A stakeholder wants a video streaming service at 99.99% availability. Walk through whether that is the right target and what it would take.

What the interviewer is testing

Whether you interrogate an availability target rather than accepting it, and whether you can price each level.

The first question

What is availability, for this service? For streaming it is almost certainly not "the API responded". It is "a member could start playing something", which is a journey spanning authentication, catalogue, licensing, personalisation and content delivery.

Defining it as journey success rather than component uptime changes the number and changes what you would build.

The second question

What does an hour of downtime cost? 99.99% is 52 minutes a year; 99.9% is 8.8 hours. The gap between them is expensive, and each additional nine is roughly an order of magnitude more expensive than the last.

For a subscription streaming service the answer is genuinely large — outages generate cancellations and are extremely visible — which may well justify it. For an internal tool the same target is over-engineering that nobody asked for.

What 99.99% requires

Multi-region active-active, because a single-region deployment cannot survive a regional event within 52 minutes a year of budget.

Standing capacity to absorb a region's traffic, which means running well below full utilisation permanently.

Soft dependencies throughout, since serial hard dependencies multiply and a fan-out of a dozen services at 99.9% each cannot reach four nines by any effort.

Rehearsed regional evacuation, because a failover capability that has never been exercised is an essay.

Deployment that cannot take out all regions, which means wave-based rollout with bake time.

What a strong answer adds

Presenting it as a costed menu rather than a yes or no: what the current architecture delivers, what 99.9% would require, what 99.99% would require, and the ongoing cost of each. That converts an argument into a decision the stakeholder can make.

And the counterweight: an error budget policy, so a consistently unspent budget is treated as evidence the target is too conservative. Over-reliability is a real cost, and it is the direction nobody ever acts on.

Common weak answers

Accepting the number and adding redundancy. Refusing without pricing the alternatives.