(I made the number 10,000 up, but you could do some statistics to figure out how many samples would be needed for a given level of confidence, and I don't think it would be prohibitively large.)
Do you have a source for that claim?
I don't see how could that possibly be true... For example, if your original points are sampled from two gaussians of centers -100 and 100, of small but slightly different variance, then the true median can be anywhere between the two centers, and you may need a humungous number of samples to get anywhere close to it.
True, in that case any point between say -90 and 90 would be equally good as a median in most applications. But this does not mean that the median can be found accurately by your method.
https://en.m.wikipedia.org/wiki/Non-uniform_random_variate_g...
You can do dynamic sampling: e.g. take double the samples, see what decimal in your result budges. Adjust.
[1] Bahadur, R. R. (1966). A note on quantiles in large samples. Annals of Mathematical Statistics, 37, 577–580.
If you want the n-9s rate of failure (eg., n=5, 99.999) for a system with a power-law performance distribution, you could be waiting much more than a billion samples to see a failure.
eg., 3E10 ms (30 bn samples) in a year, at 5-9s failure rate has 3E3 ms of failure (3 thousand samples) -- which will be lower given a sampling strategy.
Two observations: 1. With sufficiently large n, you can control your variance to an acceptable level. 2. The factor f is outside of your control. If your problem has a small density around the median, then you'll need to throw more samples to compensate for it.
I think your concern is about #2: you can always construct a pathological distribution to make the sampling-based approach unattainable. The paper provided guidance on what to expect when you apply the sampling-based approach.
That's a good point, though, that the 49th percentile and 51st percentile can be arbitrarily far from the median.
Uniform sampling also wasn't obviously simple, at least to me. There were thousands of log files involved, coming from hundreds of computers. Any single log file only had timings from a single computer. What kind of bias would be introduced by different approaches to distributing those log files to a cluster for the median calculation? Once the solution outlined in the previous comment was identified, that seemed simpler that trying to understand if we were talking about 49-51% or 40-50%. And if it was too big a margin, restructuring our infra to allow different log file distribution algorithms would have been far more complicated.
We were an infra-focused team building up a scalable data handling / computation platform in preparation to apply ML at a non-trivial scale. When we finally hired some people with deep stats knowledge 10 or 12 years ago, there was a lot of great learning for everyone about how each of our areas of expertise helped the others. I wish we'd found the right stats people earlier to help us understand things like this more deeply, more quickly.
In fact, this is often where decisions on what metrics to have in the first place come from. Ask why. You can go far without deep stats knowledge!
As an example: computing the .min() and .max() of an enumerable is two passes even though it could be done with one pass.
I'd love to see a language embrace a more efficient style similar to how a SQL does it, where you can elegant request this as a single pass over the data: "SELECT min(x), max(x) FROM y"
You could implement a similar (and simpler) fast path for types with contiguous memory by performing min and max per iteration instead.
Why go to the store and back twice to buy two things instead of buying two things in one trip? ;p
In all use-cases I've seen a close estimate of the median was enough.
But, you're right, I was lucky to work on a bunch of fun problems. That period, in particular, was pretty amazing. I was part of a fun, collaborative team working on hard problems. And management showed a lot of trust in us. We came up with some very interesting solutions, some by skill and some by luck, that set the foundation for years of growth that came after that (both revenue growth and technical platform growth).
I totally respect that—don't let me pressure you any further.
Yeah, sounds like you had the ideal work setup: cool problem requiring cool technical solutions with good management. That's wonderful!
It was a struggle until I figured out that knowledge of the precision and range of our data helped. These were timings, expressed in integer milliseconds. So they were non-negative, and I knew the 90th percentile was well under a second.
As the article mentions, finding a median typically involves something akin to sorting. With the above knowledge, bucket sort becomes available, with a slight tweak in my case. Even if the samples were floating point, the same approach could be used as long as an integer (or even fixed point) approximation that is very close to the true median is good enough, again assuming a known, relatively small range.
The idea is to build a dictionary where the keys are the timings in integer milliseconds and the values are a count of the keys' appearance in the data, i.e., a histogram of timings. The maximum timing isn't known, so to ensure the size of the dictionary doesn't get out of control, use the knowledge that the 90th percentile is well under a second and count everything over, say, 999ms in the 999ms bin. Then the dictionary will be limited to 2000 integers (keys in the range 0-999 and corresponding values) - this is the part that is different from an ordinary bucket sort. All of that is trivial to do in a single pass, even when distributed with MapReduce. Then it's easy to get the median from that dictionary / histogram.