Bootstrap
The bootstrap approximates a statistic's sampling distribution by recalculating it on samples drawn from an estimated population distribution.
In the ordinary nonparametric form, draw the original number of observations with replacement from the observed sample and recompute the statistic. Observations may repeat or be absent. A bootstrap distribution, its standard deviation and a confidence interval are different outputs. Percentile, basic and bias-corrected accelerated intervals need not agree. [1]
Resampling must reflect the sampling unit and dependence. Repeated daily records from one drive are not independent devices. A whole-cluster procedure retains members of a sampled group together. A paired bootstrap retains aligned measurements across arrays; that setting alone does not identify rack or device groups. [2] [1]
More Monte Carlo resamples can reduce simulation error but do not increase the observed fleet size. If all observed binary outcomes are zero, ordinary resampling produces only zeros, not evidence that the population failure probability is zero. Degenerate distributions can also make interval calculations undefined. [1]