On this page4 sections
In a hypothetical release meeting, the fix is ready, but it has company. The same package contains an unrelated feature, and the service has already used its error budget: the amount of failure allowed by its reliability target. Holding the package keeps the repair from reaching users. Shipping it brings along a change that those users do not need in order to recover.
An error-budget policy should help people separate those choices. It says what work continues, what waits and who can approve a bounded exception as failures consume the allowance. The useful conversation happens before this meeting: can the repair ship separately, and if it cannot, who can weigh the benefit of recovery against the extra feature’s risk?
First, agree which failures used the budget
Before anyone can apply that policy, the people in the meeting need to be looking at the same service result. Start with what customers should receive and how often they receive it. The measurement is a service-level indicator, or SLI. Its target over a specified window is a service-level objective, or SLO. An internal SLO is not automatically a contractual service-level agreement, or SLA.
For an illustrative request-based calculation, a 99.9% success target leaves 0.1% for failed eligible requests. Across one million eligible requests in the window, that is an allowance of 1,000 failures. If 1,200 failures are observed, the service has exceeded its allowance by 200. Everyone can agree on that subtraction and still disagree about whether to release the mixed package. The count describes the service result; it cannot inspect what is inside the next release.
Even the count deserves a closer look when two teams arrive with different numbers. One may include requests that timed out. Another may drop those requests because no final status arrived. They are then measuring different promises, not merely displaying the same promise differently. Agree on the eligible requests, justified exclusions, measurement sources and treatment of missing data before a release depends on the answer.
A genuine measurement defect still needs fixing. Keep the old and new interpretations available when correcting it, so someone can follow why the budget changed. Quietly redefining a bad event during an incident can make the chart look healthier without explaining whether customers are getting a better result. That leaves the next release meeting with a disputed number as well as a disputed package.
Let the balance change the work on the calendar
Once the measurement is understood, “prioritize reliability” needs to turn into something people can actually do. A discretionary release might wait. Engineering time might move to a recurring failure. An expansion might need another review. For each condition, the agreement needs a concrete decision and someone able to carry it out; otherwise the team can acknowledge an exhausted budget and keep exactly the same plans.
The balance is only part of that conversation. Remaining budget tells you how much allowance is left. Burn rate tells you how quickly the service is consuming it relative to the SLO. A service spending its allowance rapidly may need attention before the balance reaches zero. Google’s burn-rate alerting chapter explains this relationship, including why sparse traffic makes the signal harder to interpret. Those signals help decide when to respond; the policy still has to say what changes.
Return to the package on the table. The repair may reduce harm already happening, while the unrelated feature adds another possible source of trouble. Treating both as simply “a release” hides the distinction that matters. The same care applies when a dependency caused the breach. Users still need a response even if the service team did not cause the failure. Reviewing responsibility and choosing what to do for those users are related jobs, but one need not wait for the other to be settled.
Read diagram description
The diagram moves downward through four steps: SLO measurement, with eligible events, target and window; budget consumption, with remaining allowance and burn rate; agreed policy, allowing work to continue, pause or be considered for a scoped exception; and a release decision with a named decision maker, checks and review point. It describes how to reach a decision, not an automatic approval of the hypothetical mixed package.
Give the repair a path of its own
In the hypothetical meeting, separating the package would let the team assess the repair without also accepting the unrelated feature. That changes the decision itself: the expected recovery benefit no longer comes bundled with that additional exposure. It does not make the repair automatically safe or approved, but it gives the appropriate people a narrower change to evaluate.
If separation is not feasible, the tradeoff remains inside the package. Waiting holds back the expected recovery benefit; proceeding accepts the feature’s extra exposure along with the repair. Someone with the necessary authority must decide whether that benefit justifies the combined risk. Calling the package urgent does not tell that person what the feature could change.
Make the exception specific enough to reconstruct later: which package, what customer impact is happening now, what benefit is expected, who approved it and which outcome checks will follow. Bind it to this situation and give it an expiry. That way an exception for this mixed release does not silently become permission for the next one.
The people involved have different parts of the decision. Leadership can accept risk within its authority. Technical owners establish whether the change can be carried out and recovered from. An incident lead coordinates the response, but that role does not automatically supply every production permission or every right to accept risk. An agreed escalation route lets the responder bring the right people into the choice instead of inventing those divisions while trying to restore service.
A dependency failure needs a usable route too. If the service team cannot repair the dependency directly, its response might involve fallback behavior, demand limits or a changed commitment with the provider, depending on the service. Leaving that case unassigned means customer impact can continue while teams debate whose failure consumed the budget. The policy should identify who takes the response forward, even while attribution remains under review.
This is why different services may need different policies. Consistent measurement and explicit exceptions make the decisions understandable. A single release freeze imposed on every workload can overlook both the harm continuing now and the kind of change that might reduce it.
Review the decisions the rule actually produced
After the release decision, look beyond whether the gate followed its instructions. Examine held releases, granted exceptions, repeated exhaustion and customer failures that the SLI missed. A gate can enforce a written rule perfectly while blocking harmless work and admitting the same recurring failure. Those results give the team a reason to revise the agreement, rather than count consistent enforcement as success on its own.
The error-budget policy template helps record that agreement. The release-gate guide carries it into implementation, while the error-budget calculation reference keeps the arithmetic explicit. Together, the measurement and decision record should let another engineer understand why a release proceeded or stopped without needing someone’s memory of the incident.
For the mixed package, the useful result is a repair-only path that can be assessed on its own, or a deliberate exception for the inseparable bundle. Delivery and recovery owners can work out those paths before the budget is exhausted. Then the meeting can address the actual choice in front of it: how to pursue recovery, what additional exposure that requires, and who is authorized to accept it.
Put this into practice
Review the allowance behind your SLO
Bring: Your SLO, measurement window, eligible population and bad outcomes, using one consistent event-based or time-based definition.
Sources & context
Sources linked in this article. Read alongside the author’s analysis; a citation does not independently verify a publisher’s claims.
- burn-rate alerting chaptersre.google
