OCaml implementation of the T-Digest algorithm.
The T-Digest is a data structure and algorithm for constructing an approximate distribution for a collection of real numbers presented as a stream.
The median of a list of medians is not necessarily equal to the median of the whole dataset. The median (p50), p95, and p99 are critical measures that are expensive to compute due to their requirement of having the entire sorted dataset present in one place.
The T-Digest can estimate percentiles or quantiles extremely accurately even at the tails, while using a fraction of the space.
Additionally, the T-Digest is concatenable, making it a good fit for distributed systems. The internal state of a T-Digest can be exported as a binary string, and the concatenation of any number of those strings can then be imported to form a new T-Digest.
Tdigest module return a new
Tdigest.t, including "reading" ones since they may trigger expensive intermediate computations worth caching.
The API is well documented here.
opam install tdigest
On a 2018 MacBook Pro, it can incorporate 1,000,000 random floating points in just 770ms.
Exporting and importing state (
of_string) is essentially free.
>= "v0.14.0" & < "v0.15.0"