A team consistently ends every window with 90% of its error budget unspent. What does that tell you?
Show the full answer Hide the answer
What is being tested
Whether you know the symmetric half of the error budget mechanism, which almost nobody uses.
What it tells you
The team is over-invested in reliability at the expense of delivery, or the SLO is set far below what the architecture actually delivers.
Either way, the correct response is to take more risk: deploy more often, reduce staging ceremony, ship changes that were being held back, remove a manual verification step.
100% reliability is not the goal. An unspent budget is not a gold star — it is capacity that could have been spent on delivery and was not.
Why this is under-used
The budget is almost always framed as a constraint: exhaust it and feature work stops. The symmetric half — budget remaining means go faster — is rarely stated, so teams treat consistently unspent budget as success and nobody asks what it cost.
How to check which explanation applies
Is the SLO meaningfully derived? If it was set at 99.9% by aspiration rather than from what users need and what the business loses, it may simply be far below actual performance and therefore uninformative in either direction.
Set the target from the cost of unavailability: revenue lost per minute (and whether it is lost or merely deferred — an e-commerce sale during a five-minute outage may simply happen five minutes later), contractual penalties, regulatory consequences, internal cost.
Is the team holding back changes? Ask what they would ship if a bad deployment were cheap. The answer usually reveals the held capacity.
The framing that makes it actionable
The error budget is an allowance to spend, not a limit to avoid. A team spending 60–80% of it across windows is calibrated: reliable enough, and moving as fast as that reliability permits.
What must be true for any of this to work
The consequence must be genuinely agreed by the product organisation. Adopted as a metric without that agreement, the budget is a dashboard nobody acts on in either direction — and the practice is then blamed for not working.