Computing exact counts requires loading values into a set and returning its size. This doesn’t scale when working on high-cardinality sets and/or large values as the required memory usage and the need to communicate those per-shard sets between nodes would utilize too many resources of the cluster.

This COUNT_DISTINCT function is based on the HyperLogLog++ algorithm, which counts based on the hashes of the values with some interesting properties:

For a precision threshold of c, the implementation that we are using requires about c * 8 bytes.

The following chart shows how the error varies before and after the threshold:

For all 3 thresholds, counts have been accurate up to the configured threshold. Although not guaranteed, this is likely to be the case. Accuracy in practice depends on the dataset in question. In general, most datasets show consistently good accuracy. Also note that even with a threshold as low as 100, the error remains very low (1-6% as seen in the above graph) even when counting millions of items.

The HyperLogLog++ algorithm depends on the leading zeros of hashed values, the exact distributions of hashes in a dataset can affect the accuracy of the cardinality.

The COUNT_DISTINCT function takes an optional second parameter to configure the precision discussed previously.

FROM hosts
| STATS COUNT_DISTINCT(ip0, 80000), COUNT_DISTINCT(ip1, 5)