Apache Druid aggregators based on [Apache DataSketches](https://datasketches.apache.org/) library. Sketches are data structures implementing approximate streaming mergeable algorithms. Sketches can be ingested from the outside of Druid or built from raw data at ingestion time. Sketches can be stored in Druid segments as additive metrics.
* [Theta sketch](datasketches-theta.md) - approximate distinct counting with set operations (union, intersection and set difference).
* [Tuple sketch](datasketches-tuple.md) - extension of Theta sketch to support values associated with distinct keys (arrays of numeric values in this specialized implementation).
* [Quantiles sketch](datasketches-quantiles.md) - approximate distribution of comparable values to obtain ranks, quantiles and histograms. This is a specialized implementation for numeric values.
* [HLL sketch](datasketches-hll.md) - approximate distinct counting using very compact HLL sketch.