Error budgets for small teams: a practical guide
An error budget turns an uptime target into minutes you can spend. How to calculate it, decide what counts, and use burn-rate alerts without a dedicated SRE team.
· 7 min read · By the Spot Downtime team
An uptime target on its own is just a number on a slide. An error budget makes it useful: it is the amount of failure you are allowed before you miss the target, expressed in minutes or requests. Once you have one, decisions like "should we ship this risky change today?" and "is this blip worth waking someone up?" get much easier. You don't need an SRE team to use the idea. A spreadsheet and a monitor are enough.
The arithmetic
The budget is the complement of your target over a window. For a 30-day window (43,200 minutes):
| Target | Budget (30 days) | Budget (per year) |
|---|---|---|
| 99% | 7 h 12 min | 3 d 15 h 36 min |
| 99.5% | 3 h 36 min | 1 d 19 h 48 min |
| 99.9% | 43 min 12 s | 8 h 45 min 36 s |
| 99.95% | 21 min 36 s | 4 h 22 min 48 s |
| 99.99% | 4 min 19 s | 52 min 34 s |
The formula is window x (1 - target). For 99.9% over 30 days: 43,200 x 0.001 = 43.2 minutes. The uptime calculator does this for any target. For what these numbers mean in practice, see what 99.9% uptime really means.
Step 1: pick a target you can defend
Choose a target lower than you wish for. A team of three that deploys several times a day and runs on a single database will not hit 99.99%, and pretending otherwise just makes the budget meaningless. A target is good when:
- It is a little looser than what you actually achieved over the last few months.
- Users would genuinely notice and complain if you fell below it.
- It is no stricter than any promise in a customer contract. Keep the internal target tighter than the external SLA, so you have a warning zone before credits apply (see SLA credits explained).
Step 2: define what counts as downtime
This is where most budgets quietly fail. Write down the rule, so two people looking at the same incident reach the same number.
- What is "down"? A non-2xx response, a response over a latency threshold, a failed login, or a missing field in an API response all count in some definitions. Pick the checks that match user pain. A homepage returning 200 while checkout is broken is down.
- Is planned maintenance excluded? Customer contracts often exclude it; an internal budget usually shouldn't, because users still can't use the product. Decide and write it down.
- What about partial outages? Either count the full duration, or count the fraction of failing checks or requests. Counting the full duration is simpler and stricter.
- Whose failures count? A third-party outage that breaks your product still costs your users. Count it, then decide separately whether to build around it.
Measure from outside, with independent monitoring from several locations. Your own metrics miss exactly the failures that matter most: DNS, certificates, the load balancer, the CDN.
Step 3: track consumption
Keep one row per incident: start, end, duration, what was affected. Sum the durations over the rolling window and subtract from the budget.
Target: 99.9% over 30 days -> budget 43.2 min
Nov 02 deploy broke login 12 min
Nov 06 database failover 9 min
Nov 09 expired certificate on api 15 min
------
Consumed 36 min (83%)
Remaining 7.2 minA monitor's incident history provides the start and end times without anyone remembering to write them down, and keeps the arithmetic honest. Spot Downtime keeps an incident history for exactly this.
Step 4: decide what happens at each level
A budget only changes behaviour if it has consequences agreed in advance. A simple policy for a small team:
- Under 50% consumed: ship normally. You have room to take risks and run experiments.
- 50-100% consumed: slow down on risky changes. Prefer small deploys, deploy earlier in the day, and finish pending reliability fixes first.
- Budget spent: freeze feature releases that carry risk until the window recovers, and spend the time on the causes of the last incidents.
The point is not to punish failure. A budget you never spend means you are too cautious; one you always blow means the target or the system needs to change.
Alerting on the rate you burn it
Paging on every failed check wakes people for blips that cost almost nothing. Paging only when the budget is gone is too late. The useful signal is the burn rate: how fast you are consuming the budget compared with an even spend over the window. A burn rate of 1 uses exactly the whole budget by the end of the window; a rate of 10 uses it in a tenth of the time.
The Google SRE Workbook suggests pairing a fast and a slow alert. With a 30-day window:
- Page on a burn rate of 14.4 sustained for 1 hour. That consumes 2% of the monthly budget in that hour (14.4 x 1 / 720).
- Page on a burn rate of 6 sustained for 6 hours: 5% of the budget (6 x 6 / 720).
- Open a ticket, don't page, on a burn rate of 1 over 3 days: slow leaks that need fixing but not tonight.
Most small teams can't compute burn rate automatically and don't need to. A rougher version works: page for a hard outage that is ongoing, and review the budget weekly to catch the slow leaks. To keep outages from being triggered by blips, require a failure to be confirmed before alerting. Spot Downtime confirms every failure from three regions before it alerts; see how to reduce false uptime alerts and alert escalation policies for routing.
A worked example
At 99.9%, a 12-minute outage uses 28% of the 30-day budget. Two of those in a month leaves you with about 19 minutes. A single 50-minute outage exceeds the budget by itself, which is why one long incident hurts more than many short ones, and why fast detection and fast rollback matter more than preventing every failure.
Detection time is part of the outage
Common mistakes
- Setting the target from a marketing page instead of from measured history.
- Measuring with internal health endpoints that stay green during a real outage.
- Never writing down the definition of downtime, so every incident starts an argument.
- Having a budget with no policy attached, so nothing changes when it runs out.
- Chasing extra nines. Each additional nine costs roughly ten times more effort and rarely matters to your users as much as you think.
Get started
Set up website monitoring on the pages and API calls your users depend on, pick a target, and start the incident log. After one window you will know your real numbers, and the first review will tell you more than any amount of planning.
Downtime cost calculatorEstimate what an hour of downtime costs you, to help decide how tight the target should be.Get new guides by email
Practical posts on monitoring, outages, error fixes and on-call, when they're published. No spam.
Keep reading
- Uptime · SLAsSLA credits explained: what you get when a provider misses its uptime targetHow SLA service credits work, how providers measure downtime, why you must claim them, and how to keep your own evidence so the claim isn't denied.November 8, 2026 · 7 min read
- Uptime · SLAs99.99% uptime: how much downtime is that?Four nines allows 52 minutes of downtime a year, or 4 minutes 23 seconds a month. What that budget means for deploys, on-call and monitoring, and when 99.99% is worth promising.October 27, 2026 · 5 min read