)
    # cannot instantly satisfy immobile_s, expire every track, or reset an immobility countdown.
    # Wall time is used only where a real timestamp is required: the HMAC X-Timestamp headers,
    # which the server validates against its own wall clock, and Event.occurred_at.
    mono = time.monotonic
    wall = time.time
    base = cfg.cafe24_url.rstrip("/")
    client = ServerClient(base, cfg.device_id, cfg.secret, nonce, wall)
    post = PostClient(base + "/api/events", cfg.device_id, cfg.secret, nonce, wall)
    queue = OfflineQueue(queue_db_path(dict(os.environ)), post)

    pose = build_pose()  # shared across cameras

    def make_runtime(cam, source):
        # Per-camera config so events are stamped with that camera's id.
        cam_cfg = replace(cfg, camera_id=cam, engine=replace(cfg.engine, camera_id=cam))
        return PipelineRuntime(cam_cfg, source, pose, queue, wall_clock=wall)

    # A missing/unseated camera is skipped (logged), not fatal — the others still run.
    cameras = parse_cameras(dict(os.environ), cfg.camera_id)
    runtimes = build_runtimes(cameras, build_source, make_runtime)

    run_loop(runtimes, client, queue, clock=mono, sleep=time.sleep, iterations=None,
             frame_interval=frame_interval_from_env(dict(os.environ)),
             heartbeat_interval_s=30,  # predictable cadence; server marks offline after 120s (4x)
             configured_cameras=len(cameras),  # so a camera that failed to start stays visible
             zones_for=lambda c, rt: zones_from_config(c, rt.camera_id))