Exponential histogram (DGIM)
A DGIM exponential histogram approximates recent-stream counts using timestamped buckets whose sizes grow by powers of two.
The original counting problem concerns 1-bits among the latest N stream elements. Older equal-size buckets merge, retaining their more recent timestamp. Only the oldest surviving bucket may straddle the window boundary; its partial contribution is estimated. The paper derives a relative-count-error bound with O((1/epsilon) log^2 N) bits of memory. This is a deterministic approximation bound, not a sampling confidence interval. [1]
Here, bucket size means a count of 1-bits, not a range of measured latencies. The similarly named OpenTelemetry representation partitions measurement magnitudes instead. Choosing either representation does not by itself specify wall-clock windows, handling of late events or the population to count.
For storage telemetry, first decide whether the question is how many events occurred recently or how measured durations were distributed. A stream position and a duration value are different coordinates.