Expire presence so users can't get stuck online (#77)

Presence was a per-user refcount in Redis, decremented only when a
connection's cleanup ran. Anything that ended a connection without
cleanup (a crash, an unclean restart, a laptop asleep with the page
frozen) leaked its count forever, so those users never showed offline.
Production had 19 counts for 2 live connections.

Presence is now one expiring entry per connection (90s TTL) for global,
per-room and focus presence, in new presence:v2:* keys. Clients renew
theirs by sending any frame, which the existing 30s ping/pong already
does. The server closes a human connection that sends nothing for 90s
(code 4408), which catches a frozen page whose browser still answers
transport pings. Bot connections can't be expected to answer pings, so
the server renews theirs itself while the handler is alive.

An expiry has no disconnect event to announce it, so a sweeper on each
worker (one per 15s overall, via a Redis lock) prunes expired entries
and broadcasts the offline transition. The old presence:* hashes are
deleted at startup.

Tests cover the expiry store, the sweeper, and the websocket cases.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

backend/app/main.py

@@ -1,5 +1,6 @@
 import asyncio
 import contextlib
+import logging
 import pathlib
 from collections.abc import AsyncIterator
 from contextlib import asynccontextmanager
@@ -11,6 +12,7 @@ from redis.asyncio import Redis
 from starlette.middleware.sessions import SessionMiddleware
 
 from app.config import settings
+from app.database import async_session_factory
 from app.routers import (
     admin,
     auth,
@@ -31,6 +33,9 @@ from app.ws.connection_manager import ConnectionManager
 from app.ws.focus_presence import FocusPresence
 from app.ws.global_presence import GlobalPresence
 from app.ws.presence import Presence
+from app.ws.presence_sweeper import purge_legacy_presence, run_presence_sweeper
+
+logger = logging.getLogger(__name__)
 
 # backend/app/main.py -> backend/ -> repo root -- matches both the local
 # monorepo layout and the production layout (/srv/ds-chat/backend,
@@ -65,11 +70,23 @@ async def lifespan(app: FastAPI) -> AsyncIterator[None]:
     # restarts on failure instead.
     listener_task = asyncio.create_task(app.state.broadcaster.run_forever())
 
+    # #77: presence entries now expire, and an expiry has no disconnect event
+    # to announce it -- this loop is what notices. The purge clears the old
+    # refcount hashes (and the counts that had leaked into them); it's
+    # best-effort, since Redis being down at startup shouldn't stop the app.
+    try:
+        await purge_legacy_presence(app.state.redis)
+    except Exception:
+        logger.warning("Could not purge legacy presence keys", exc_info=True)
+    sweeper_task = asyncio.create_task(run_presence_sweeper(app))
+
     yield
 
-    listener_task.cancel()
-    with contextlib.suppress(asyncio.CancelledError):
-        await listener_task
+    for task in (sweeper_task, listener_task):
+        task.cancel()
+    for task in (sweeper_task, listener_task):
+        with contextlib.suppress(asyncio.CancelledError):
+            await task
     await app.state.redis.aclose()
 
 
@@ -85,6 +102,7 @@ def create_app() -> FastAPI:
     )
 
     app.state.connection_manager = ConnectionManager()
+    app.state.session_factory = async_session_factory
     app.state.redis = Redis.from_url(settings.redis_url, decode_responses=True)
     app.state.presence = Presence(app.state.redis)
     app.state.global_presence = GlobalPresence(app.state.redis)

backend/app/ws/chat.py

@@ -1,4 +1,7 @@
 import asyncio
+import contextlib
+import logging
+import time
 import uuid
 
 from fastapi import APIRouter, Depends, WebSocket, WebSocketDisconnect
@@ -28,6 +31,8 @@ from app.services.message_service import (
 from app.services.room_service import mark_room_read
 from app.services.session_service import resolve_session
 
+logger = logging.getLogger(__name__)
+
 router = APIRouter(tags=["ws"])
 
 WS_UNAUTHENTICATED = 4401
@@ -44,6 +49,22 @@ WS_UNAUTHENTICATED = 4401
 # itself actually recovering).
 WS_PING_INTERVAL_SECONDS = 30
 
+# #77: a human client that has sent nothing -- not even the "pong" to the ping
+# above -- for this long is treated as gone and disconnected, so its presence
+# stops being renewed and cleanup runs. Needed on top of the transport's own
+# keepalive: a browser answers *protocol-level* pings from its network layer
+# even while the page's JavaScript is frozen (a suspended or discarded tab,
+# a sleeping laptop that keeps its network up), so the connection looks
+# healthy forever while nobody is actually there. Only the page itself can
+# send the app-level "pong". The frontend reconnects on any non-4401 close.
+WS_DEAD_AFTER_SECONDS = 90
+WS_HEARTBEAT_TIMEOUT = 4408
+
+# Presence entries (see ExpiringPresenceSet) are renewed on incoming frames,
+# at most this often -- a chatty client shouldn't turn every message into
+# several extra Redis writes.
+PRESENCE_REFRESH_MIN_INTERVAL_SECONDS = 10
+
 
 class ClientEnvelope(BaseModel):
     type: str
@@ -74,6 +95,21 @@ def _missing_scope(api_token: ApiToken | None, scope: str) -> bool:
     return api_token is not None and scope not in api_token.scopes
 
 
+async def _broadcast_offline(db: AsyncSession, broadcaster, user_id: uuid.UUID) -> None:
+    await broadcast_member_updated(db, broadcaster, user_id)
+    await broadcast_dm_presence_update(db, broadcaster, user_id, online=False)
+
+
+async def _best_effort(label: str, awaitable):
+    # Disconnect cleanup is several independent Redis/DB steps; one failing
+    # must not skip the ones after it.
+    try:
+        return await awaitable
+    except Exception:
+        logger.warning("Presence cleanup step failed: %s", label, exc_info=True)
+        return None
+
+
 @router.websocket("/ws/chat")
 async def chat_endpoint(websocket: WebSocket, db: AsyncSession = Depends(get_db)) -> None:
     api_token: ApiToken | None = None
@@ -114,13 +150,19 @@ async def chat_endpoint(websocket: WebSocket, db: AsyncSession = Depends(get_db)
     # refcount if the socket closes while still blurred -- mirroring how
     # joined_rooms tracks per-connection room membership for its own cleanup.
     is_blurred = False
+    # Identifies this connection's own presence entries (#77) -- see
+    # ExpiringPresenceSet -- so one connection ending can't affect another
+    # belonging to the same user, and a dead one can simply stop renewing.
+    conn_id = uuid.uuid4().hex
+    last_frame_at = time.monotonic()
+    last_presence_refresh = last_frame_at
     manager.register_user(user.id, websocket)
     # Only broadcast on a genuine offline->online transition (this user's
     # first open connection), not for every extra tab -- broadcast_member_
     # updated tells every room this user's in to refresh, which would be
     # wasted churn on a transition that didn't actually change anything
     # visible.
-    if await global_presence.connect(user.id):
+    if await global_presence.connect(user.id, conn_id):
         await broadcast_member_updated(db, broadcaster, user.id)
         await broadcast_dm_presence_update(db, broadcaster, user.id, online=True)
     # This session is shared for the connection's entire lifetime (which can
@@ -135,6 +177,18 @@ async def chat_endpoint(websocket: WebSocket, db: AsyncSession = Depends(get_db)
     # waiting for the next one.
     await db.commit()
 
+    async def refresh_presence() -> None:
+        # Renews every presence entry this connection holds. If the global
+        # one had already lapsed (a long stall) this is also the moment the
+        # user comes back online, and has to be announced like a fresh connect.
+        if await global_presence.refresh(user.id, conn_id):
+            await broadcast_member_updated(db, broadcaster, user.id)
+            await broadcast_dm_presence_update(db, broadcaster, user.id, online=True)
+        for room_id in joined_rooms:
+            await presence.refresh(room_id, user.id, conn_id)
+        if is_blurred:
+            await focus_presence.refresh(user.id, conn_id)
+
     try:
         while True:
             try:
@@ -147,9 +201,30 @@ async def chat_endpoint(websocket: WebSocket, db: AsyncSession = Depends(get_db)
                 # send (raising, caught below same as any other disconnect);
                 # a merely quiet one just gets proded and the loop continues
                 # waiting.
+                if api_token is None and time.monotonic() - last_frame_at >= WS_DEAD_AFTER_SECONDS:
+                    # Gone (see WS_DEAD_AFTER_SECONDS). The socket may be too
+                    # dead to even take a close frame; either way, leaving the
+                    # loop runs the normal cleanup below.
+                    with contextlib.suppress(Exception):
+                        await websocket.close(code=WS_HEARTBEAT_TIMEOUT)
+                    break
+                if api_token is not None:
+                    # A bot isn't expected to speak our app-level heartbeat
+                    # (it may just ignore the ping), so it can't be dropped
+                    # for not answering and its presence can't depend on
+                    # answers. While the handler is alive the transport's own
+                    # keepalive is what vouches for it -- a dead bot socket
+                    # ends this handler, and cleanup does the rest.
+                    await refresh_presence()
+                    last_presence_refresh = time.monotonic()
+                    await db.commit()
                 await websocket.send_json({"type": "ping"})
                 continue
+            last_frame_at = time.monotonic()
             try:
+                if last_frame_at - last_presence_refresh >= PRESENCE_REFRESH_MIN_INTERVAL_SECONDS:
+                    await refresh_presence()
+                    last_presence_refresh = last_frame_at
                 try:
                     envelope = ClientEnvelope.model_validate(raw)
                 except ValidationError:
@@ -166,7 +241,7 @@ async def chat_endpoint(websocket: WebSocket, db: AsyncSession = Depends(get_db)
                         )
                         continue
                     manager.join(envelope.room_id, websocket)
-                    await presence.join(envelope.room_id, user.id)
+                    await presence.join(envelope.room_id, user.id, conn_id)
                     joined_rooms.add(envelope.room_id)
                     await websocket.send_json({"type": "joined", "room_id": str(envelope.room_id)})
 
@@ -183,7 +258,7 @@ async def chat_endpoint(websocket: WebSocket, db: AsyncSession = Depends(get_db)
                         await websocket.send_json({"type": "error", "detail": "room_id required"})
                         continue
                     manager.leave(envelope.room_id, websocket)
-                    await presence.leave(envelope.room_id, user.id)
+                    await presence.leave(envelope.room_id, user.id, conn_id)
                     joined_rooms.discard(envelope.room_id)
 
                 elif envelope.type == "focus":
@@ -198,11 +273,11 @@ async def chat_endpoint(websocket: WebSocket, db: AsyncSession = Depends(get_db)
                         continue
                     if envelope.focused:
                         if is_blurred:
-                            await focus_presence.mark_focused(user.id)
+                            await focus_presence.mark_focused(user.id, conn_id)
                             is_blurred = False
                     else:
                         if not is_blurred:
-                            await focus_presence.mark_blurred(user.id)
+                            await focus_presence.mark_blurred(user.id, conn_id)
                             is_blurred = True
 
                 elif envelope.type == "message":
@@ -396,9 +471,11 @@ async def chat_endpoint(websocket: WebSocket, db: AsyncSession = Depends(get_db)
         manager.leave_all(websocket)
         manager.unregister_user(user.id, websocket)
         for room_id in joined_rooms:
-            await presence.leave(room_id, user.id)
+            await _best_effort("room presence", presence.leave(room_id, user.id, conn_id))
         if is_blurred:
-            await focus_presence.mark_focused(user.id)
-        if await global_presence.disconnect(user.id):
-            await broadcast_member_updated(db, broadcaster, user.id)
-            await broadcast_dm_presence_update(db, broadcaster, user.id, online=False)
+            await _best_effort("focus presence", focus_presence.mark_focused(user.id, conn_id))
+        if await _best_effort("global presence", global_presence.disconnect(user.id, conn_id)):
+            await _best_effort(
+                "offline broadcast",
+                _broadcast_offline(db, broadcaster, user.id),
+            )

backend/app/ws/expiring_presence.py

@@ -0,0 +1,103 @@
+import time
+import uuid
+from collections.abc import Callable
+
+from redis.asyncio import Redis
+
+# #77: an entry that isn't renewed for this long stops counting. Chosen as 3x
+# the app-level ping interval (chat.py's WS_PING_INTERVAL_SECONDS): a healthy
+# connection renews roughly every ping, so two missed beats are tolerated
+# before someone flips offline.
+PRESENCE_TTL_SECONDS = 90
+
+
+def _text(value: str | bytes) -> str:
+    return value.decode() if isinstance(value, bytes) else value
+
+
+def _user_part(member: str) -> str:
+    return member.split(":", 1)[0]
+
+
+class ExpiringPresenceSet:
+    """"Which users have at least one live connection" for one Redis key,
+    where every connection is its own entry with its own expiry.
+
+    This replaced per-user refcounts (#77): a refcount only goes down when a
+    connection's cleanup code actually runs, so any connection that ended
+    without it -- a server crash or reboot, a process killed mid-restart, a
+    client that vanished without the server ever noticing -- left a +1 that
+    nothing could ever subtract, showing that person online forever (and in
+    the per-room case, suppressing their push notifications). Here a dead
+    connection simply stops being renewed and ages out on its own, so there
+    is no state that can leak.
+
+    One sorted set per key: member `<user_id>:<connection_id>`, score = the
+    moment it expires. "Live" means score > now. Scans are O(connections) --
+    fine for a self-hosted team chat; only the global key grows with the
+    whole server rather than one room.
+    """
+
+    def __init__(
+        self,
+        redis: Redis,
+        ttl_seconds: float = PRESENCE_TTL_SECONDS,
+        clock: Callable[[], float] = time.time,
+    ) -> None:
+        self._redis = redis
+        self._ttl = ttl_seconds
+        self._clock = clock
+
+    @staticmethod
+    def _member(user_id: uuid.UUID, conn_id: str) -> str:
+        return f"{user_id}:{conn_id}"
+
+    async def _live_members(self, key: str) -> list[str]:
+        members = await self._redis.zrangebyscore(key, f"({self._clock()}", "+inf")
+        return [_text(m) for m in members]
+
+    async def add(self, key: str, user_id: uuid.UUID, conn_id: str) -> bool:
+        """Creates or renews this connection's entry. Returns True iff the
+        user had no live entry beforehand -- a genuine offline->online
+        transition worth telling anyone about. Renewing an entry that's still
+        live (the common case, every heartbeat) is a single cheap write."""
+        member = self._member(user_id, conn_id)
+        now = self._clock()
+        score = await self._redis.zscore(key, member)
+        came_online = False
+        if score is None or score <= now:
+            came_online = not await self.is_live(key, user_id)
+        await self._redis.zadd(key, {member: now + self._ttl})
+        # Lets a key nobody renews any more (an idle room) disappear on its
+        # own instead of sitting there holding only expired entries.
+        await self._redis.expire(key, int(self._ttl * 2))
+        return came_online
+
+    async def remove(self, key: str, user_id: uuid.UUID, conn_id: str) -> bool:
+        """Drops this connection's entry. Returns True iff that left the user
+        with no live entry -- a genuine online->offline transition. False if
+        the entry was already gone: whoever removed it (see prune) already
+        reported the transition, and reporting it twice would be noise."""
+        removed = await self._redis.zrem(key, self._member(user_id, conn_id))
+        if not removed:
+            return False
+        return not await self.is_live(key, user_id)
+
+    async def is_live(self, key: str, user_id: uuid.UUID) -> bool:
+        prefix = f"{user_id}:"
+        return any(m.startswith(prefix) for m in await self._live_members(key))
+
+    async def live_user_ids(self, key: str) -> set[uuid.UUID]:
+        return {uuid.UUID(_user_part(m)) for m in await self._live_members(key)}
+
+    async def prune(self, key: str) -> set[uuid.UUID]:
+        """Deletes expired entries. Returns the users whose last live entry
+        was among them -- the offline transitions nobody was told about,
+        since an expiry has no disconnect event to hang a broadcast on."""
+        now = self._clock()
+        expired = [_text(m) for m in await self._redis.zrangebyscore(key, "-inf", now)]
+        if not expired:
+            return set()
+        await self._redis.zremrangebyscore(key, "-inf", now)
+        gone = {_user_part(m) for m in expired} - {_user_part(m) for m in await self._live_members(key)}
+        return {uuid.UUID(u) for u in gone}

backend/app/ws/focus_presence.py

@@ -2,44 +2,45 @@ import uuid
 
 from redis.asyncio import Redis
 
+from app.ws.expiring_presence import ExpiringPresenceSet
+
 
 class FocusPresence:
     """Cross-instance "is this desktop-mode user's window currently
     unfocused," used only to widen desktop_notification/push eligibility
-    beyond plain room-connection state (#59). A Redis hash (field =
-    user_id, value = refcount of that user's currently-blurred desktop
-    connections), parallel to Presence/GlobalPresence.
+    beyond plain room-connection state (#59). One expiring entry per
+    currently-blurred desktop connection (see ExpiringPresenceSet),
+    parallel to Presence/GlobalPresence.
 
-    Absence from this hash is the default and means "focused." That default
+    Absence from the set is the default and means "focused." That default
     is also exactly right for every browser-tab connection: only the
     desktop client ever sends focus/blur frames at all (see chat.py's
     "focus" envelope handling), so a browser user never appears here --
     their attention is already fully captured by Presence's room-connection
     state, which stays visibility-gated with no separate focus signal.
 
-    Refcounted for the same multi-connection reason as Presence/
-    GlobalPresence, with the same known simplification: two desktop windows
-    for one user, one focused and one blurred, count as "unfocused" here
-    (refcount > 0) even though the user does have attention somewhere. That
-    errs toward notifying rather than silently missing one, which is the
-    safer failure mode for a notification.
+    Same known simplification as before: two desktop windows for one user,
+    one focused and one blurred, count as "unfocused" (any live blurred
+    entry) even though the user does have attention somewhere. That errs
+    toward notifying rather than silently missing one, which is the safer
+    failure mode for a notification. Entries expire (#77) -- a desktop
+    window that died while blurred used to leave its user "unfocused"
+    forever.
     """
 
-    def __init__(self, redis: Redis) -> None:
-        self._redis = redis
+    KEY = "presence:v2:unfocused"
 
-    def _key(self) -> str:
-        return "presence:unfocused"
+    def __init__(self, redis: Redis, store: ExpiringPresenceSet | None = None) -> None:
+        self._store = store or ExpiringPresenceSet(redis)
 
-    async def mark_blurred(self, user_id: uuid.UUID) -> None:
-        await self._redis.hincrby(self._key(), str(user_id), 1)
+    async def mark_blurred(self, user_id: uuid.UUID, conn_id: str) -> None:
+        await self._store.add(self.KEY, user_id, conn_id)
 
-    async def mark_focused(self, user_id: uuid.UUID) -> None:
-        key = self._key()
-        field = str(user_id)
-        remaining = await self._redis.hincrby(key, field, -1)
-        if remaining <= 0:
-            await self._redis.hdel(key, field)
+    async def mark_focused(self, user_id: uuid.UUID, conn_id: str) -> None:
+        await self._store.remove(self.KEY, user_id, conn_id)
+
+    async def refresh(self, user_id: uuid.UUID, conn_id: str) -> None:
+        await self._store.add(self.KEY, user_id, conn_id)
 
     async def is_unfocused(self, user_id: uuid.UUID) -> bool:
-        return await self._redis.hexists(self._key(), str(user_id))
+        return await self._store.is_live(self.KEY, user_id)

backend/app/ws/global_presence.py

@@ -2,50 +2,55 @@ import uuid
 
 from redis.asyncio import Redis
 
+from app.ws.expiring_presence import ExpiringPresenceSet
+
 
 class GlobalPresence:
     """Cross-instance "does this user have the app open at all right now,"
     independent of which (if any) room they currently have open -- backing
     the online/offline presence dot shown wherever a user's avatar renders.
-    A single Redis hash (field = user_id, value = connection refcount),
-    parallel to but separate from Presence's per-room hashes.
 
-    Refcounted for the same reason as Presence: multiple tabs/instances for
-    one user shouldn't flip them offline until the last connection closes.
+    Each connection has its own expiring entry (see ExpiringPresenceSet), so
+    a user with several tabs/devices stays online until the last one goes,
+    and a connection that dies without cleaning up ages out instead of
+    leaving them online forever (#77). The key is versioned (`v2`) because
+    the old refcount hash lives at `presence:global` and its leaked counts
+    must not be read as live entries.
     """
 
-    def __init__(self, redis: Redis) -> None:
-        self._redis = redis
+    KEY = "presence:v2:global"
 
-    def _key(self) -> str:
-        return "presence:global"
+    def __init__(self, redis: Redis, store: ExpiringPresenceSet | None = None) -> None:
+        self._store = store or ExpiringPresenceSet(redis)
 
-    async def connect(self, user_id: uuid.UUID) -> bool:
-        """Returns True iff this was the user's first open connection --
-        a genuine offline->online transition worth telling anyone about."""
-        count = await self._redis.hincrby(self._key(), str(user_id), 1)
-        return count == 1
+    async def connect(self, user_id: uuid.UUID, conn_id: str) -> bool:
+        """Returns True iff this made the user newly online -- a genuine
+        offline->online transition worth telling anyone about."""
+        return await self._store.add(self.KEY, user_id, conn_id)
 
-    async def disconnect(self, user_id: uuid.UUID) -> bool:
-        """Returns True iff this was the user's last open connection -- a
+    async def refresh(self, user_id: uuid.UUID, conn_id: str) -> bool:
+        """Heartbeat. Same as connect() on purpose: a connection whose entry
+        had already expired (a long stall) comes back online here, and the
+        caller needs to announce that exactly like a fresh connect."""
+        return await self._store.add(self.KEY, user_id, conn_id)
+
+    async def disconnect(self, user_id: uuid.UUID, conn_id: str) -> bool:
+        """Returns True iff that was the user's last live connection -- a
         genuine online->offline transition."""
-        key = self._key()
-        field = str(user_id)
-        remaining = await self._redis.hincrby(key, field, -1)
-        if remaining <= 0:
-            await self._redis.hdel(key, field)
-            return True
-        return False
+        return await self._store.remove(self.KEY, user_id, conn_id)
+
+    async def sweep(self) -> set[uuid.UUID]:
+        """Drops expired entries and returns the users that went offline as a
+        result -- the transitions no disconnect ever reported."""
+        return await self._store.prune(self.KEY)
 
     async def is_online(self, user_id: uuid.UUID) -> bool:
-        return await self._redis.hexists(self._key(), str(user_id))
+        return await self._store.is_live(self.KEY, user_id)
 
     async def online_user_ids(self, user_ids: list[uuid.UUID]) -> set[uuid.UUID]:
         if not user_ids:
             return set()
-        values = await self._redis.hmget(self._key(), [str(u) for u in user_ids])
-        return {uid for uid, v in zip(user_ids, values) if v is not None}
+        return set(user_ids) & await self._store.live_user_ids(self.KEY)
 
     async def all_online_user_ids(self) -> set[uuid.UUID]:
-        fields = await self._redis.hkeys(self._key())
-        return {uuid.UUID(f) for f in fields}
+        return await self._store.live_user_ids(self.KEY)

backend/app/ws/presence.py

@@ -2,38 +2,40 @@ import uuid
 
 from redis.asyncio import Redis
 
+from app.ws.expiring_presence import ExpiringPresenceSet
+
 
 class Presence:
-    """Cross-instance "who's connected to this room," backed by a Redis hash
-    per room (field = user_id, value = connection refcount).
+    """Cross-instance "who's connected to this room," one expiring set per
+    room (see ExpiringPresenceSet).
 
-    Refcounted rather than a plain set so a user with two connections to the
-    same room -- two tabs, or one per app instance -- doesn't get marked
-    offline when only one of those connections closes.
-
-    Known limitation: a hard crash (not a clean disconnect) leaks that
-    connection's increment forever, since there's no heartbeat/TTL here to
-    reclaim it -- out of scope for this phase, same category of
-    simplification as the "no server-side session revocation" note in the
-    README.
+    This is what the offline-push check reads, so a leaked entry used to mean
+    a user who'd gone away was still treated as "watching the room" and never
+    got a push notification (#77). Entries now expire unless the connection
+    keeps renewing them, so that can't outlive a dead connection by more than
+    the TTL. Keys are versioned (`v2`) so the old refcount hashes
+    (`presence:{room_id}`), with whatever they leaked, are never read.
     """
 
-    def __init__(self, redis: Redis) -> None:
-        self._redis = redis
+    def __init__(self, redis: Redis, store: ExpiringPresenceSet | None = None) -> None:
+        self._store = store or ExpiringPresenceSet(redis)
 
     def _key(self, room_id: uuid.UUID) -> str:
-        return f"presence:{room_id}"
+        return f"presence:v2:room:{room_id}"
 
-    async def join(self, room_id: uuid.UUID, user_id: uuid.UUID) -> None:
-        await self._redis.hincrby(self._key(room_id), str(user_id), 1)
-
-    async def leave(self, room_id: uuid.UUID, user_id: uuid.UUID) -> None:
+    async def join(self, room_id: uuid.UUID, user_id: uuid.UUID, conn_id: str) -> None:
         key = self._key(room_id)
-        field = str(user_id)
-        remaining = await self._redis.hincrby(key, field, -1)
-        if remaining <= 0:
-            await self._redis.hdel(key, field)
+        # Per-room offline transitions aren't broadcast anywhere, so unlike
+        # the global key there's nothing to lose by pruning here -- it just
+        # keeps a busy room's set from accumulating expired entries.
+        await self._store.prune(key)
+        await self._store.add(key, user_id, conn_id)
+
+    async def refresh(self, room_id: uuid.UUID, user_id: uuid.UUID, conn_id: str) -> None:
+        await self._store.add(self._key(room_id), user_id, conn_id)
+
+    async def leave(self, room_id: uuid.UUID, user_id: uuid.UUID, conn_id: str) -> None:
+        await self._store.remove(self._key(room_id), user_id, conn_id)
 
     async def connected_user_ids(self, room_id: uuid.UUID) -> set[uuid.UUID]:
-        fields = await self._redis.hkeys(self._key(room_id))
-        return {uuid.UUID(f) for f in fields}
+        return await self._store.live_user_ids(self._key(room_id))

backend/app/ws/presence_sweeper.py

@@ -0,0 +1,83 @@
+import asyncio
+import logging
+import uuid
+from collections.abc import Callable
+
+from redis.asyncio import Redis
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.services.message_events import broadcast_dm_presence_update, broadcast_member_updated
+from app.ws.broadcaster import Broadcaster
+from app.ws.global_presence import GlobalPresence
+
+logger = logging.getLogger(__name__)
+
+PRESENCE_SWEEP_INTERVAL_SECONDS = 15
+SWEEP_LOCK_KEY = "presence:v2:sweeper-lock"
+
+
+async def sweep_once(
+    redis: Redis,
+    global_presence: GlobalPresence,
+    broadcaster: Broadcaster,
+    session_factory: Callable[[], AsyncSession],
+    lock_seconds: int | None = max(1, PRESENCE_SWEEP_INTERVAL_SECONDS - 1),
+) -> set[uuid.UUID]:
+    """Expires dead connections' presence entries and tells everyone about
+    the users that went offline as a result.
+
+    A clean disconnect broadcasts its own offline transition (chat.py), but a
+    connection that just stops being renewed has no such event -- this is the
+    only thing that notices. Every app worker runs this loop; the Redis lock
+    makes it one sweep per interval across the whole deployment rather than
+    N workers each doing (and broadcasting) the same thing. `lock_seconds=None`
+    skips the lock (tests).
+    """
+    if lock_seconds is not None and not await redis.set(SWEEP_LOCK_KEY, "1", nx=True, ex=lock_seconds):
+        return set()
+    went_offline = await global_presence.sweep()
+    for user_id in went_offline:
+        # One user's failed broadcast must not cost everyone after them
+        # theirs -- their entry is already gone, so this is the only chance.
+        try:
+            async with session_factory() as db:
+                await broadcast_member_updated(db, broadcaster, user_id)
+                await broadcast_dm_presence_update(db, broadcaster, user_id, online=False)
+        except Exception:
+            logger.error("Failed to broadcast offline for %s", user_id, exc_info=True)
+    return went_offline
+
+
+async def run_presence_sweeper(app, interval: float = PRESENCE_SWEEP_INTERVAL_SECONDS) -> None:
+    # Same lesson as Broadcaster.run_forever (#76): a background task that
+    # dies on its first error does so silently, and here that would quietly
+    # bring the stuck-online bug back.
+    while True:
+        await asyncio.sleep(interval)
+        try:
+            await sweep_once(
+                app.state.redis,
+                app.state.global_presence,
+                app.state.broadcaster,
+                app.state.session_factory,
+                lock_seconds=max(1, int(interval) - 1),
+            )
+        except asyncio.CancelledError:
+            raise
+        except Exception:
+            logger.error("Presence sweep failed", exc_info=True)
+
+
+async def purge_legacy_presence(redis: Redis) -> int:
+    """Deletes the old refcount hashes (`presence:global`, `presence:unfocused`,
+    `presence:{room_id}`). Nothing reads them any more -- the live data is
+    under `presence:v2:` -- and they hold whatever counts leaked before #77,
+    so there's nothing worth keeping. Safe to run on every startup."""
+    deleted = 0
+    async for key in redis.scan_iter("presence:*"):
+        name = key.decode() if isinstance(key, bytes) else key
+        if name.startswith("presence:v2:"):
+            continue
+        await redis.delete(name)
+        deleted += 1
+    return deleted