feat(api): Introduce Broadcast Channel (#27835)

This PR introduces a `BroadcastChannel` abstraction with broadcasting and at-most once delivery semantics, serving as the communication component between celery worker and API server.

It also includes a reference implementation backed by Redis PubSub.

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
Co-authored-by: autofix-ci[bot] <114827586+autofix-ci[bot]@users.noreply.github.com>

api/libs/broadcast_channel/channel.py

@@ -0,0 +1,134 @@
+"""
+Broadcast channel for Pub/Sub messaging.
+"""
+
+import types
+from abc import abstractmethod
+from collections.abc import Iterator
+from contextlib import AbstractContextManager
+from typing import Protocol, Self
+
+
+class Subscription(AbstractContextManager["Subscription"], Protocol):
+    """A subscription to a topic that provides an iterator over received messages.
+    The subscription can be used as a context manager and will automatically
+    close when exiting the context.
+
+    Note: `Subscription` instances are not thread-safe. Each thread should create its own
+    subscription.
+    """
+
+    @abstractmethod
+    def __iter__(self) -> Iterator[bytes]:
+        """`__iter__` returns an iterator used to consume the message from this subscription.
+
+        If the caller did not enter the context, `__iter__` may lazily perform the setup before
+        yielding messages; otherwise `__enter__` handles it.”
+
+        If the subscription is closed, then the returned  iterator exits without
+        raising any error.
+        """
+        ...
+
+    @abstractmethod
+    def close(self) -> None:
+        """close closes the subscription, releases any resources associated with it."""
+        ...
+
+    def __enter__(self) -> Self:
+        """`__enter__` does the setup logic of the subscription (if any), and return itself."""
+        return self
+
+    def __exit__(
+        self,
+        exc_type: type[BaseException] | None,
+        exc_value: BaseException | None,
+        traceback: types.TracebackType | None,
+    ) -> bool | None:
+        self.close()
+        return None
+
+    @abstractmethod
+    def receive(self, timeout: float | None = 0.1) -> bytes | None:
+        """Receive the next message from the broadcast channel.
+
+        If `timeout` is specified, this method returns `None` if no message is
+        received within the given period. If `timeout` is `None`, the call blocks
+        until a message is received.
+
+        Calling receive with `timeout=None` is highly discouraged, as it is impossible to
+        cancel a blocking subscription.
+
+        :param timeout: timeout for receive message, in seconds.
+
+        Returns:
+            bytes: The received message as a byte string, or
+            None: If the timeout expires before a message is received.
+
+        Raises:
+            SubscriptionClosed: If the subscription has already been closed.
+        """
+        ...
+
+
+class Producer(Protocol):
+    """Producer is an interface for message publishing. It is already bound to a specific topic.
+
+    `Producer` implementations must be thread-safe and support concurrent use by multiple threads.
+    """
+
+    @abstractmethod
+    def publish(self, payload: bytes) -> None:
+        """Publish a message to the bounded topic."""
+        ...
+
+
+class Subscriber(Protocol):
+    """Subscriber is an interface for subscription creation. It is already bound to a specific topic.
+
+    `Subscriber` implementations must be thread-safe and support concurrent use by multiple threads.
+    """
+
+    @abstractmethod
+    def subscribe(self) -> Subscription:
+        pass
+
+
+class Topic(Producer, Subscriber, Protocol):
+    """A named channel for publishing and subscribing to messages.
+
+    Topics provide both read and write access. For restricted access,
+    use as_producer() for write-only view or as_subscriber() for read-only view.
+
+    `Topic` implementations must be thread-safe and support concurrent use by multiple threads.
+    """
+
+    @abstractmethod
+    def as_producer(self) -> Producer:
+        """as_producer creates a write-only view for this topic."""
+        ...
+
+    @abstractmethod
+    def as_subscriber(self) -> Subscriber:
+        """as_subscriber create a read-only view for this topic."""
+        ...
+
+
+class BroadcastChannel(Protocol):
+    """A broadcasting channel is a channel supporting broadcasting semantics.
+
+    Each channel is identified by a topic, different topics are isolated and do not affect each other.
+
+    There can be multiple subscriptions to a specific topic. When a publisher publishes a message to
+    a specific topic, all subscription should receive the published message.
+
+    There are no restriction for the persistence of messages. Once a subscription is created, it
+    should receive all subsequent messages published.
+
+    `BroadcastChannel` implementations must be thread-safe and support concurrent use by multiple threads.
+    """
+
+    @abstractmethod
+    def topic(self, topic: str) -> "Topic":
+        """topic returns a `Topic` instance for the given topic name."""
+        ...

api/libs/broadcast_channel/exc.py

@@ -0,0 +1,12 @@
+class BroadcastChannelError(Exception):
+    """`BroadcastChannelError` is the base class for all exceptions related
+    to `BroadcastChannel`."""
+
+    pass
+
+
+class SubscriptionClosedError(BroadcastChannelError):
+    """SubscriptionClosedError means that the subscription has been closed and
+    methods for consuming messages should not be called."""
+
+    pass

api/libs/broadcast_channel/redis/__init__.py

@@ -0,0 +1,3 @@
+from .channel import BroadcastChannel
+
+__all__ = ["BroadcastChannel"]

api/libs/broadcast_channel/redis/channel.py

@@ -0,0 +1,200 @@
+import logging
+import queue
+import threading
+import types
+from collections.abc import Generator, Iterator
+from typing import Self
+
+from libs.broadcast_channel.channel import Producer, Subscriber, Subscription
+from libs.broadcast_channel.exc import SubscriptionClosedError
+from redis import Redis
+from redis.client import PubSub
+
+_logger = logging.getLogger(__name__)
+
+
+class BroadcastChannel:
+    """
+    Redis Pub/Sub based broadcast channel implementation.
+
+    Provides "at most once" delivery semantics for messages published to channels.
+    Uses Redis PUBLISH/SUBSCRIBE commands for real-time message delivery.
+
+    The `redis_client` used to construct BroadcastChannel should have `decode_responses` set to `False`.
+    """
+
+    def __init__(
+        self,
+        redis_client: Redis,
+    ):
+        self._client = redis_client
+
+    def topic(self, topic: str) -> "Topic":
+        return Topic(self._client, topic)
+
+
+class Topic:
+    def __init__(self, redis_client: Redis, topic: str):
+        self._client = redis_client
+        self._topic = topic
+
+    def as_producer(self) -> Producer:
+        return self
+
+    def publish(self, payload: bytes) -> None:
+        self._client.publish(self._topic, payload)
+
+    def as_subscriber(self) -> Subscriber:
+        return self
+
+    def subscribe(self) -> Subscription:
+        return _RedisSubscription(
+            pubsub=self._client.pubsub(),
+            topic=self._topic,
+        )
+
+
+class _RedisSubscription(Subscription):
+    def __init__(
+        self,
+        pubsub: PubSub,
+        topic: str,
+    ):
+        # The _pubsub is None only if the subscription is closed.
+        self._pubsub: PubSub | None = pubsub
+        self._topic = topic
+        self._closed = threading.Event()
+        self._queue: queue.Queue[bytes] = queue.Queue(maxsize=1024)
+        self._dropped_count = 0
+        self._listener_thread: threading.Thread | None = None
+        self._start_lock = threading.Lock()
+        self._started = False
+
+    def _start_if_needed(self) -> None:
+        with self._start_lock:
+            if self._started:
+                return
+            if self._closed.is_set():
+                raise SubscriptionClosedError("The Redis subscription is closed")
+            if self._pubsub is None:
+                raise SubscriptionClosedError("The Redis subscription has been cleaned up")
+
+            self._pubsub.subscribe(self._topic)
+            _logger.debug("Subscribed to channel %s", self._topic)
+
+            self._listener_thread = threading.Thread(
+                target=self._listen,
+                name=f"redis-broadcast-{self._topic}",
+                daemon=True,
+            )
+            self._listener_thread.start()
+            self._started = True
+
+    def _listen(self) -> None:
+        pubsub = self._pubsub
+        assert pubsub is not None, "PubSub should not be None while starting listening."
+        while not self._closed.is_set():
+            raw_message = pubsub.get_message(ignore_subscribe_messages=True, timeout=0.1)
+
+            if raw_message is None:
+                continue
+
+            if raw_message.get("type") != "message":
+                continue
+
+            channel_field = raw_message.get("channel")
+            if isinstance(channel_field, bytes):
+                channel_name = channel_field.decode("utf-8")
+            elif isinstance(channel_field, str):
+                channel_name = channel_field
+            else:
+                channel_name = str(channel_field)
+
+            if channel_name != self._topic:
+                _logger.warning("Ignoring message from unexpected channel %s", channel_name)
+                continue
+
+            payload_bytes: bytes | None = raw_message.get("data")
+            if not isinstance(payload_bytes, bytes):
+                _logger.error("Received invalid data from channel %s, type=%s", self._topic, type(payload_bytes))
+                continue
+
+            self._enqueue_message(payload_bytes)
+
+        _logger.debug("Listener thread stopped for channel %s", self._topic)
+        pubsub.unsubscribe(self._topic)
+        pubsub.close()
+        _logger.debug("PubSub closed for topic %s", self._topic)
+        self._pubsub = None
+
+    def _enqueue_message(self, payload: bytes) -> None:
+        while not self._closed.is_set():
+            try:
+                self._queue.put_nowait(payload)
+                return
+            except queue.Full:
+                try:
+                    self._queue.get_nowait()
+                    self._dropped_count += 1
+                    _logger.debug(
+                        "Dropped message from Redis subscription, topic=%s, total_dropped=%d",
+                        self._topic,
+                        self._dropped_count,
+                    )
+                except queue.Empty:
+                    continue
+        return
+
+    def _message_iterator(self) -> Generator[bytes, None, None]:
+        while not self._closed.is_set():
+            try:
+                item = self._queue.get(timeout=0.1)
+            except queue.Empty:
+                continue
+
+            yield item
+
+    def __iter__(self) -> Iterator[bytes]:
+        if self._closed.is_set():
+            raise SubscriptionClosedError("The Redis subscription is closed")
+        self._start_if_needed()
+        return iter(self._message_iterator())
+
+    def receive(self, timeout: float | None = None) -> bytes | None:
+        if self._closed.is_set():
+            raise SubscriptionClosedError("The Redis subscription is closed")
+        self._start_if_needed()
+
+        try:
+            item = self._queue.get(timeout=timeout)
+        except queue.Empty:
+            return None
+
+        return item
+
+    def __enter__(self) -> Self:
+        self._start_if_needed()
+        return self
+
+    def __exit__(
+        self,
+        exc_type: type[BaseException] | None,
+        exc_value: BaseException | None,
+        traceback: types.TracebackType | None,
+    ) -> bool | None:
+        self.close()
+        return None
+
+    def close(self) -> None:
+        if self._closed.is_set():
+            return
+
+        self._closed.set()
+        # NOTE: PubSub is not thread-safe. More specifically, the `PubSub.close` method and the `PubSub.get_message`
+        # method should NOT be called concurrently.
+        #
+        # Due to the restriction above, the PubSub cleanup logic happens inside the consumer thread.
+        listener = self._listener_thread
+        if listener is not None:
+            listener.join(timeout=1.0)
+            self._listener_thread = None