Typical/default tolerance values for "known" quantiles

[daluu]

I came across this from https://github.com/tsenart/vegeta, which at the moment follows the default tolerance values and known quantiles presented in the example docs https://godoc.org/github.com/streadway/quantile#example-Estimator.

Curious to how the typical/default tolerance values were determined for the known quantiles:

• 50th percentile or median
• 95th percentile
• 99th percentile

and how to extrapolate that for other typical additional percentiles or known quantiles, for example 90th percentile is often another quantile or percentile of interest. This method of figuring out tolerance values would be helpful to figure out for given quantiles, X,Y,Z, how to figure out the tolerances for new quantiles A,B,C. (e.g. other percentiles like 25th, 75th, 80th, etc.).

In the case of 90th percentile (for vegeta), based on this project's comment // We know we want to query the 95th and 99th, with the 95th a little less accurately. and current known quantile tolerances (pattern), I came up with this remapping to fit in the 90th percentile, based on gut instinct. Is it reasonable?

quantile.Known(0.50, 0.01),
quantile.Known(0.90, 0.001),
quantile.Known(0.95, 0.0005),
quantile.Known(0.99, 0.0001),