fix(avatar): prevent BOXIM polling starvation

This commit is contained in:
stefanfeng
2026-09-01 14:03:35 +08:00
parent a7eb6ac2a5
commit 540edb58c4
6 changed files with 196 additions and 11 deletions
@@ -25,7 +25,8 @@ logger = logging.getLogger(__name__)
ACTIVE_TASK_STATUSES = ("pending", "generating", "ready", "sending")
GENERATABLE_TASK_STATUSES = ("pending",)
MAX_PROMPT_LENGTH = 4000
MAX_STALE_SECONDS = 120
DEFAULT_MAX_MESSAGE_AGE_SECONDS = 600
MAX_SEND_OVERDUE_SECONDS = 120
STUCK_LOCK_SECONDS = 90
TAKEOVER_PERMISSION = "takeover"
TAKEOVER_DELAY_KEY = "takeoverReplyDelaySeconds"
@@ -115,11 +116,15 @@ class TakeoverService:
boxim_client: BoxIMClient,
*,
reply_delay_seconds: int | None = None,
poll_concurrency: int = 8,
max_message_age_seconds: int = DEFAULT_MAX_MESSAGE_AGE_SECONDS,
now: Callable[[], datetime] = _utcnow,
):
self.session_factory = session_factory
self.boxim = boxim_client
self.reply_delay_seconds = reply_delay_seconds
self.poll_concurrency = max(1, min(int(poll_concurrency), 64))
self.max_message_age_seconds = max(60, int(max_message_age_seconds))
self.now = now
self._sessions: dict[str, dict] = {}
self._poll_lock = asyncio.Lock()
@@ -138,8 +143,48 @@ class TakeoverService:
self._recover_stuck_tasks()
avatar_ids = self._enabled_avatar_ids()
self._cancel_disabled_tasks(set(avatar_ids))
for avatar_id in avatar_ids:
await self._sync_avatar(avatar_id)
self._ensure_takeover_cursors(avatar_ids)
semaphore = asyncio.Semaphore(self.poll_concurrency)
async def sync(avatar_id: str):
async with semaphore:
return await self._sync_avatar(avatar_id)
results = await asyncio.gather(
*(sync(avatar_id) for avatar_id in avatar_ids),
return_exceptions=True,
)
for avatar_id, result in zip(avatar_ids, results):
if isinstance(result, Exception):
logger.warning("BOXIM poll crashed for avatar %s: %s", avatar_id, result)
def _ensure_takeover_cursors(self, avatar_ids: list[str]):
"""Create durable cursors before concurrent network polling starts."""
if not avatar_ids:
return
db = self.session_factory()
try:
existing = {
row[0]
for row in db.query(TakeoverCursor.avatar_id)
.filter(TakeoverCursor.avatar_id.in_(avatar_ids))
.all()
}
avatars = (
db.query(Avatar.id, Avatar.owner_id)
.filter(
Avatar.id.in_(
[avatar_id for avatar_id in avatar_ids if avatar_id not in existing]
)
)
.all()
)
for avatar_id, owner_id in avatars:
db.add(TakeoverCursor(avatar_id=avatar_id, owner_id=owner_id))
if avatars:
db.commit()
finally:
db.close()
async def process_reply_tasks(self):
"""Generate and send replies independently from BOXIM's long poll."""
@@ -290,7 +335,10 @@ class TakeoverService:
if not cursor:
cursor = TakeoverCursor(avatar_id=avatar.id, owner_id=avatar.owner_id)
db.add(cursor)
db.flush()
db.commit()
else:
# Release SQLite's read transaction before the long network poll.
db.commit()
if not user or not user.huihui_token:
self._record_connection_failure(
db,
@@ -452,16 +500,37 @@ class TakeoverService:
return
if not schedule_reply or event.message_type != 0 or not event.content.strip():
return
if (now - send_time).total_seconds() > MAX_STALE_SECONDS:
if (now - send_time).total_seconds() > self.max_message_age_seconds:
logger.info(
"Ignored stale BOXIM message %s for avatar %s (age=%ss)",
message_id,
avatar.id,
int((now - send_time).total_seconds()),
)
return
if is_avatar:
self._cancel_conversation(db, avatar.owner_id, peer_id, "peer_avatar_message")
logger.info(
"Skipped BOXIM reply for avatar %s message %s: peer_avatar_message",
avatar.id,
message_id,
)
return
if self._human_pause_active(db, avatar.owner_id, peer_id, now):
self._cancel_conversation(db, avatar.owner_id, peer_id, "owner_active")
logger.info(
"Skipped BOXIM reply for avatar %s message %s: owner_active",
avatar.id,
message_id,
)
return
if self._conversation_rate_limited(db, avatar.owner_id, peer_id, now):
self._cancel_conversation(db, avatar.owner_id, peer_id, "rate_limited")
logger.info(
"Skipped BOXIM reply for avatar %s message %s: rate_limited",
avatar.id,
message_id,
)
return
self._schedule_reply(db, avatar, event)
@@ -540,8 +609,10 @@ class TakeoverService:
prompt_parts.append(event.content.strip())
source_ids.append(event.boxim_message_id)
prompt = "\n".join(part for part in prompt_parts if part).strip()[-MAX_PROMPT_LENGTH:]
due_at = event.send_time + timedelta(
seconds=_configured_reply_delay(avatar, self.reply_delay_seconds)
due_at = max(
event.send_time
+ timedelta(seconds=_configured_reply_delay(avatar, self.reply_delay_seconds)),
self.now(),
)
task_id = secrets.token_hex(16)
local_id = _avatar_local_id(avatar.owner_id, event.boxim_message_id)
@@ -701,7 +772,7 @@ class TakeoverService:
task.cancel_reason = "takeover_disabled"
db.commit()
return False
if (self.now() - task.scheduled_at).total_seconds() > MAX_STALE_SECONDS:
if (self.now() - task.scheduled_at).total_seconds() > MAX_SEND_OVERDUE_SECONDS:
task.status = "cancelled"
task.cancel_reason = "stale_reply"
db.commit()