A lock-free latency histogram: an :atomics array of log-spaced buckets that many processes write to concurrently, with no per-op allocation and no shared GenServer. The load generator records request latency in it, and Malachi.Metrics records storage flush latency in it.

Latencies are recorded in microseconds. Bucket b covers (2^((b-1)/scale), 2^(b/scale)] us, so the resolution is 2^(1/scale) - 1 (about 4.4% at scale = 16), fine enough for tail percentiles while keeping the array tiny. Percentiles return the representative us of the bucket the rank falls in.

Writers pay two :atomics.add/3 calls on the caller (the bucket and the running sum), which is why this is safe on a hot path: routing the samples through a GenServer would serialize every writer behind one process instead.

cumulative/1 exports the counts at a coarser, fixed set of edges (edges/0, four per octave), which is what a Prometheus histogram needs: a stable le set whose counts can be subtracted between two scrapes. Every edge is the upper bound of an internal bucket, so those counts are exact rather than interpolated. A bucket includes its upper bound, so a count is of samples at or below its edge, which is what a Prometheus le means; the only integer samples that can equal an edge are exact powers of two.