Reliability Perfected: 99.99% Uptime for DeHub Streaming

By DeHub Team — 2025-03-30

In Q1 2025 we reached 99.99% uptime across all streaming services.

What the extra nine costs

Our 2022 stream app shipped at 99.9% — roughly eight and a half hours of downtime a year. 99.99% is about 52 minutes.

Getting from the first to the second is not ten percent more work. It is a different engineering posture. At three nines you can recover from most failures by responding to them competently. At four, the response has to be automatic, because the entire annual budget is smaller than the time it takes a human to be paged, orient and act.

That means redundancy at every point where a single failure would be visible, failover that happens without a decision, and — the expensive part — regularly proving the failover works, because untested failover is a belief rather than a property.

Why bother

Because of the asymmetry we identified in 2022 and have never revised.

A creator who goes live to a scheduled audience and cannot broadcast has not been delayed. That audience is gone, and the creator spends their own credibility explaining a failure that was not theirs. Do it twice and they stop scheduling on your platform, which is a rational response.

Fifty-two minutes a year is roughly one bad stream. Eight and a half hours is potentially dozens, and they do not land conveniently — they land during the events that matter, because that is when load is highest.

"All streaming services" is the load-bearing phrase

Not the streams. Everything a live broadcast depends on: ingest, transcoding, delivery, chat, the tipping path, the authentication in front of all of it.

A stream that stays up while tipping is down is not up in any sense a creator cares about. Composite availability is much harder than any single component's, because the failure probabilities compound, and quoting a number for the easy component is a common way to publish a figure that is technically true and practically meaningless.

What made it possible

Everything shipped around it. Livepeer for concurrency, the custom CDN, the 200% backend improvement the previous quarter, and the distributed delivery work begun with DePIN Phase 1.

Distribution is what made four nines realistic. Independently operated infrastructure has no single component whose failure takes everything with it, which is the property centralised delivery cannot offer at any price.

The caveat we keep

Uptime is measured, and measurement is a choice. Ours covers the composite service rather than the most flattering component, and we would rather report a harder number honestly than a soft one impressively.

The standard has not changed since 2022: if a creator's stream fails, it is our failure, whatever the cause.

What we are not claiming

That nothing will go wrong. Four nines is 52 minutes a year, not zero, and a creator who loses a stream inside those 52 minutes is not consoled by the annual figure.

We are also not claiming that the number will hold automatically. Availability degrades quietly as a system grows — new dependencies, new failure paths, the slow accumulation of things that have never been tested together. Reaching four nines once and reporting it forever is a common and dishonest pattern.

It is measured continuously and it is reported as measured, including in periods where it dips. An availability figure that only appears in announcements is a marketing artefact, and the difference between that and an operating standard is whether it survives a bad quarter.