DECNET/decnet/collector/worker.py

"""
Host-side Docker log collector.

Streams stdout from all running decky service containers via the Docker SDK,
writes RFC 5424 lines to <log_file> and parsed JSON records to <log_file>.json.
The ingester tails the .json file; rsyslog can consume the .log file independently.
"""

import asyncio
import contextlib
import json
import os
import re
import threading
import time
from concurrent.futures import ThreadPoolExecutor
from datetime import datetime
from pathlib import Path
from typing import Any, Callable, Optional

from decnet.bus import topics as _topics
from decnet.bus.factory import get_bus
from decnet.bus.publish import (
    make_thread_safe_publisher,
    run_control_listener_signal,
    run_health_heartbeat,
)
from decnet.logging import get_logger
from decnet.telemetry import traced as _traced, get_tracer as _get_tracer, inject_context as _inject_ctx

# Collector publish signature: ``publish_fn(parsed_event_dict)``.  Callable
# from the container-stream threads; the worker wraps it around a thread-safe
# bus publisher that marshals onto the asyncio loop.
CollectorPublishFn = Callable[[dict[str, Any]], None]

logger = get_logger("collector")

# ─── Ingestion rate limiter ───────────────────────────────────────────────────
#
# Rationale: connection-lifecycle events (connect/disconnect/accept/close) are
# emitted once per TCP connection. During a portscan or credential-stuffing
# run, a single attacker can generate hundreds of these per second from the
# honeypot services themselves — each becoming a tiny WAL-write transaction
# through the ingester, starving reads until the queue drains.
#
# The collector still writes every line to the raw .log file (forensic record
# for rsyslog/SIEM). Only the .json path — which feeds SQLite — is deduped.
#
# Dedup key: (attacker_ip, decky, service, event_type)
# Window:    DECNET_COLLECTOR_RL_WINDOW_SEC seconds (default 1.0)
# Scope:     DECNET_COLLECTOR_RL_EVENT_TYPES comma list
#            (default: connect,disconnect,connection,accept,close)
# Events outside that set bypass the limiter untouched.

def _parse_float_env(name: str, default: float) -> float:
    raw = os.environ.get(name)
    if raw is None:
        return default
    try:
        value = float(raw)
    except ValueError:
        logger.warning("collector: invalid %s=%r, using default %s", name, raw, default)
        return default
    return max(0.0, value)


_RL_WINDOW_SEC: float = _parse_float_env("DECNET_COLLECTOR_RL_WINDOW_SEC", 1.0)
_RL_EVENT_TYPES: frozenset[str] = frozenset(
    t.strip()
    for t in os.environ.get(
        "DECNET_COLLECTOR_RL_EVENT_TYPES",
        "connect,disconnect,connection,accept,close",
    ).split(",")
    if t.strip()
)
_RL_MAX_ENTRIES: int = 10_000

# APP-NAMEs we never want to see in the ingestion stream — native unix
# daemons that share a container with a DECNET service. Their logs are
# noise: sshd's "Failed password for root from X" duplicates the
# auth-helper's structured `auth_attempt` event, pam_unix repeats it
# again, and CRON/systemd/etc. say nothing about attacker behavior.
# Override or extend with DECNET_COLLECTOR_DROP_APPS (comma list).
_DROP_APPS: frozenset[str] = frozenset(
    a.strip()
    for a in os.environ.get(
        "DECNET_COLLECTOR_DROP_APPS",
        "sshd,pam_unix,sudo,su,CRON,cron,systemd,kernel,rsyslogd,dbus-daemon",
    ).split(",")
    if a.strip()
)

_rl_lock: threading.Lock = threading.Lock()
_rl_last: dict[tuple[str, str, str, str], float] = {}


def _should_ingest(parsed: dict[str, Any]) -> bool:
    """
    Return True if this parsed event should be written to the JSON ingestion
    stream. Drops native unix daemon noise (sshd, pam_unix, …) outright;
    rate-limits connection-lifecycle events within a dedup window.
    """
    if parsed.get("service", "") in _DROP_APPS:
        return False
    event_type = parsed.get("event_type", "")
    if _RL_WINDOW_SEC <= 0.0 or event_type not in _RL_EVENT_TYPES:
        return True
    key = (
        parsed.get("attacker_ip", "Unknown"),
        parsed.get("decky", ""),
        parsed.get("service", ""),
        event_type,
    )
    now = time.monotonic()
    with _rl_lock:
        last = _rl_last.get(key, 0.0)
        if now - last < _RL_WINDOW_SEC:
            return False
        _rl_last[key] = now
        # Opportunistic GC: when the map grows past the cap, drop entries older
        # than 60 windows (well outside any realistic in-flight dedup range).
        if len(_rl_last) > _RL_MAX_ENTRIES:
            cutoff = now - (_RL_WINDOW_SEC * 60.0)
            stale = [k for k, t in _rl_last.items() if t < cutoff]
            for k in stale:
                del _rl_last[k]
    return True


def _reset_rate_limiter() -> None:
    """Test-only helper — clear dedup state between test cases."""
    with _rl_lock:
        _rl_last.clear()

# ─── RFC 5424 parser ──────────────────────────────────────────────────────────

_RFC5424_RE = re.compile(
    r"^<\d+>1 "
    r"(\S+) "       # 1: TIMESTAMP
    r"(\S+) "       # 2: HOSTNAME (decky name)
    r"(\S+) "       # 3: APP-NAME (service)
    r"\S+ "         # PROCID — NILVALUE ("-") for syslog_bridge emitters,
                    # real PID for native syslog callers like sshd/sudo
                    # routed through rsyslog. Accept both; we don't consume it.
    r"(\S+) "       # 4: MSGID (event_type)
    r"(.+)$",       # 5: SD element + optional MSG
)

# Honeypot SSH containers export a ``PROMPT_COMMAND`` that calls
# ``logger --rfc5424 --msgid command -p user.info -t bash "CMD …"``.
# That inner RFC 5424 line lands on the container's stdout, where the
# Docker stream reader prepends ANOTHER RFC 5424 envelope (PRI=14,
# HOSTNAME=<decky>, APP-NAME=1, MSGID=NIL). The outer parse therefore
# sees ``event_type == "-"`` while the real MSGID (``command``) is
# inside the body. We detect that case and re-extract the inner
# ``HOSTNAME APP-NAME PROCID MSGID rest`` so downstream consumers see
# ``event_type == "command"`` plus the real source hostname.
#
# Anchored on an ISO-8601 timestamp at the head of the body so we
# don't false-match free-form prose like "Connection from 1.2.3.4".
_INNER_RFC5424_RE = re.compile(
    r"^(\d{4}-\d{2}-\d{2}T\S+)\s+"  # 1: inner TIMESTAMP
    r"(\S+)\s+"                       # 2: inner HOSTNAME
    r"(\S+)\s+"                       # 3: inner APP-NAME
    r"\S+\s+"                         # PROCID (NIL or PID)
    r"(\S+)\s+"                       # 4: inner MSGID
    r"(.+)$",                         # 5: inner SD/MSG remainder
)
_SD_BLOCK_RE = re.compile(r'\[relay@55555\s+(.*?)\]', re.DOTALL)
_PARAM_RE = re.compile(r'(\w+)="((?:[^"\\]|\\.)*)"')
_IP_FIELDS = ("src_ip", "src", "client_ip", "remote_ip", "remote_addr", "target_ip", "ip")

# Free-form `key=value` pairs in the MSG body. Used for lines that bypass the
# syslog_bridge SD format — e.g. the SSH container's PROMPT_COMMAND which
# calls `logger -t bash "CMD uid=0 user=root src=1.2.3.4 pwd=/root cmd=…"`.
# Values run until the next whitespace, so `cmd=…` at end-of-line is preserved
# as one unit; we only care about IP-shaped fields here anyway.
_MSG_KV_RE = re.compile(r'(\w+)=(\S+)')

# Native sshd / pam syslog lines arrive without an SD block and without
# key=value pairs. The remote address shows up as free prose:
#   "Failed password for root from 1.2.3.4 port 42772 ssh2"
#   "Connection from 1.2.3.4 port 42772 on 10.0.0.2 port 22"
#   "pam_unix(sshd:auth): authentication failure; … rhost=1.2.3.4 user=root"
# Anchored patterns first so we never confuse the attacker with the
# local listener IP ("on 10.0.0.2"). Bare IP scan is the last resort.
_IPV4 = r"\d{1,3}(?:\.\d{1,3}){3}"
_IPV6 = r"[0-9a-fA-F:]+:[0-9a-fA-F:]+"
_IP = rf"(?:{_IPV4}|{_IPV6})"
_MSG_IP_ANCHORED_RE = re.compile(
    rf"\b(?:from|rhost[:=]|client[:=]|src[:=])\s*({_IP})",
    re.IGNORECASE,
)
_MSG_IP_BARE_RE = re.compile(rf"\b({_IPV4})\b")


def parse_rfc5424(line: str) -> Optional[dict[str, Any]]:
    """
    Parse an RFC 5424 DECNET log line into a structured dict.
    Returns None if the line does not match the expected format.
    """
    m = _RFC5424_RE.match(line)
    if not m:
        return None
    ts_raw, decky, service, event_type, sd_rest = m.groups()

    fields: dict[str, str] = {}

    # Honeypot SSH PROMPT_COMMAND lines are double-wrapped (Docker
    # stdout envelope around the inner ``logger --msgid command`` line).
    # Outer MSGID is NIL; the real MSGID is inside the body. Detect
    # the inner shape and re-extract HOSTNAME / APP-NAME / MSGID /
    # remainder so downstream extraction sees the real header.
    if event_type == "-" and sd_rest.startswith("-"):
        body = sd_rest[1:].lstrip()
        inner = _INNER_RFC5424_RE.match(body)
        if inner is not None:
            _i_ts, i_host, i_app, i_msgid, i_rest = inner.groups()
            decky = i_host
            service = i_app
            event_type = i_msgid
            sd_rest = i_rest

    msg: str = ""
    if sd_rest.startswith("-"):
        msg = sd_rest[1:].lstrip()
    elif sd_rest.startswith("["):
        block = _SD_BLOCK_RE.search(sd_rest)
        if block:
            for k, v in _PARAM_RE.findall(block.group(1)):
                fields[k] = v.replace('\\"', '"').replace("\\\\", "\\").replace("\\]", "]")
        # Always recover the post-SD message tail, even when the SD
        # block isn't ``relay@55555`` (e.g. the ``timeQuality`` block
        # syslog auto-emits on bash CMD lines). Without this the body
        # of unwrapped PROMPT_COMMAND lines stays empty and the
        # attacker_ip kv-fallback below has nothing to scan.
        msg_match = re.search(r'\]\s+(.+)$', sd_rest)
        if msg_match:
            msg = msg_match.group(1).strip()
    else:
        msg = sd_rest

    attacker_ip = "Unknown"
    for fname in _IP_FIELDS:
        if fname in fields:
            attacker_ip = fields[fname]
            break

    # Fallback for plain `logger` callers that don't use SD params (notably
    # the SSH container's bash PROMPT_COMMAND: `logger -t bash "CMD … src=IP …"`).
    # Scan the MSG body for IP-shaped `key=value` tokens ONLY — don't fold
    # them into `fields`, because the frontend's parseEventBody already
    # renders kv pairs from the msg and doubling them up produces noisy
    # duplicate pills. This keeps attacker attribution working without
    # changing the shape of `fields` for non-SD lines.
    if attacker_ip == "Unknown" and msg:
        for k, v in _MSG_KV_RE.findall(msg):
            if k in _IP_FIELDS:
                attacker_ip = v
                break

    # Final fallback for native syslog producers that emit free-form prose
    # (notably sshd and pam_unix routed via rsyslog without the relay@55555
    # SD wrapper). Prefer anchored matches so the local listener address in
    # "Connection from X port Y on Z port 22" never wins over X.
    if attacker_ip == "Unknown" and msg:
        anchored = _MSG_IP_ANCHORED_RE.search(msg)
        if anchored:
            attacker_ip = anchored.group(1)
        else:
            bare = _MSG_IP_BARE_RE.search(msg)
            if bare:
                attacker_ip = bare.group(1)

    try:
        ts_formatted = datetime.fromisoformat(ts_raw).strftime("%Y-%m-%d %H:%M:%S")
    except ValueError:
        ts_formatted = ts_raw

    # Free-form bash PROMPT_COMMAND lines (MSGID=NIL, body starts with
    # "CMD ") get event_type rewritten to "command". `fields` stays empty
    # so the frontend's msg-based pill rendering doesn't double up.
    if event_type == "-" and msg.startswith("CMD "):
        event_type = "command"

    return {
        "timestamp": ts_formatted,
        "decky": decky,
        "service": service,
        "event_type": event_type,
        "attacker_ip": attacker_ip,
        "fields": fields,
        "msg": msg,
        "raw_line": line,
    }


# ─── Container helpers ────────────────────────────────────────────────────────

def _load_service_container_names() -> set[str]:
    """
    Return the exact set of service container names from decnet-state.json.
    Format: {decky_name}-{service_name}, e.g. 'omega-decky-smtp'.
    Returns an empty set if no state file exists.
    """
    from decnet.config import load_state
    state = load_state()
    if state is None:
        return set()
    config, _ = state
    names: set[str] = set()
    for decky in config.deckies:
        for svc in decky.services:
            names.add(f"{decky.name}-{svc.replace('_', '-')}")
    return names


_TOPOLOGY_SERVICE_LABEL = "decnet.topology.service"
_FLEET_SERVICE_LABEL = "decnet.fleet.service"


def _has_decnet_service_label(labels: Optional[dict]) -> bool:
    """Recognize both fleet (``decnet.fleet.service``, set by
    ``decnet/composer.py``) and MazeNET topology (``decnet.topology.service``,
    set by ``decnet/topology/compose.py``) containers.

    Label-based detection is the canonical path: it's stateless and avoids
    the race between ``docker compose up`` and the ``decnet-state.json``
    write that previously caused freshly-deployed fleet containers to be
    silently dropped by the docker-events watcher.
    """
    if not labels:
        return False
    return (
        labels.get(_TOPOLOGY_SERVICE_LABEL) == "true"
        or labels.get(_FLEET_SERVICE_LABEL) == "true"
    )


def is_service_container(container) -> bool:
    """Return True if this Docker container is a known DECNET service container.

    Label-based detection is preferred (works for both fleet and MazeNET
    topology containers without touching decnet-state.json). The
    state-file name match remains as a fallback so containers built from
    older composes — which predate the ``decnet.fleet.service`` label —
    are still picked up.
    """
    if isinstance(container, str):
        return container.lstrip("/") in _load_service_container_names()
    labels: Optional[dict] = None
    attrs = getattr(container, "attrs", None)
    if isinstance(attrs, dict):
        labels = (attrs.get("Config") or {}).get("Labels")
    if labels is None:
        labels = getattr(container, "labels", None)
    if _has_decnet_service_label(labels):
        return True
    # Fallback: legacy containers without labels still match by name.
    name = container.name.lstrip("/")
    return name in _load_service_container_names()


def is_service_event(attrs: dict) -> bool:
    """Return True if a Docker start event is for a known DECNET service container.

    Docker start-event attrs flatten every container label alongside the
    ``name``/``image`` keys — no separate ``labels`` sub-dict — so label
    detection happens directly on ``attrs``.

    Prefer the label path because it's race-free with respect to the
    ``decnet-state.json`` write that ``decnet deploy`` performs around
    ``docker compose up``: a freshly-started container's start event can
    arrive before the state file has been updated, and the legacy
    name-based fallback would then drop the event.
    """
    if _has_decnet_service_label(attrs):
        return True
    name = attrs.get("name", "").lstrip("/")
    return name in _load_service_container_names()


# ─── Blocking stream worker (runs in a thread) ────────────────────────────────

def _reopen_if_needed(path: Path, fh: Optional[Any]) -> Any:
    """Return fh if it still points to the same inode as path; otherwise close
    fh and open a fresh handle.  Handles the file being deleted (manual rm) or
    rotated (logrotate rename + create)."""
    try:
        if fh is not None and os.fstat(fh.fileno()).st_ino == os.stat(path).st_ino:
            return fh
    except OSError:
        pass
    # File gone or inode changed — close stale handle and open a new one.
    if fh is not None:
        try:
            fh.close()
        except Exception:  # nosec B110 — best-effort file handle cleanup
            pass
    path.parent.mkdir(parents=True, exist_ok=True)
    return open(path, "a", encoding="utf-8")


@_traced("collector.stream_container")
def _stream_container(
    container_id: str,
    log_path: Path,
    json_path: Path,
    publish_fn: CollectorPublishFn | None = None,
) -> None:
    """Stream logs from one container and append to the host log files."""
    import docker

    lf: Optional[Any] = None
    jf: Optional[Any] = None
    try:
        client = docker.from_env()
        container = client.containers.get(container_id)
        log_stream = container.logs(stream=True, follow=True, stdout=True, stderr=False)
        buf = ""
        for chunk in log_stream:
            buf += chunk.decode("utf-8", errors="replace")
            while "\n" in buf:
                line, buf = buf.split("\n", 1)
                line = line.rstrip()
                if not line:
                    continue
                lf = _reopen_if_needed(log_path, lf)
                lf.write(line + "\n")
                lf.flush()
                parsed = parse_rfc5424(line)
                if parsed:
                    if _should_ingest(parsed):
                        _tracer = _get_tracer("collector")
                        with _tracer.start_as_current_span("collector.event") as _span:
                            _span.set_attribute("decky", parsed.get("decky", ""))
                            _span.set_attribute("service", parsed.get("service", ""))
                            _span.set_attribute("event_type", parsed.get("event_type", ""))
                            _span.set_attribute("attacker_ip", parsed.get("attacker_ip", ""))
                            _inject_ctx(parsed)
                            logger.debug("collector: event written decky=%s type=%s", parsed.get("decky"), parsed.get("event_type"))
                            jf = _reopen_if_needed(json_path, jf)
                            jf.write(json.dumps(parsed) + "\n")
                            jf.flush()
                            if publish_fn is not None:
                                try:
                                    publish_fn(parsed)
                                except Exception as exc:
                                    logger.debug(
                                        "collector: bus publish failed: %s", exc,
                                    )
                    else:
                        logger.debug(
                            "collector: rate-limited decky=%s service=%s type=%s attacker=%s",
                            parsed.get("decky"), parsed.get("service"),
                            parsed.get("event_type"), parsed.get("attacker_ip"),
                        )
                else:
                    logger.debug("collector: malformed RFC5424 line snippet=%r", line[:80])
    except Exception as exc:
        logger.debug("collector: log stream ended container_id=%s reason=%s", container_id, exc)
    finally:
        for fh in (lf, jf):
            if fh is not None:
                try:
                    fh.close()
                except Exception:  # nosec B110 — best-effort file handle cleanup
                    pass


# ─── Bus plumbing ─────────────────────────────────────────────────────────────

def _make_system_log_publisher(
    bus: Any, loop: asyncio.AbstractEventLoop,
) -> CollectorPublishFn:
    """Factory: returns a ``publish_fn(parsed)`` for use by stream threads.

    When *bus* is ``None`` the returned callable is a no-op, so the stream
    thread can call it unconditionally.  Otherwise each call is marshalled
    onto *loop* (the asyncio event loop that owns the bus socket) via
    ``make_thread_safe_publisher``.
    """
    raw_publish = make_thread_safe_publisher(bus, loop) if bus is not None else None
    if raw_publish is None:
        return lambda _parsed: None

    topic = _topics.system(_topics.SYSTEM_LOG)

    def _publish(parsed: dict[str, Any]) -> None:
        event_type = parsed.get("event_type", "")
        raw_publish(
            topic,
            {
                "decky": parsed.get("decky", ""),
                "service": parsed.get("service", ""),
                "event_type": event_type,
                "attacker_ip": parsed.get("attacker_ip", "Unknown"),
                "timestamp": parsed.get("timestamp", ""),
            },
            event_type,
        )

    return _publish


# ─── Async collector ──────────────────────────────────────────────────────────

async def log_collector_worker(log_file: str) -> None:
    """
    Background task: streams Docker logs from all running decky service
    containers, writing RFC 5424 lines to log_file and parsed JSON records
    to log_file.json for the ingester to consume.

    Watches Docker events to pick up containers started after initial scan.
    """
    import docker

    log_path = Path(log_file)
    json_path = log_path.with_suffix(".json")
    log_path.parent.mkdir(parents=True, exist_ok=True)

    active: dict[str, asyncio.Task[None]] = {}
    loop = asyncio.get_running_loop()

    # Optional bus wiring — per-line system.log publish.  Fan-in from many
    # container-stream threads is handled by make_thread_safe_publisher,
    # which marshals each publish onto this loop.
    bus = None
    try:
        bus = get_bus(client_name="collector")
        await bus.connect()
    except Exception as exc:
        logger.warning("collector: bus unavailable, continuing without publish: %s", exc)
        bus = None

    _publish_log = _make_system_log_publisher(bus, loop)

    # Workers panel health heartbeat + bus-driven stop control.  The
    # heartbeat beacons on system.collector.health every 30s; the
    # control listener translates a bus stop intent into a SIGTERM to
    # this process (collector's main loop is a blocking thread pool, so
    # self-signalling is cleaner than threading a shutdown event).
    heartbeat_task = asyncio.create_task(run_health_heartbeat(bus, "collector"))
    control_task = asyncio.create_task(run_control_listener_signal(bus, "collector"))

    # Periodic re-scan of running containers. Belt to the event-watcher's
    # suspenders: if dockerd or the SDK ever drops a start event during a
    # reconnect window (the retry loop in ``_watch_events`` covers the
    # restart itself, but events fired *during* the gap are lost), this
    # loop picks up the orphan within ``RECONCILE_INTERVAL_S``. Also
    # prunes finished futures so ``active`` doesn't accumulate over the
    # agent's lifetime as topology mutations churn containers.
    _reconcile_interval_s = float(
        os.environ.get("DECNET_COLLECTOR_RECONCILE_S", "30")
    )

    # Dedicated thread pool so long-running container log streams don't
    # saturate the default asyncio executor and starve short-lived
    # to_thread() calls elsewhere (e.g. load_state in the web API).
    collector_pool = ThreadPoolExecutor(
        max_workers=64, thread_name_prefix="decnet-collector",
    )

    def _spawn(container_id: str, container_name: str) -> None:
        if container_id not in active or active[container_id].done():
            active[container_id] = asyncio.ensure_future(
                loop.run_in_executor(
                    collector_pool, _stream_container,
                    container_id, log_path, json_path, _publish_log,
                ),
                loop=loop,
            )
            logger.info("collector: streaming container=%s", container_name)

    try:
        logger.info("collector started log_path=%s", log_path)
        client = docker.from_env()

        async def _reconcile_loop() -> None:
            while True:
                try:
                    await asyncio.sleep(_reconcile_interval_s)
                    # Drop done futures so the dict's bounded by the
                    # current container count, not lifetime churn.
                    for cid in [c for c, t in active.items() if t.done()]:
                        active.pop(cid, None)
                    containers = await loop.run_in_executor(
                        collector_pool,
                        lambda: list(client.containers.list()),
                    )
                    for container in containers:
                        if container.id in active:
                            continue
                        if is_service_container(container):
                            _spawn(container.id, container.name.lstrip("/"))
                except asyncio.CancelledError:
                    raise
                except Exception as exc:  # noqa: BLE001 — keep loop alive across SDK transients
                    logger.warning("collector: reconcile pass failed: %s", exc)

        reconcile_task = asyncio.create_task(_reconcile_loop())

        for container in client.containers.list():
            if is_service_container(container):
                _spawn(container.id, container.name.lstrip("/"))

        def _watch_events() -> None:
            # The dockerd event stream is the fast path for picking up
            # newly-started service containers. It can break in two ways:
            # (a) dockerd restart / reload severs the long-poll, (b) the
            # SDK's JSON-stream decoder occasionally raises on a partial
            # frame. Both used to make this thread return cleanly, leaving
            # the collector "running" with no event subscription — future
            # container starts were silently dropped until an operator
            # restarted the unit. Retry with exponential backoff (cap at
            # 30s, matching the heartbeat cadence) so dockerd hiccups are
            # invisible to the operator. The reconcile loop is the safety
            # net for any events lost during the reconnect window.
            backoff = 1.0
            while True:
                try:
                    for event in client.events(
                        decode=True,
                        filters={"type": "container", "event": "start"},
                    ):
                        attrs = event.get("Actor", {}).get("Attributes", {})
                        cid  = event.get("id", "")
                        name = attrs.get("name", "")
                        if cid and is_service_event(attrs):
                            loop.call_soon_threadsafe(_spawn, cid, name)
                    # Clean iterator exhaustion: real dockerd doesn't
                    # close the stream voluntarily, so this only
                    # happens in tests with mocked iterators or in
                    # genuinely unrecoverable daemon states. Either
                    # way, returning lets the worker shut down
                    # cleanly — the reconciler is the safety net for
                    # productive cases.
                    return
                except Exception as exc:  # noqa: BLE001 — SDK leaks bare Exceptions on stream-decode errors
                    logger.warning(
                        "collector: event stream broke (%s: %s); reconnecting in %.1fs",
                        type(exc).__name__, exc, backoff,
                    )
                    time.sleep(backoff)
                    backoff = min(backoff * 2, 30.0)

        await loop.run_in_executor(collector_pool, _watch_events)

    except asyncio.CancelledError:
        logger.info("collector shutdown requested cancelling %d tasks", len(active))
        for task in active.values():
            task.cancel()
        collector_pool.shutdown(wait=False)
        raise
    except Exception as exc:
        logger.error("collector error: %s", exc)
    finally:
        collector_pool.shutdown(wait=False)
        # `reconcile_task` may not exist if startup failed before
        # `client = docker.from_env()` returned; tolerate that.
        _maintenance_tasks = [heartbeat_task, control_task]
        if "reconcile_task" in locals():
            _maintenance_tasks.append(reconcile_task)
        for t in _maintenance_tasks:
            t.cancel()
            with contextlib.suppress(Exception, asyncio.CancelledError):
                await t
        if bus is not None:
            with contextlib.suppress(Exception):
                await bus.close()