takes a large and continuous data stream and continuously only retain a reduced empirical CDF approximation
small, simple, no dependencies
// recorder with 5 retained samplesvar recorder = 55362718console // minimum:0console // median:4console // maximum:8
- constant memory use, no compression steps and/or triggered garbage collection
- significantly faster than other implementations (about 3-5x faster)
- mean-preserving compression (empirical cdf with same average as samples)
- No value interpolation. Values are kept as-is, but ranks are interpolated.
- other than the mean, the other moments (variance, skew, kurtosis) are not preserved
- The moments are those of the generated curve that approximate those of the source samples
Background and related projects
- tdigest based on the work of Dunning
- h-digest based on the above but keeping values and adjusting ranks only
This module attempts to match the underlying CDF as closely as possible by adjusting local ranks to keep the local average when discarding values.
By some measure, the root mean square error for each original sample value is 10 times better than the 2 other implementations above.
npm run compare or see
./util/compare.js for a benchmark and error comparison for multiple distribution types.
var recorder = new Recorder(L): creates a recorder that will keep
Lvalues and associated
Properties - read-only getters
.Nnumber: total samples received
.Enumber: average of samples received and of the resulting approximated cdf
.Snumber: standard variation of the resulting aproximated cdf
.Vnumber: variance of the resulting aproximated cdf
Properties - internal
.vsarray: internal store of retained sample values
.rsarray: internal store of retained value approximated ranks
.push(number)void: sample value(s) to be added
.F(value:number)number: cummulative probability (cdf) of the specified value
.f(value:number)number: probability density (pdf) of the specified value
.Q(probability:number)number: estimated value for specified probability
.M(order:number)number: estimated origine moment (ie. E = M(1), V=M(2)-E^2)