Error budgets are useless without a policy. 'We're out of error budget' should trigger consequences. If it doesn't, you don't have an error budget — you have a vanity metric.
Here's a policy that actually works.
The four states
Healthy (< 70% of budget used). Business as usual. Feature development proceeds at full speed.
Watch (70-90% used). Feature velocity continues but new risky changes require explicit sign-off from an SRE. No gate, just attention.
Constrained (90-100% used). Feature freezes. Only reliability work and critical bug fixes until we're back below 90%.
Breached (> 100% used). Incident-level response. Leadership informed. Post-mortem for why we blew through. Feature work stays frozen until we recover and identify systemic causes.
The part most policies miss
The feature freeze in 'constrained' state is the part that actually changes behavior. Everything else is documentation. Without consequences, teams ignore the budget.
The freeze has to be real. Leadership can't override it for a 'really important feature' — that's exactly the time the freeze matters. The only exception is a legitimate emergency fix, and those should be rare.
Selling this to leadership
Executives hate feature freezes. They see it as slowing the business. Counter-argument: feature freezes during budget exhaustion protect the business. Shipping features onto broken infrastructure creates more breakage, which burns more budget, which is a doom loop.
Frame it as: 'the feature freeze is a safety valve. When it triggers, it's because something's wrong and we need to fix it before making it worse.'
Also: a good policy lets you spend the budget aggressively when you have it. Feature teams should be encouraged to experiment, deploy fast, and take risks when you're at 30% budget used. The freeze is only for when the safety margin is gone.
The review cadence
Weekly error budget review, 15 minutes max. Who attended: SRE lead, engineering manager, maybe a PM. Decisions: are we in healthy/watch/constrained? Any actions for the coming week?
Monthly broader review with leadership. Trends over time. Investment decisions.
The escalation
If a team enters 'constrained' state three times in a quarter, that's a systemic issue. Escalate to engineering leadership with a proposal: either invest in reliability or accept a lower SLO formally.
The endgame
A mature organization uses error budget policy to balance feature velocity against reliability automatically. Nobody is negotiating individual decisions. The framework does the work.
Getting there takes 6-12 months of discipline. The first few freezes will feel painful. After that, they become routine, and something surprising happens: you stop having them as often. The policy is working.
Written by Dr. Samson Tanimawo
BSc · MSc · MBA · PhD
Founder & CEO, Nova AI Ops. https://novaaiops.com
Top comments (0)