KllSketchAggFloat¶

Overview¶

KllSketchAggFloat is a typed imperative aggregate function that builds a KLL (K, L, L) sketch from float values. It creates a compact probabilistic data structure that can be used for approximate quantile calculations and percentile estimations on streaming or large datasets.

Syntax¶

kll_sketch_agg_float(column [, k])

Arguments¶

Argument	Type	Description
column	FloatType	The float column or expression to build the sketch from
k	IntegerType (optional)	The accuracy parameter controlling sketch size and precision (defaults to implementation-defined value)

Return Type¶

BinaryType - Returns a serialized byte array representation of the KLL sketch.

Supported Data Types¶

FloatType only - Integer types are explicitly not supported to avoid precision loss from integer-to-float conversion
Users should use kll_sketch_agg_bigint for integer types

Algorithm¶

Creates a heap-based KllFloatsSketch instance with the specified k parameter
Iterates through input rows, updating the sketch with each non-null float value
Ignores null values during aggregation without affecting the sketch
Merges multiple sketch instances during distributed computation phases
Serializes the final sketch into a compact byte array format

Partitioning Behavior¶

This expression requires data shuffling for final aggregation:

Does not preserve input partitioning due to its aggregate nature
Requires shuffle operation to merge partial sketches from different partitions
Uses custom merge logic to combine KllFloatsSketch instances across executors

Edge Cases¶

Null values are completely ignored and do not affect the sketch construction
Empty input results in an empty sketch buffer created via createAggregationBuffer()
Invalid sketch buffers during merge operations throw QueryExecutionErrors.kllInvalidInputSketchBuffer
Deserialization failures from corrupted byte arrays are caught and converted to error exceptions
The function is marked as non-nullable, always returning a valid byte array

Code Generation¶

This expression does not support Tungsten code generation and falls back to interpreted mode, as it extends TypedImperativeAggregate which requires custom object manipulation for the KLL sketch state.

Examples¶

-- Basic usage with default k parameter
SELECT kll_sketch_agg_float(price) FROM sales;

-- Usage with custom accuracy parameter
SELECT kll_sketch_agg_float(temperature, 256) FROM weather_data;

// DataFrame API usage with default k
df.agg(expr("kll_sketch_agg_float(price)"))

// DataFrame API usage with custom k parameter  
df.agg(expr("kll_sketch_agg_float(temperature, 256)"))