Improve Audio Recording Performance by Removing Blocking I/O from Callback Path

[nikhil1205-ai]

This PR refactors the audio recording pipeline in AudioSink to eliminate blocking disk I/O from the real-time audio callback path.

The previous implementation performed synchronous writes (wave.writeframes) inside the write() method under a global lock, which could introduce latency and packet loss during multi-user voice activity.

Problem #19

The existing implementation:

with self._lock:
    writer.writeframes(data)

Issues:

• Blocking disk I/O in real-time audio ingestion path
• Global lock causes contention across multiple users
• Poor scalability for concurrent speakers
• Risk of dropped or delayed audio packets

Solution

This PR introduces a queue-based buffering system with a background worker thread:

Changes:

• Audio data is now enqueued in write() instead of written directly
• A dedicated worker thread processes the queue and performs disk writes
• Lock usage is minimized to only protect shared state (_user_writers)
• Disk I/O is performed outside the lock

Implementation Details

1. Queue-based buffering
   self._queue = queue.Queue(maxsize=1000)

2. Non-blocking write path
   self._queue.put((user_id, user, data), block=False)

3. Background worker thread

def _worker(self):
    while True:
        item = self._queue.get()
        if item is None:
            break
        ...
        writer.writeframes(data)

4. Graceful shutdown

self._queue.put(None)
self._thread.join()

Benefits

• Eliminates blocking operations from audio callback path
• Improves performance in multi-user voice scenarios
• Reduces risk of audio packet loss
• Scales better under concurrent load
• Cleaner separation of concerns (ingestion vs I/O)