SLA Risk Analyzer Uptime Targets

Availability is the fraction of a measurement window in which a service met its success criteria, and an uptime target converts directly into a downtime allowance once the window length is fixed. Using an average month of 43,833.6 minutes, 99 percent permits 438.3 minutes of outage, 99.9 percent permits 43.8 minutes, 99.95 percent permits 21.9 minutes, 99.99 percent permits 4.4 minutes and 99.999 percent permits about 26 seconds. Annually those become roughly 3.65 days, 8.8 hours, 4.4 hours, 52.6 minutes and 5.3 minutes. That allowance is the error budget: the quantity of unreliability a team is permitted to spend on releases, migrations and experiments before it must stop shipping and work on stability instead. Availability can also be written as mean time between failures divided by the sum of mean time between failures and mean time to recovery, which shows that halving recovery time buys as much availability as halving failure frequency. Two structural facts govern realistic targets. Dependencies in series multiply, so a service built on five components each at 99.9 percent has a theoretical ceiling near 99.5 percent unless it degrades gracefully; redundancy works the other way, taking the failure probability to the power of the number of independent replicas, which is why single-region architectures rarely sustain four nines. The financial layer is separate from the engineering one. SLA credits are almost always service credits rather than damages, are commonly 10 to 25 percent of the monthly fee, and are usually capped at 100 percent of that fee, with carve-outs for scheduled maintenance, force majeure and customer-caused faults. Two important caveats: computed availability depends heavily on the measurement granularity and on whether failure is counted by minutes or by failed requests, and real failures are correlated, so the independence assumed by redundancy math overstates protection.

Quantified availability targets came from telephony and defense electronics. Bell System engineers designing electronic switching systems in the 1960s and 1970s worked to a goal of no more than a couple of hours of outage over a forty year service life, the origin of the five nines shorthand that later spread to computing, while formal reliability prediction was codified for military hardware in handbooks such as MIL-HDBK-217, first issued in the 1960s. Data centers received their own graded standard when Kenneth Brill founded the Uptime Institute in 1993 and published the Tier classification, which ranks facilities from Tier I through Tier IV by redundancy of power and cooling paths and attaches expected annual availability to each level. The contractual form followed the outsourcing of infrastructure. Telecommunications carriers offered service level agreements with credits through the 1990s, and public cloud made them universal: Amazon Web Services published an SLA for EC2 in October 2008 promising 99.95 percent annual availability backed by service credits, and competitors matched the structure. The engineering counterweight came from Google, where the Site Reliability Engineering discipline established by Ben Treynor Sloss in 2003 reframed the leftover unreliability as an error budget to be spent deliberately, a practice described publicly in the 2016 O'Reilly book Site Reliability Engineering edited by Betsy Beyer and colleagues. That book also popularized the distinction still used today between service level indicators, the measurements, service level objectives, the internal targets, and service level agreements, the external promises with money attached.