A streaming platform is adding scheduled live events to a catalogue service that has always served on-demand content. Which quality attributes change, and why is this not "the same system with more traffic"?
Show the full answer Hide the answer
Why it is a different system
On-demand traffic is statistically smooth: viewers arrive independently, caches warm gradually, and a capacity error is absorbed by the fact that no two users want the same byte at the same instant.
A live event inverts every one of those properties:
- Arrival is correlated. Everyone joins within a few minutes of the same moment. There is no gradual warm-up, so cold caches, connection establishment, and authentication all peak simultaneously.
- The content cannot be pre-positioned in the usual sense. The video does not exist yet. Only the manifests, players, keys and static assets can be pre-cached; the segments are produced seconds before they are requested, which puts the origin and ingest path on the critical path in a way on-demand never does.
- There is no retry. A failure during on-demand playback is an annoyance; the user restarts. A failure during the deciding minute of a live event is unrecoverable — the moment is gone.
- Failure is correlated and public. Everyone fails at the same time, at the same moment, and says so.
The quality attributes that move
| Attribute | On-demand | Live event |
|---|---|---|
| Peak-to-average | Modest, smooth | Extreme, scheduled, correlated |
| Latency that matters | Start-up time | Glass-to-glass delay and rebuffer ratio |
| Availability window | Continuous | A specific three-hour window that dominates the year |
| Degradation options | Few needed | Essential and must be designed in advance |
| Cacheability | High | Segments have seconds of usable life |
The design consequences
Pre-scaling rather than autoscaling, because autoscaling reacts on a timescale longer than the arrival spike. Ingest redundancy with independent paths, since the encode chain is now a single point of failure for the whole event. A shortened bitrate ladder with a defined floor, so that congestion produces lower quality rather than rebuffering. A rehearsed degradation list — chat, statistics overlays, recommendations, personalised rails — with owners and kill switches, exercised before the event rather than discovered during it.
The judgement
The forecast is the least reliable input. Live audiences have repeatedly exceeded planned concurrency by large multiples. The architecture should be built so that exceeding the forecast degrades quality gradually rather than failing a fraction of viewers completely — which means the interesting design work is in the behaviour above the plan, not at it.