Add compute functions to ExponentialHistogramState #136749
Conversation
elasticsearchmachine
added
external-contributor
|
Pinging @elastic/es-analytical-engine (Team:Analytics) |
elasticsearchmachine
added
the
Team:Analytics
...r/src/main/java/org/elasticsearch/search/aggregations/metrics/ExponentialHistogramState.java
Outdated
Show resolved
Hide resolved
...r/src/main/java/org/elasticsearch/search/aggregations/metrics/ExponentialHistogramState.java
Show resolved
Hide resolved
| public double cdf(double x) { | ||
| ExponentialHistogram histogram = histogram(); | ||
| long numValuesLess = ExponentialHistogramQuantile.estimateRank(histogram, x, false); | ||
| long numValuesLessOrEqual = ExponentialHistogramQuantile.estimateRank(histogram, x, true); |
There was a problem hiding this comment.
Just noticed, these don't check if the value is outside min/max? You should do that separately, return 0 for < min and 1 for > max.
There was a problem hiding this comment.
Correctness wise, we already do this here.
So you are suggesting to do this for performance reasons?
I'm not sure if this is worth it, because it increases the complexity: We currently clamp the bucket POLRE to the histogram min / max. So if the POLRE was clamped, we need to return different values based on whether the requested rank was inclusive or exclusive equal values.
I don't think that this is a common enough case to justify this extra code
There was a problem hiding this comment.
Why does this increase the complexity? Isn't this a simple condition to return early, that can be added at the top of ExponentialHistogramQuantile.estimateRank? It should be uncontroversial, in terms of correctness, and more efficient.
Btw, this is unrelated to this pr so not a blocker.
...r/src/main/java/org/elasticsearch/search/aggregations/metrics/ExponentialHistogramState.java
Outdated
Show resolved
Hide resolved
...r/src/main/java/org/elasticsearch/search/aggregations/metrics/ExponentialHistogramState.java
Show resolved
Hide resolved
4 participants
Part of #135625 , follow up of #136075.
Makes
ExponentialHistogramStatemimic theTDigestStateprovided functionality used by aggregations, so that we can build the drop-in replacementHistogramStateconsisting of both. This will then allow us to apply e.g. percentile aggregation on exponential histograms in addition to T-Digests and a mix of both.We also had to implement a
centroids()functionality, which is implemented by returning the mean values of the populated histogram buckets. Based on my research,centroids()is only used in the boxplot aggregation in order to define the length of the whiskers, where this usage should be fine.