LIVEKIT SERVER LIVEKIT INGRESS and LIVEKIT EGRESS WORK EFFICIENTLY FOR WEBRTC

Self Hosting
1

Hey everyone!

Just wanted to share a quick update and some real-world benchmark metrics from a live production run with my self-hosted setup (itg-media-engine PWA).

We just wrapped up two live interviews using the PWA to pull remote guests into OBS Studio, routing the feed back into the LiveKit room via Ingress, and handling multiplatform live streams out via Egress.

Everything ran remarkably smooth. Outside of a minor user error on my end (forgot to toggle my wide format layout, so the social stream went out in default portrait!), the underlying engine and LiveKit infrastructure performed flawlessly.

18-Minute Broadcast Log Highlights

Here is the log overview from our latest 18-minute live session:

  • Video Processing:

    • Frames Processed: ~1.5M frames

    • Queue & Buffer Drops: 0 drops

    • Delivery: Perfect frame delivery with zero gaps or sync drift.

  • Audio Processing:

    • Frames Processed: ~54k frames

    • Buffer Drops: Only 2 buffers dropped (~17ms total—completely inaudible)

    • Queue Drops: 0 drops

  • Ingress Stream Metrics:

    • Video Bitrate: ~12 Mbps H.264

    • Audio Bitrate: ~255 kbps Opus

    • Video Jitter (p50): 0.01ms

    • Picture Loss: Zero

  • System Health & Teardown:

    • Shutdown: Clean Source closed with Exit Code 0.

    • Process Management (PM2): All background services remained completely stable (Django/Daphne production services sitting on 2+ days uptime without a hitch).

Takeaways

Seeing 0.01ms median video jitter over a 12 Mbps H.264 ingress stream while simultaneously running Egress pipeline distribution shows just how efficient the LiveKit ecosystem is when properly tuned on bare metal / self-hosted environments.

Huge thanks to the LiveKit team and community for building such a robust suite of tools. If anyone is working on similar custom WebRTC → OBS → Multicast pipelines, happy to share notes on the setup!

2

Now, when I can bump up to a server with a GPU, I will have my full app running. I have an ai assistant for analyzing Football Video, and I have a talking avatar agent, which is powered by my Open Ai Qwen model. This model is currently ony used for my home page chat bubble, and I have the other developed. Though, still working on the Video Analysis. Yet, my first code for it is on my github, and it does run. Just too much for my current server.

I am sure excited to get what I have running, and when I am able to run: pm2 start ecosystem.fullconfig.js I will be thrilled. Talking Ai Avatar Agents are no doubt a Huge Market, and I believe the talking avatar Ocean has enough fish for us all to eat well. I’m still a firm believer that LiveKit has the market for powering these talking Avatars. Which means they are holding the golden ticket. BRAVO LIVEKIT!