B10 life
B10 life is the tenth percentile of a specified population's time to a defined failure event.
Specify population, failure event, time origin, units and conditions. For a continuous, strictly increasing lifetime CDF F, F(B10) = 0.10 and S(B10) = 0.90. More generally, B10 is the earliest time F reaches or exceeds 0.10; a discrete jump need not give exactly 10% failed. This is not an individual device's failure date. [1] [2]
For an unshifted two-parameter Weibull model with scale eta > 0 and shape k > 0, B10 = eta * (-ln(0.9))^(1/k). Eta has time units; k is dimensionless. Eta is the characteristic life at about 63.2% cumulative failures, not B10 and not generally the mean. State the parameter convention; the formula does not establish a Weibull fit to drives. [3]
The 90% survival fraction is not a 90% confidence level. An estimated B10 may separately have a confidence interval or one-sided bound. Report its method and level: the percentile identifies the population quantity, confidence concerns its estimation. [4]
Retain right-censored observations rather than taking the tenth percentile of failures alone. In fixed-time censored likelihood, failed units contribute density terms and survivors contribute survival terms. Report failures, sample size and censoring assumptions, including non-informative censoring. A fitted B10 need not be an observed quantile. [5]
Cluster replacement gaps are not device lifetimes. Schroeder and Gibson's FAST 2007 section 5.3 fits Weibull to gaps between HPC1 replacements; those parameters cannot supply individual disk-lifetime B10. [6]
Relations
Citations
- NIST e-Handbook 1.4.4: B10 as the estimated 10% point of failure times
- R stats documentation: sample quantiles and inverse empirical CDF
- NIST e-Handbook 8.1.6.2: Weibull formulas and parameter conventions
- NIST e-Handbook 1.3.5.2: confidence-interval interpretation
- NIST e-Handbook 8.4.1.2: censored likelihood and estimation limits
- Schroeder and Gibson: Disk failures in the real world, FAST 2007, sections 2.1 and 5.3