DataSketches HLL Sketch module

This module provides Apache Druid aggregators for distinct counting based on HLL sketch from Apache DataSketches library. At ingestion time, this aggregator creates the HLL sketch objects to store in Druid segments. By default, Druid reads and merges sketches at query time. The default result is the estimate of the number of distinct values presented to the sketch. You can also use post aggregators to produce a union of sketch columns in the same row. You can use the HLL sketch aggregator on any column to estimate its cardinality.

To use this aggregator, make sure you include the extension in your config file:

  1. druid.extensions.loadList=["druid-datasketches"]

For additional sketch types supported in Druid, see DataSketches extension.

Aggregators

PropertyDescriptionRequired?
typeEither HLLSketchBuild or HLLSketchMerge.yes
nameString representing the output column to store sketch values.yes
fieldNameThe name of the input field.yes
lgKlog2 of K that is the number of buckets in the sketch, parameter that controls the size and the accuracy. Must be between 4 and 21 inclusively.no, defaults to 12
tgtHllTypeThe type of the target HLL sketch. Must be HLL_4, HLL_6 or HLL_8no, defaults to HLL_4
roundRound off values to whole numbers. Only affects query-time behavior and is ignored at ingestion-time.no, defaults to false
shouldFinalizeReturn the final double type representing the estimate rather than the intermediate sketch type itself. In addition to controlling the finalization of this aggregator, you can control whether all aggregators are finalized with the query context parameters finalize and sqlFinalizeOuterSketches.no, defaults to true

The default lgK value has proven to be sufficient for most use cases; expect only very negligible improvements in accuracy with lgK values over 16 in normal circumstances.

HLLSketchBuild aggregator

  1. {
  2.   "type": "HLLSketchBuild",
  3.   "name": <output name>,
  4.   "fieldName": <metric name>,
  5.   "lgK": <size and accuracy parameter>,
  6.   "tgtHllType": <target HLL type>,
  7.   "round": <false | true>
  8.  }

The HLLSketchBuild aggregator builds an HLL sketch object from the specified input column. When used during ingestion, Druid stores pre-generated HLL sketch objects in the datasource instead of the raw data from the input column. When applied at query time on an existing dimension, you can use the resulting column as an intermediate dimension by the post-aggregators.

It is very common to use HLLSketchBuild in combination with rollup to create a metric on high-cardinality columns. In this example, a metric called userid_hll is included in the metricsSpec. This will perform a HLL sketch on the userid field at ingestion time, allowing for highly-performant approximate COUNT DISTINCT query operations and improving roll-up ratios when userid is then left out of the dimensionsSpec.

  1. "metricsSpec": [
  2.   {
  3.     "type": "HLLSketchBuild",
  4.     "name": "userid_hll",
  5.     "fieldName": "userid",
  6.     "lgK": 12,
  7.     "tgtHllType": "HLL_4"
  8.   }
  9. ]

HLLSketchMerge aggregator

  1. {
  2.   "type": "HLLSketchMerge",
  3.   "name": <output name>,
  4.   "fieldName": <metric name>,
  5.   "lgK": <size and accuracy parameter>,
  6.   "tgtHllType": <target HLL type>,
  7.   "round": <false | true>
  8. }

You can use the HLLSketchMerge aggregator to ingest pre-generated sketches from an input dataset. For example, you can set up a batch processing job to generate the sketches before sending the data to Druid. You must serialize the sketches in the input dataset to Base64-encoded bytes. Then, specify HLLSketchMerge for the input column in the native ingestion metricsSpec.

Post aggregators

Estimate

Returns the distinct count estimate as a double.

  1. {
  2.   "type": "HLLSketchEstimate",
  3.   "name": <output name>,
  4.   "field": <post aggregator that returns an HLL Sketch>,
  5.   "round": <if true, round the estimate. Default is false>
  6. }

Estimate with bounds

Returns a distinct count estimate and error bounds from an HLL sketch. The result will be an array containing three double values: estimate, lower bound and upper bound. The bounds are provided at a given number of standard deviations (optional, defaults to 1). This must be an integer value of 1, 2 or 3 corresponding to approximately 68.3%, 95.4% and 99.7% confidence intervals.

  1. {
  2.   "type": "HLLSketchEstimateWithBounds",
  3.   "name": <output name>,
  4.   "field": <post aggregator that returns an HLL Sketch>,
  5.   "numStdDev": <number of standard deviations: 1 (default), 2 or 3>
  6. }

Union

  1. {
  2.   "type": "HLLSketchUnion",
  3.   "name": <output name>,
  4.   "fields": <array of post aggregators that return HLL sketches>,
  5.   "lgK": <log2 of K for the target sketch>,
  6.   "tgtHllType": <target HLL type>
  7. }

Sketch to string

Human-readable sketch summary for debugging.

  1. {
  2.   "type": "HLLSketchToString",
  3.   "name": <output name>,
  4.   "field": <post aggregator that returns an HLL Sketch>
  5. }