Anyone who has compared entry categories for a trip abroad knows the pattern: a tourist route, a business route and a transit route can all look similar on the surface, yet each has different documents, timing and cost, and choosing the wrong one creates headaches. Technology teams face a parallel decision when they pick an availability target for an online service. Each extra “nine” belongs to a different category, with its own price and its own fit. This comparison sets four common targets side by side, following the framework of the SRE entry in the Ops Atlas. It is general information about engineering practice, not travel or immigration guidance.
The yardstick: what a target really measures
Before comparing categories, you need a common unit. In the discipline known as site reliability engineering, a target applies to an indicator, defined as good events divided by valid events and measured as near the user as possible. Over a window, usually a rolling 28 days according to the Atlas, the target converts into an error budget, which is simply one minus the target. The budget is expressed in bad minutes or failed requests. With that yardstick, we can compare tiers fairly.
Tier A: 99 percent, the entry category
Over 28 days, a 99 percent target tolerates about 6.7 hours of bad service, a figure derived by applying the same method. It suits internal tools, batch jobs and early-stage products where users can wait. The cost is low and the requirements are light: basic monitoring, a simple rota and a modest recovery routine.
- Suited to: internal dashboards, back-office tasks, prototypes.
- Documents needed: a measured indicator and a named owner.
- Risk: users of customer-facing pages will notice the downtime.
Tier B: 99.9 percent, the mainstream category
This is the working example used throughout the Atlas. Over 28 days it permits 40.3 bad minutes, or a thousand failed requests per million. It suits most customer-facing services, and it is where the full practice starts to pay off: burn-rate alerting, a signed budget policy and regular reviews. The sheet’s illustrations include a service with 50 million requests and a budget of 50,000 failures.
- Suited to: shops, portals, content sites, ordinary business apps.
- Documents needed: indicator, objective, written policy with a signature, tested alerts.
- Risk: if the budget stays nearly full every month, the target may be loose or releases too cautious.
Tier C: 99.99 percent, the premium category
Four nines leaves around four minutes in 28 days. That is shorter than most human response times, so a person paged at 3 a.m. cannot meaningfully protect the budget. Teams need automated failover, staged rollouts and well-tested capacity margins. The Atlas notes a rule of thumb that dependencies should offer one extra nine beyond the service they support, which means a four-nines product needs five-nines parts underneath. That demand is where costs jump.
- Suited to: payments, authentication, core infrastructure other teams depend on.
- Documents needed: everything from Tier B plus automation, load testing and redundancy planning.
- Risk: the entry cautions that every further step up can run to about a hundred times the price of the one before.
Tier D: 99.999 percent, the exceptional category
The Atlas notes that at five nines a complete outage can use up the whole budget in about 26 seconds. Nobody can respond to that manually, which turns the target into an architecture question rather than an operations question. The source lists “five nines everywhere” as an anti-pattern, because few organisations can justify the expense for more than a tiny handful of services.
- Suited to: a very small number of safety-critical or foundational systems.
- Documents needed: all of the above plus engineered automatic recovery and an honest cost case.
- Risk: spending money users cannot perceive.
Side-by-side summary
- Cost and effort: rises slowly from A to B, then steeply from C onward.
- Human involvement: people can defend A and B; C and D need machines to act first.
- Alert style: A and B can use simple pages and tickets; C and D need rapid burn detection, and at a loose 90 percent target some alerts can never fire at all, according to the Atlas.
- Typical misuse: choosing D for everything, or choosing A for something customers pay for.
A worked comparison of two services
Imagine two services in one company. The first is a staff-only expense report page used a few times a week. The second is the login screen customers use to reach their accounts. Placing both at Tier B would waste effort on the first and under-protect the second, while placing both at Tier D would be ruinous. A sounder arrangement gives the report page Tier A, with a reminder that people can simply retry later, and gives the login screen Tier C, backed by automation and staged releases. The difference in annual spend between those plans is usually far larger than the difference in the total number of nines, which is why the choice deserves a written justification.
The same exercise reveals dependencies. If the login screen relies on a directory service that only promises three nines, the stricter target cannot be met however carefully the front end is built. Mapping those relationships before choosing a tier prevents promises the underlying parts cannot keep.
How to choose the right category
Start from the user, not from prestige. Ask what customers actually perceive. If a service is used rarely and tolerates retries, a lower tier is fine. If it blocks a purchase or a sign-in, step up. Check whether the organisation will sign a policy; the Atlas treats unsigned targets as decorative. Check the on-call arithmetic, too: it recommends around eight people for continuous single-site cover, and a ceiling of roughly two incidents per twelve-hour shift.
Choosing is a matter of purpose, evidence and capacity, much as choosing between entry categories depends on the purpose of travel. If you want to see how that comparison logic reads in the immigration field, the guides to a Spain short-stay tourism visa and how to apply and to a Spain short-stay business visa with its requirements and fees show how similar-looking categories are separated by purpose and paperwork. Always check current rules with the official authority before relying on them.
A note on movement between tiers
Targets are not permanent. A product may begin at Tier A, rise to B as customers arrive, and later raise one component to C. Moving upward is a deliberate investment, and moving downward can be a sensible saving if the evidence shows nobody notices. Review each target at least once a year, and use the error budget history to support the change.
The same logic applies to the paperwork analogy: circumstances shift, and the right category shifts with them. What matters is that the choice is made on evidence, with the purpose clear and the costs understood in advance.