Memory leak on Ubuntu with Pipecat 0.0.85

[draftin1]

pipecat version

0.0.85

Python version

3.12.8

Operating System

Ubuntu 24.04.3 LTS

Use Case Description

I am running pipecat (for a year), with a pipeline of deepgram, openai, elevenlabs through livekit.

Current Approach

This issue occurs only when running in the pods on the cluster. 
I can run version 0.0.85 just fine from a macOS.

The memory leak occurs with 0.0.86 and 0.0.92. (Those are the only two other version I've checked.

I incremented from v0.0.80, where I had had it, up one at a time, and .81, .82, .83 and .84 are all fine.

With the exact same code, except for the change in pipecat version, the memory leak occurs.

I updated the deepgram, openai, elevenlabs, and livekit packages from where they were to the latest, and there's still the leak.

Errors or Unexpected Behavior

When I go from pipecat version 0.0.84 to 0.0.85 there is a memory issue that makes the instance increase its memory usage about 3GB/minute.

Additional Context

Pipeline Details Services:

• livekit
• deepgram
• openai
• elevenlabs
• krisp noise filter
• silero vad analyzer

I wonder if anyone here has experienced anything similar, or know what could be happening

[markbackman]

There are no known memory leaks in Pipecat. To look into this, you'll need to provide more detailed information on what you're doing and what you're seeing.

[draftin1]

I think I located the memory leak to exist in the LiveKitTransportClient.

v0.0.85 introduced video streaming support to LiveKitTransport (via CHANGELOG).

I am using livekit with a video track, and this is why I am seeing the issue. I have written some code to ignore the video track, and it has fixed the issue for me.

I used this prompt with the whole transports.livekit.transport file: Do you see any potential issues in this code that may cause memory issues if a video track was subscribed? and it surfaced this issue.

Issue:

• _process_video_stream is always started when a video track is subscribed (via _on_track_subscribed_wrapper), and it always puts into _video_queue.
• The consumer (_video_in_task_handler) only runs if params.video_in_enabled is True.

So if a video track is subscribed and video_in_enabled is False (or your input pipeline is stopped/blocked), _video_queue becomes an unbounded buffer for raw VideoFrameEvents → memory will keep growing as long as the track is active.

Even with video_in_enabled=True, if your downstream processing (conversion, push_video_frame, etc.) can’t keep up with LiveKit’s video rate, the queue can still grow without bound.