org.apache.beam.sdk.extensions.sketching

Class ApproximateDistinct

java.lang.Object

org.apache.beam.sdk.extensions.sketching.ApproximateDistinct

@Experimental
public final class ApproximateDistinct
extends java.lang.Object

PTransforms for computing the approximate number of distinct elements in a stream.

This class relies on the HyperLogLog algorithm, and more precisely HyperLogLog+, the improved version of Google.

References

The implementation comes from Addthis' Stream-lib library.
The original paper of the HyperLogLog is available here.
A paper from the same authors to have a clearer view of the algorithm is available here.
Google's HyperLogLog+ version is detailed in this paper.

Parameters

Two parameters can be tuned in order to control the computation's accuracy:

• Precision: p
  Controls the accuracy of the estimation. The precision value will have an impact on the number of buckets used to store information about the distinct elements.
  In general one can expect a relative error of about 1.1 / sqrt(2^p). The value should be of at least 4 to guarantee a minimal accuracy.
  By default, the precision is set to 12 for a relative error of around 2%.
• Sparse Precision: sp
  Used to create a sparse representation in order to optimize memory and improve accuracy at small cardinalities.
  The value of sp should be greater than p, but lower than 32.
  By default, the sparse representation is not used (sp = 0). One should use it if the cardinality may be less than 12000.

Examples

There are 2 ways of using this class:

• Use the PTransforms that return PCollection<Long> corresponding to the estimate number of distinct elements in the input PCollection of objects or for each key in a PCollection of KVs.
• Use the ApproximateDistinct.ApproximateDistinctFn CombineFn that is exposed in order to make advanced processing involving the HyperLogLogPlus structure which resumes the stream.

Using the Transforms

Example 1: globally default use

 PCollection<Integer> input = ...;
 PCollection<Long> hllSketch = input.apply(ApproximateDistinct.<Integer>globally());
 

Example 2: per key default use

 PCollection<Integer, String> input = ...;
 PCollection<Integer, Long> hllSketches = input.apply(ApproximateDistinct
                .<Integer, String>perKey());
 

Example 3: tune precision and use sparse representation

One can tune the precision and sparse precision parameters in order to control the accuracy and the memory. The tuning works exactly the same for globally() and perKey().

 int precision = 15;
 int sparsePrecision = 25;
 PCollection<Double> input = ...;
 PCollection<Long> hllSketch = input.apply(ApproximateDistinct
                .<Double>globally()
                .withPrecision(precision)
                .withSparsePrecision(sparsePrecision));
 

Using the ApproximateDistinct.ApproximateDistinctFn CombineFn

The CombineFn does the same thing as the transform but it can be used in cases where you want to manipulate the HyperLogLogPlus sketch, for example if you want to store it in a database to have a backup. It can also be used in stateful processing or in CombineFns.ComposedCombineFn.

Example 1: basic use

This example is not really interesting but show how you can properly create an ApproximateDistinct.ApproximateDistinctFn. One must always specify a coder using the ApproximateDistinct.ApproximateDistinctFn.create(Coder) method.

 PCollection<Integer> input = ...;
 PCollection<HyperLogLogPlus> output = input.apply(Combine.globally(ApproximateDistinctFn
                 .<Integer>create(BigEndianIntegerCoder.of()));
 

Example 2: use the Combine.CombineFn in a stateful ParDo

One may want to use the ApproximateDistinct.ApproximateDistinctFn in a stateful ParDo in order to make some processing depending on the current cardinality of the stream.
For more information about stateful processing see the blog spot on this topic here.

Here is an example of DoFn using an ApproximateDistinct.ApproximateDistinctFn as a CombiningState:

  class StatefulCardinality<V> extends DoFn<V, OutputT> {
    @StateId("hyperloglog")
    private final StateSpec<CombiningState<V, HyperLogLogPlus, HyperLogLogPlus>>
      indexSpec;

    public StatefulCardinality(ApproximateDistinctFn<V> fn) {
     indexSpec = StateSpecs.combining(fn);
   }

   @ProcessElement
   public void processElement(
      ProcessContext context,
       @StateId("hllSketch")
       CombiningState<V, HyperLogLogPlus, HyperLogLogPlus> hllSketch) {
     long current = MoreObjects.firstNonNull(hllSketch.getAccum().cardinality(), 0L);
     hllSketch.add(context.element());
     context.output(...);
   }
 }
 

Then the DoFn can be called like this:

 PCollection<V> input = ...;
 ApproximateDistinctFn<V> myFn = ApproximateDistinctFn.create(input.getCoder());
 PCollection<V> = input.apply(ParDo.of(new StatefulCardinality<>(myFn)));
 

Example 3: use the RetrieveCardinality utility class

One may want to retrieve the cardinality as a long after making some advanced processing using the HyperLogLogPlus structure.
The RetrieveCardinality utility class provides an easy way to do so:

 PCollection<MyObject> input = ...;
 PCollection<HyperLogLogPlus> hll = input.apply(Combine.globally(ApproximateDistinctFn
                  .<MyObject>create(new MyObjectCoder())
                  .withSparseRepresentation(20)));

  // Some advanced processing
  PCollection<SomeObject> advancedResult = hll.apply(...);

  PCollection<Long> cardinality = hll.apply(ApproximateDistinct.RetrieveCardinality.globally());

 

Consider using the HllCount.Init transform in the zetasketch extension module if you need to create sketches compatible with Google Cloud BigQuery. For more details about using HllCount and the zetasketch extension module, see https://s.apache.org/hll-in-beam#bookmark=id.v6chsij1ixo7

Warning: this class is experimental. Its API is subject to change in future versions of Beam. For example, it may be merged with the ApproximateUnique transform.

Nested Class Summary

Nested Classes

Modifier and Type	Class and Description
static class	ApproximateDistinct.ApproximateDistinctFn<InputT> Implements the Combine.CombineFn of ApproximateDistinct transforms.
static class	ApproximateDistinct.GloballyDistinct<InputT> Implementation of globally().
static class	ApproximateDistinct.HyperLogLogPlusCoder Coder for HyperLogLogPlus class.
static class	ApproximateDistinct.PerKeyDistinct<K,V> Implementation of perKey().

Constructor Summary

Constructors

Constructor and Description
ApproximateDistinct()

Method Summary

Modifier and Type	Method and Description
static <InputT> ApproximateDistinct.GloballyDistinct<InputT>	globally() Computes the approximate number of distinct elements in the input PCollection<InputT> and returns a PCollection<Long>.
static <K,V> ApproximateDistinct.PerKeyDistinct<K,V>	perKey() Like globally() but per key, i.e computes the approximate number of distinct values per key in a PCollection<KV<K, V>> and returns PCollection<KV<K, Long>>.
static long	precisionForRelativeError(double relativeError) Computes the precision based on the desired relative error.
static double	relativeErrorForPrecision(int p)

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Constructor Detail

ApproximateDistinct

public ApproximateDistinct()

Method Detail

globally

public static <InputT> ApproximateDistinct.GloballyDistinct<InputT> globally()

Computes the approximate number of distinct elements in the input PCollection<InputT> and returns a PCollection<Long>.

Type Parameters:

InputT - the type of the elements in the input PCollection

perKey

public static <K,V> ApproximateDistinct.PerKeyDistinct<K,V> perKey()

Like globally() but per key, i.e computes the approximate number of distinct values per key in a PCollection<KV<K, V>> and returns PCollection<KV<K, Long>>.

Type Parameters:

K - type of the keys mapping the elements

V - type of the values being combined per key

precisionForRelativeError

public static long precisionForRelativeError(double relativeError)

Computes the precision based on the desired relative error.

According to the paper, the mean squared error is bounded by the following formula:

b(m) / sqrt(m)
 Where m is the number of buckets used ( p = log2(m))
 and  b(m) < 1.106 for  m > 16 (and p > 4).
 

WARNING:
This does not mean relative error in the estimation can't be higher.
This only means that on average the relative error will be lower than the desired relative error.
Nevertheless, the more elements arrive in the PCollection, the lower the variation will be.
Indeed, this is like when you throw a dice millions of time: the relative frequency of each different result {1,2,3,4,5,6} will get closer to 1/6.

Parameters:

relativeError - the mean squared error should be in the interval ]0,1]

Returns:

the minimum precision p in order to have the desired relative error on average.

relativeErrorForPrecision

public static double relativeErrorForPrecision(int p)

Parameters:

p - the precision i.e. the number of bits used for indexing the buckets

Returns:

the Mean squared error of the Estimation of cardinality to expect for the given value of p.