️(backend) refactor presence cache to bound key lookups per room

The previous presence cache lookup keyed off a scan over the whole
cache, so its cost was O(db_size) rather than O(room_size).
Combined with the recent switch to cursor-based `SCAN` at an
inappropriate page size, this caused a lot of Redis round-trips and
noticeably slowed down the backend pods under load.

Refactor the presence cache to keep a per-room set of all its
participant keys. Lookups now iterate that set instead of scanning
the whole database.

Complexity is now bounded by room size, not database size, which
should restore the backend performance to its previous levels while
keeping the lobby behavior unchanged.
This commit is contained in:
lebaudantoine
2026-09-04 23:02:23 +02:00
parent 2d392d31f8
commit c614085b81
4 changed files with 55 additions and 31 deletions
+1
View File
@@ -29,6 +29,7 @@ and this project adheres to
- ⚡️(frontend) increase lobby polling interval on both sides
- ⚡️(frontend) add trailing slash on the /me endpoint call
- ⚡️(backend) refactor lobby storage to bound key lookups per room
- ⚡️(backend) refactor presence cache to bound key lookups per room
## [1.30.0] - 2026-09-01
+48 -21
View File
@@ -1,25 +1,11 @@
"""Presence cache.
Redis-backed memo of "this identity is currently connected to this room".
This module is intentionally a *pure cache store* with no dependency on other
services, so that `participants_management` (which talks to LiveKit) can
import it without creating an import cycle. The composition of "check cache,
fall back to LiveKit" lives in
`ParticipantsManagement.check_if_in_meeting_cached`.
Only positive answers are stored: a sticky negative would lock out someone
who joins right after a miss for the whole TTL. The TTL is a safety net in
case an invalidation webhook is lost.
"""
"""Presence cache."""
from typing import FrozenSet
from uuid import UUID
from django.conf import settings
from django.core.cache import cache
from core.utils import CACHE_SCAN_ITERSIZE
class PresenceCache:
"""Store and invalidate (room, identity) presence entries."""
@@ -29,24 +15,65 @@ class PresenceCache:
"""Cache key for a (room, identity) presence entry."""
return f"{settings.PRESENCE_KEY_PREFIX}_{room_id!s}_{identity}"
@staticmethod
def _get_index_key(room_id: UUID | str) -> str:
"""Raw Redis key of the per-room identity index (a native SET).
Built through django-redis' make_key so it lives under the same
KEY_PREFIX/version namespace as the presence entries.
"""
return cache.client.make_key(
f"{settings.PRESENCE_KEY_PREFIX}-index_{room_id!s}"
)
@staticmethod
def _redis(write: bool = True):
"""Raw redis-py client.
SADD/SREM/SMEMBERS are not exposed by the Django cache API; this is
the documented django-redis escape hatch.
"""
return cache.client.get_client(write=write)
def _index_members(self, room_id: UUID | str) -> FrozenSet[str]:
"""All identities currently indexed for the room."""
members = self._redis(write=False).smembers(self._get_index_key(room_id))
return frozenset(
member.decode() if isinstance(member, bytes) else member
for member in members
)
def is_marked_present(self, room_id: UUID | str, identity: str) -> bool:
"""Return True if a positive presence entry exists in cache."""
return bool(cache.get(self._get_cache_key(room_id, identity)))
def mark_present(self, room_id: UUID | str, identity: str) -> None:
"""Record that `identity` is in `room_id`."""
"""Record that `identity` is in `room_id` and index it for the room."""
cache.set(
self._get_cache_key(room_id, identity),
True,
timeout=settings.PRESENCE_CACHE_TIMEOUT,
)
index_key = self._get_index_key(room_id)
pipe = self._redis().pipeline(transaction=False)
pipe.sadd(index_key, identity)
pipe.expire(index_key, settings.PRESENCE_CACHE_TIMEOUT)
pipe.execute()
def clear(self, room_id: UUID | str, identity: str) -> None:
"""Forget presence for one participant (e.g. on participant_left)."""
cache.delete(self._get_cache_key(room_id, identity))
self._redis().srem(self._get_index_key(room_id), identity)
def clear_room(self, room_id: UUID | str) -> None:
"""Forget presence for every participant of a room (on room_finished)."""
cache.delete_pattern(
self._get_cache_key(room_id, "*"), itersize=CACHE_SCAN_ITERSIZE
)
"""Forget presence for every participant of a room (on room_finished).
Deletes the indexed entries and the index itself with targeted
commands instead of a full-keyspace pattern scan.
"""
identities = self._index_members(room_id)
if identities:
cache.delete_many(
[self._get_cache_key(room_id, identity) for identity in identities]
)
self._redis().delete(self._get_index_key(room_id))
@@ -90,19 +90,18 @@ def test_presence_clear_and_clear_room():
assert presence.is_marked_present(other_room, "a") is True
def test_presence_clear_room_scans_in_pages():
"""clear_room removes every match, even across several SCAN pages,
and only within the room."""
def test_presence_clear_room_removes_many_entries_and_the_index():
"""clear_room removes every entry of the room through the index — never
a keyspace scan — and leaves other rooms untouched."""
room_id, other_room = str(uuid4()), str(uuid4())
presence = PresenceCache()
for i in range(7):
presence.mark_present(room_id, f"user-{i}")
presence.mark_present(other_room, "user-0")
# An itersize smaller than the match count forces delete_pattern to
# page through several SCAN cursors rather than finish in one pass.
with mock.patch("core.utils.CACHE_SCAN_ITERSIZE", 3):
presence.clear_room(room_id)
presence.clear_room(room_id)
assert all(not presence.is_marked_present(room_id, f"user-{i}") for i in range(7))
assert presence.is_marked_present(other_room, "user-0") is True
assert presence._index_members(room_id) == frozenset([])
assert presence._index_members(other_room) == frozenset(["user-0"])
-3
View File
@@ -499,6 +499,3 @@ def build_telephony_config():
"default_country": country,
"international_phone_number": international,
}
CACHE_SCAN_ITERSIZE = 500