Live Streaming: What Breaks at Scale

Live video is unforgiving in a way on-demand is not. There is no retry, the audience arrives all at once, and every problem is witnessed live.

The load shape

On-demand traffic is smooth. Live traffic is a step function: almost nothing, then everyone, at a time announced in advance. Capacity planned against averages fails at the join.

The join is the worst moment — every viewer requests the manifest and first segments simultaneously, the heaviest thing they will do all session.

Predictable failure modes

• Origin overload at kickoff. If the CDN has not cached the first segments, every request reaches your origin at once.
• Latency drift. Viewers fall behind live, which matters enormously when social media is reacting ahead of them.
• Bitrate ladder gaps. Too few rungs and viewers on poor connections oscillate instead of settling.
• Ad break desync. Server-side insertion that drifts returns viewers mid-action.
• The tail after the event. Demand for clips spikes exactly when the live team is exhausted.

What helps

• Pre-warm the CDN before the announced start.
• Test at the join, not in steady state. Steady state is the easy part.
• Degrade deliberately. Decide in advance what to shed under load — usually quality rungs before availability.
• Plan the clip pipeline before the event, because that demand is guaranteed and time-limited.

The clip problem

That last point is where most operations fall down. The moments people want are worth most within minutes, which is shorter than a manual editing pass. Working it out during the event is too late; it has to exist beforehand.