HyperLogLog Functions

Presto implements the approx_distinct() function using the HyperLogLog data structure.

Data Structures

Presto implements HyperLogLog data sketches as a set of 32-bit buckets which store a maximum hash. They can be stored sparsely (as a map from bucket ID to bucket), or densely (as a contiguous memory block). The HyperLogLog data structure starts as the sparse representation, switching to dense when it is more efficient. The P4HyperLogLog structure is initialized densely and remains dense for its lifetime.

HyperLogLog implicitly casts to P4HyperLogLog, while one can explicitly cast HyperLogLog to P4HyperLogLog:

  1. cast(hll AS P4HyperLogLog)

Serialization

Data sketches can be serialized to and deserialized from varbinary. This allows them to be stored for later use. Combined with the ability to merge multiple sketches, this allows one to calculate approx_distinct() of the elements of a partition of a query, then for the entirety of a query with very little cost.

For example, calculating the HyperLogLog for daily unique users will allow weekly or monthly unique users to be calculated incrementally by combining the dailies. This is similar to computing weekly revenue by summing daily revenue. Uses of approx_distinct() with GROUPING SETS can be converted to use HyperLogLog. Examples:

  1. CREATE TABLE visit_summaries (
  2. visit_date date,
  3. hll varbinary
  4. );
  5. INSERT INTO visit_summaries
  6. SELECT visit_date, cast(approx_set(user_id) AS varbinary)
  7. FROM user_visits
  8. GROUP BY visit_date;
  9. SELECT cardinality(merge(cast(hll AS HyperLogLog))) AS weekly_unique_users
  10. FROM visit_summaries
  11. WHERE visit_date >= current_date - interval '7' day;

Functions

approx_set(x) → HyperLogLog

Returns the HyperLogLog sketch of the input data set of x. The value of the maximum standard error is defaulted to 0.01625. This data sketch underlies approx_distinct() and can be stored and used later by calling cardinality().

approx_set(x, e) → HyperLogLog

Returns the HyperLogLog sketch of the input data set of x, with a maximum standard error of e. The current implementation of this function requires that e be in the range of [0.0040625, 0.26000]. This data sketch underlies approx_distinct() and can be stored and used later by calling cardinality().

cardinality(hll) → bigint

This will perform approx_distinct() on the data summarized by the hll HyperLogLog data sketch.

empty_approx_set() → HyperLogLog

Returns an empty HyperLogLog. The value of the maximum standard error is defaulted to 0.01625.

empty_approx_set(e) → HyperLogLog

Returns an empty HyperLogLog with a maximum standard error of e. The current implementation of this function requires that e be in the range of [0.0040625, 0.26000].

merge(HyperLogLog) → HyperLogLog

Returns the HyperLogLog of the aggregate union of the individual hll HyperLogLog structures.

merge_hll(array(HyperLogLog)) → HyperLogLog

Returns the HyperLogLog of the union of an array hll HyperLogLog structures.