cumulativeness of classic histogram bucket counts
ClassicHistogramBuckets' javadoc asserts that bucket counts are not cumulative. But, it seems in practice they are cumulative. Indeed, the only way classic histogram buckets can be cleared today is if the native histogram reset interval is configured. Cumulativeness is further consistent with the generic metric type documentation, which says that histogram buckets are cumulative.
Internally, buckets are not cumulative - but that shouldn't be an issue. When scraping, they are cumulative.
It occurred to me that with histograms there's two interpretations of the word "cumulative". There's the notion of cumulative in time. E.g., counters are cumulative in time because they're monotonic. With a histogram though, cumulative could also mean a higher bucket includes all counts in lower buckets. I.e., a cumulative histogram could mean the range of the buckets overlaps such that they bucket intervals are $[-\infty, b_1), [-\infty, b_2), [-\infty, b_3],\ldots$ rather than the disjoint $[-\infty, b_1), [b_1, b_2), [b_2, b_3),\ldots$. Do you think the ClassicHistogramBuckets javadoc is referring to this second notion of cumulativeness?
.. to this second notion of cumulativeness?
that's correct - see https://en.wikipedia.org/wiki/Histogram#Cumulative_histogram and https://prometheus.io/docs/practices/histograms/#apdex-score