Mean time between failures (MTBF)
Mean time between failures (MTBF) tells you how long a system usually runs before it breaks again. You take the time it was up in a period and divide by the number of failures in that period. It is a number for things you repair and put back into service, like an API, a server or a database cluster.
How it works
MTBF sits next to two other averages. Mean time to recovery (MTTR) is how long each failure lasts until service is back. Mean time to failure (MTTF) is for things you replace instead of repair, like a disk or a bulb. Some teams measure MTBF from the start of one failure to the start of the next, which adds the repair time. Say which one your dashboard uses.
Together they give availability: MTBF / (MTBF + MTTR). Take a payments API over 30 days, or 720 hours. It failed 4 times, and the outages added up to 2 hours. So MTBF is 718 / 4 = 179.5 hours, MTTR is 30 minutes, and availability comes out at 99.72%.
Its service level objective is 99.9%, which allows 43 minutes of downtime in 30 days. The team spent 120. Raising MTBF alone means at most one 30-minute outage a month. Cutting MTTR to 10 minutes gets there with the same 4 failures, at 40 minutes.
Hardware MTBF is a fleet number. A disk rated at 2 million hours does not last 228 years. Run 1,000 of them and you should expect one to fail roughly every 12 weeks.
Why it matters in production
That disk math assumes failures are random and independent. Software failures are neither. They bunch up after deploys, config changes and traffic peaks, and one bad release can take down every replica at once. So an MTBF of 179.5 hours can hide that 3 of the 4 outages followed Friday deploys. Count failures per change as well.
In One2N's post on evaluating AI SRE agents, Hemant Kumar describes MTBF as the preventative work after the page. That work comes out of incident management reviews, and it pays off in high availability only when the fixes ship.
See also
On the One2N blog
References
