NB: Post author here.

Yeah, that was one of the reasons we chose it as one of the ones to implement, seemed like that was a really interesting tradeoff, we also used uddsketch[1] which provides relative error guarantees, which is pretty nifty. We thought they provided different enough tradeoffs that we wanted to implement both.

[1]: https://arxiv.org/abs/1908.10693

Is it using https://github.com/tvondra/tdigest under the hood, or a separate implementation?

If folks are interested:

https://github.com/timescale/timescaledb-toolkit/blob/main/e...

(The TimescaleDB Toolkit is also implemented in Rust)

Facebook seems to have an even better performance implementation using sqrt. Might make sense to port that over to Rust. https://github.com/facebook/folly/blob/master/folly/stats/TD...

a separate implementation

Hi, in an unrelated nitpick: The relative error should be calculated by dividing the error by the true value, not by it's approximation. Still, very nice writeup!

Thanks! messed up the formula but had it right in the text :( Fixed now.

Not an expert on this topic but I noticed that the KLL algorithm (published in 2016) was not mentioned in this thread, which provides theoretically optimal performance for streaming quantiles with guaranteed worst case performance: http://courses.csail.mit.edu/6.854/20/sample-projects/B/stre... (And is pretty fast in practice).

NB: Post author here.

Interesting will have to take a look! Thanks for sharing!

Also relevant: Single-Pass Online Statistics Algorithms

[1] http://www.numericalexpert.com/articles/single_pass_stat/

That's pretty neat! Can these be used to efficiently compute rolling percentiles (over windows of the data), or just incremental?

The UDDSketch (default) implementation will allow rolling percentiles, though we still need a bit of work on our end to support it. There isn't a way to do this with TDigest however.

Sure there is. You simply maintain N phases of digests, and every T time you evict a phase and recompute the summary (because T-digests are easily merged).

I think this would be a tumbling window rather than a true "rolling" tdigest. I suppose you could decrement the buckets, but it gets a little weird as splits can't really be unsplit. The tumbling window one would probably work, though Tdigest is a little weird on merge etc as it's not completely deterministic with respect to ordering and merging (Uddsketch is) so it's likely you get something that is more than good enough, but wouldn't be the same as if you just calculated it directly so it gets a little confusing and difficult.

(NB: Post author here).

This is what I do, it's not a true rolling digest but it works well enough for my purposes.

yep I had to implement t-digest in a monitoring library. another alternative (although older) that the prometheus libraries use is CKMS quantiles [0].

[0] http://dimacs.rutgers.edu/~graham/pubs/papers/bquant-icde.pd...

i think the new ones started wtih Greenwald-Khanna. but i definately agree - p^2 can be a little silly and misleading. in particular it is really poor at finding those little modes on the tail that correspond to interesting system behaviours.

That sounds familiar, I remember reading about Greenwald-Khanna before I found T-Digest after I ran into the "how to find a percentile of a massive data set" problem myself.