Alexander Saydakov

> a number of inaccuracies that could have been repaired with a bit of dialog This issue was opened (more than three years ago!) precisely to have that bit of...

> Assuming that somebody in the world would see a random JIRA on a project that they don't know about isn't much of a way to encourage dialog. I am...

> Next, when accuracy is examined, it appears that only the median is considered This is not so. We consider all ranks. The testing method is as follows: 1. for...

> benchmarks that show about 50 ns amortized insertion times At what input size? I am not sure if I understand your point here. Do you mean that you disagree...

> Since the error is largest at the median for the t-digest, this is equivalent to considering only the median. I think I see your objection. Perhaps we ought to...

I don't see any basis for this statement that our benchmarking is flawed. We carefully construct our measurements. We can see fine details of behavior of particular algorithms. For instance,...

Here is a fresh measurement of update time of t-digest with compression=200 Uniform random input (Random.nextFloat)

Much closer than what? This is quite consistent with that report on our web site 3 years ago. It was a bit higher - around 120 ns, but my laptop...

And your code must have changed too. I don't see that problem with rank we have seen 3 years ago. And also, if you look closer at these update time...

@tdunning is this change of the transition point from 1000 to 2000 expected?