Metrics data model

The data model of metrics is an extension of emit's events. Metric events are points or buckets in a time-series. They don't model the underlying instruments collecting metrics like counters or gauges. They instead model the aggregation of readings from those instruments over their lifetime. Metric events include the following well-known properties:

• evt_kind: with a value of "metric" to indicate that the event is a metric sample.
• metric_agg: the aggregation over the underlying data stream that produced the sample.
  • "count": a monotonic sum of 1's for defined values, and 0's for undefined values.
  • "sum": a potentially non-monotonic sum of defined values.
  • "min": the lowest ordered value.
  • "max": the largest ordered value.
  • "last": the most recent value.
• metric_name: the name of the underlying data stream.
• metric_value: the sample itself. These values are expected to be numeric.
• metric_unit: the unit of the metric sample.
• metric_description: an end-user description of the data stream.

The type of metric_value

The type of metric_value isn't constrained in any way, but will enjoy the broadest compatibility when it's numeric. Integral counts are best represented as usize, and derived statistics and other fractional values as f64. Keep in mind the potential for rounding errors and precision loss particularly when processing f64 samples.

Distributions

Metric samples can optionally carry additional properties that describe the distribution of the data that produced them. The well-known properties for distributions are:

• dist_count: the count of values, like metric_value when metric_agg is "count".
• dist_sum: the sum of values, like metric_value when metric_agg is "sum".
• dist_min: the minimum value, like metric_value when metric_agg is "min".
• dist_count: the maximum value, like metric_value when metric_agg is "max".
• Exponential histograms as described in Distributions, using both of the following properties:
  • dist_exp_scale: the scale of the histogram. The value is expected to be an integer.
  • dist_exp_buckets: the bucket midpoints and counts as a [(f64, u64)] sequence.