Opens in a new tab
Guide

An error-budget template the next engineer can use

In brief

Work through request and time allowances, then fill in the release rules, exception route and follow-up record that make the worksheet usable.

7 min read

Is this the right starting point?
Before you read
Basic percentages and familiarity with the service being discussed; SLO, SLI and error-budget terms are defined.
What you will take away
Complete a draft error-budget worksheet with reproducible event or time arithmetic, missing-data rules, decision owners and scoped exceptions.
Sources
A cream panel with a blue bar and three blue circles points to a blank blue-bordered rectangle, connected by a blue line to a rust-colored tile bearing a cream wrench.
On this page4 sections

Picture a hypothetical release review with a worksheet open between two engineers. One has calculated that 350 failed requests remain in the service's allowance. The other has a repair ready to ship. They agree on the subtraction. They still need to know whether this change can go ahead.

An error-budget template earns its place by keeping those two questions together: how much of the agreed allowance remains, and what the team has agreed to do with that information. The tables below hold the measurement and worked examples; the copyable policy block holds the release rules. Filling in the number is the easy part to recognize as finished.

What are those 350 requests part of?

Before choosing a percentage, name the operation someone needs the service to perform. A request that finishes promptly may be useful; the same request arriving too late may not be. The target for that experience is a service-level objective, or SLO. The share of outcomes it allows to fall short is the error budget.

This record keeps the promise attached to its calculation. It gives the next person enough information to count the same work, rather than merely copy the same percentage.

FieldWhat belongs here
Service and user operationNamed service, environment, and operation covered
Accountable ownerTeam, escalation route, and business stakeholder
Eligible eventsExact population and justified exclusions
Good eventSuccess condition, including any latency threshold
Objective and windowTarget percentage; rolling or calendar period; time zone
MeasurementSource, query, freshness limit, and missing-data behavior
PolicyDecision thresholds, approvers, exceptions, and review date

The measurement itself is a service-level indicator, or SLI. Suppose a good event is an eligible request finishing within 300 milliseconds. A target of 95% leaves an allowance of 5% bad requests. That promise says something about requests, not permission to shut the whole service down for 5% of the month.

“Eligible” deserves a little attention here. It names the population of requests covered by the promise, with exclusions justified rather than convenient. Leave troublesome requests out and the percentage can improve while the service stays exactly as troublesome. Keep that population, the time window and the freshness limit beside the query so another person can reproduce the count.

If the owners disagree about which experience matters, leave that disagreement visible in the record. A more precise percentage will not choose between their promises.

How the allowance becomes a remainder

For the first illustrative row, the service handled one million eligible requests. A 99.9% success target leaves 0.1%, or 1,000 requests, available for bad outcomes. Subtract the 650 observed failures and the worksheet shows 350 remaining. Each step still counts the same thing.

Illustrative objectiveEligible requestsAllowed bad eventsObserved bad eventsRemaining
99.9% request success1,000,000 requests1,000 requests650 bad requests350 requests
95% within 300 ms100,000 requests5,000 requests3,200 slow requests1,800 requests

The second row changes the promise, not the method. Here a request slower than 300 milliseconds is bad. Five percent of 100,000 is 5,000; after 3,200 slow requests, 1,800 remain. The eligible count, allowed count, observed count and remainder belong together, where a reader can follow the subtraction.

In general, multiply eligible events by one minus the target expressed as a fraction, then subtract observed bad events. A negative remainder means the allowance has been exceeded and the objective was not met for that window. There is no need to hide the minus sign to make the sheet look tidier.

A time-based availability promise uses a different ruler. Exactly thirty days contain 43,200 minutes. At 99.9%, the allowance is 43.2 bad minutes; at 99.95%, it is 21.6; at 99.99%, it is 4.32. A different month's length changes those amounts. A count of partially failed requests cannot simply be relabeled as outage minutes: that would change what the service is measuring.

Now suppose collection missed part of the window. The subtraction can still work perfectly on the records that arrived, but those records may not cover the intended requests. Note the missing period and whether it affects the bad-event count, the total count, or both. These are the numerator and denominator of the error fraction. The policy needs an explicit response to missing or stale observations because a neat remainder cannot supply the missing work.

The 350 requests left in this example
Illustrative request calculation: 1,000,000 eligible requests at 99.9% success allow 1,000 bad requests. Subtracting 650 observed bad requests leaves 350. The counts apply to the defined SLO window.
In this illustrative calculation, 1,000 allowed bad requests minus 650 observed bad requests leaves 350. The units remain requests; the result also depends on collection covering the intended population.
Follow the request calculation

The diagram starts with 1,000,000 eligible requests and a 99.9% success target. That allows 1,000 bad requests. After 650 observed bad requests, 350 remain. The drawing keeps those counts within the defined SLO window. The surrounding explanation adds the collection condition: correct subtraction does not establish that all the required observations arrived.

Fill in the release rules while nobody is waiting

Return to the repair beside the worksheet. The remaining allowance informs the decision; it does not grant production permission or the right to accept a new risk. The release policy names the people and conditions that connect the count to action.

The block below has room for normal work, extra review and a hold on specified changes. Its telemetry field concerns the monitoring data used to judge those conditions. “Burn” means how quickly the allowance is being consumed. The fields deliberately leave thresholds and permitted changes blank, because those choices belong to this service.

Policy owner: [team and accountable decision maker]
Applies to: [services and release paths]
Normal state: [conditions and permitted changes]
Review state: [burn/remaining-budget conditions and required review]
Hold state: [conditions and changes paused]
Missing or stale telemetry: [hold, review, or other explicit behavior]
Permitted exceptions: [bounded categories and approver]
Resumption: [health evidence, observation window, and approval]
Policy review date: [date and triggers for earlier review]

Choose those thresholds from the workload and incident evidence. Google's SLO alerting guidance explains how several observation windows can help distinguish patterns of budget consumption, and why low request volumes need particular care. Its examples help with the reasoning; another service's numbers are not a completed policy for yours.

For a second hypothetical worksheet, move the service beyond the first example: the allowance is now exhausted, and a security repair reaches the release hold. Its calculation is correct, but the exception-approver field is empty. The repair waits while people work out who may authorize it. The sheet has stopped the release as designed and then left the people without a way to resolve it.

Walking through that case before enforcement exposes a missing agreement that more arithmetic would never find. Name which work pauses, how the repair reaches an authorized approver and what health observations, observation period and approval allow normal releases to resume. Coordinating the discussion is not automatically permission to approve every consequence.

Then make the normal approver unavailable in the rehearsal. Can the team reach a named fallback and show why the budget was exhausted? Remove fresh monitoring data next and follow the written rule for that condition. A small, familiar case may need only a concise entry; extra structure should earn its place by resolving uncertainty or coordination the team actually faces.

Leave the next reader the decision, too

An approved exception will outlive the conversation that produced it. Keep its change identifier, reason, scope, supporting evidence, expected benefit, approver, expiry and later outcome together. Someone returning to the repair should be able to see what was permitted and what happened, without reconstructing the meeting.

If follow-up remains, name who will do it and what observation would establish improvement. Permission to release and evidence of improvement are different entries. An unsupported improvement percentage cannot fill the second one.

Before the worksheet controls releases, have another person reproduce its arithmetic, try the no-traffic case and walk through the repair that needs an exception. Check the contact route as well as the name. The release-gate guide takes the next step into implementation once this agreement can actually be followed.

The useful handoff is now more than a cell containing 350. It includes which requests were counted, what is still unknown, and where a proposed change goes for a decision. When the security repair reaches an exhausted budget, the engineer can turn to the agreed exception route instead of beginning a negotiation with an empty field.

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.

Report an error or outdated detail

Keep reading

Where to go next