Datary keeps ordering and distribution operations separate. Observation order is used for change metrics; a sorted view is used only for median and percentile calculations.
For finite values (x_1,\ldots,x_n):
count: all records considered;valid_count: finite numeric values;missing_count: records without a finite numeric value;minimum,maximum,sum, and arithmeticmean;medianand linearly interpolated percentiles at 5, 25, 50, 75, 95, and 99%;- sample
varianceand samplestandard_deviation(zero for one value); root_mean_square: (\sqrt{\sum x_i^2/n});rate_of_change: (x_n-x_1), the net ordered change per record series;mean_absolute_difference: (\sum_{i=2}^{n}|x_i-x_{i-1}|/(n-1)).
Missing and non-numeric values are omitted from numeric order. Integer values beyond ±2^53 are
preserved as records but excluded from binary64 analysis and receive an invalid-values finding;
this avoids silently rounding identifiers or counters. Their record positions remain available to
quality analysis.
Time accepts finite numeric seconds or timezone-aware ISO 8601 strings. Naive date-times are not treated as time because their offset is ambiguous.
For positive adjacent intervals:
- mean, median, minimum, and maximum interval;
- population standard deviation as
jitter; - effective sample rate (1/\text{mean interval});
- a gap count for intervals at least twice the median;
- global duplicate-timestamp count, including non-adjacent duplicates;
- backwards-timestamp count.
The start, end, and duration are epoch-second values when ISO 8601 input is used.
Required roles are time, target, and response. Supply them with --time-field,
--target-field, and --response-field.
- Rise time: elapsed time between the first recorded 10% and 90% response crossings.
- Peak: maximum response for an upward step or minimum response for a downward step.
- Percentage overshoot: peak excursion beyond the final target divided by absolute step span.
- Settling time: first recorded point after the last excursion outside a ±2% band. The band uses the larger of absolute step span and target magnitude.
- Steady-state error: final target minus final response.
- MAE and RMSE: ordinary sample error summaries.
- IAE and ISE: trapezoidal integration over non-decreasing time intervals.
Rise, overshoot, and settling are reported only when the recorded target is stable. Crossings are not interpolated. Warnings disclose changing targets and excluded backwards intervals.
Required roles are sequence and latency; bytes and time are optional.
- Packet-loss estimate: missing integer IDs inside the observed inclusive sequence range divided by the expected range.
- Duplicate-packet rate: repeated valid sequence IDs divided by valid IDs.
- Out-of-order count: descending adjacent valid IDs.
- Mean, median, percentile latency, and latency jitter. Jitter is the mean absolute difference between consecutive valid latency observations.
- Throughput: total non-negative byte count divided by positive observed duration.
These definitions assume contiguous integer sequence identifiers and consistent latency/time units. They are measurement summaries, not transport-protocol truth.