The business asks for "100% uptime" for a new customer portal. Walk me through the conversation that ends in an agreed SLO.
Show the full answer Hide the answer
What the interviewer is testing
Whether you can run a negotiation that ends in a number both sides own, rather than either capitulating or lecturing about nines.
The conversation, in order
1. Do not argue with the request. Price it. "100%" is not a bad-faith answer; it is what anyone says when reliability is free. Make it not free:
| Target | Downtime / month | Roughly what it requires |
|---|---|---|
| 99% | 7h 18m | One region, single instances, manual recovery |
| 99.9% | 43m | Multi-AZ, redundancy, automated in-region failover |
| 99.95% | 21m | The above, plus tight change control |
| 99.99% | 4m 21s | Automated cross-region failover, no single points of failure |
Then: "Each additional nine costs roughly an order of magnitude more. Which of these is worth what it costs?" That reframes it from a virtue to a purchase.
2. Establish what an outage actually costs them. Revenue per hour, support call volume, contractual penalties, reputational or regulatory consequence. Now both sides are working with the same numbers, and the target follows from arithmetic rather than from who is more insistent.
3. Show the dependency ceiling. If the portal depends synchronously on three internal services at 99.9% each, its ceiling is around 99.6% no matter what the portal team does. This is usually the moment the conversation becomes productive, because it turns "make it more reliable" into a specific, actionable question: which of these dependencies do we remove from the critical path?
4. Define the SLI before the SLO. "Uptime" is not measurable. Agree what a good event is — "returns a valid response in under 500 ms" — what counts as valid, and where it is measured. The business often cares about something different from what engineering measures: a portal that is technically up but where search is broken is down, to them.
5. Propose a target below the pain threshold, and an error budget policy. "99.9%, measured monthly, with the agreement that while budget remains we ship features, and when it is exhausted we stop feature work until it recovers." The policy is the part that makes the number mean anything, and agreeing it in advance is what stops it being relitigated during an incident.
6. Keep the SLA looser than the SLO. If the customer contract promises 99.9% and the internal target is also 99.9%, the first internal miss is a commercial event and the team has no room to learn.
7. Agree a review date. Start conservatively, measure for a quarter, tighten if it is comfortably met. A target never missed is a target set too low.
The sentence that usually lands
"We can build to any of these. What I need from you is which one is worth the cost — and the agreement that when we have spent the budget for the month, we stop shipping and stabilise."
What a strong answer adds
Distinguishing SLI, SLO and SLA precisely, and noting that the negotiation is really about the error budget policy rather than the number. Two organisations with identical 99.9% targets behave completely differently depending on whether anything happens when the budget runs out.