org.apache.beam.sdk.extensions.zetasketch

Class HllCount

java.lang.Object

org.apache.beam.sdk.extensions.zetasketch.HllCount

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

PTransforms to compute HyperLogLogPlusPlus (HLL++) sketches on data streams based on the ZetaSketch implementation.

HLL++ is an algorithm implemented by Google that estimates the count of distinct elements in a data stream. HLL++ requires significantly less memory than the linear memory needed for exact computation, at the cost of a small error. Cardinalities of arbitrary breakdowns can be computed using the HLL++ sketch. See this published paper for details about the algorithm.

HLL++ functions are also supported in Google Cloud BigQuery. The HllCount PTransforms provided here produce and consume sketches compatible with BigQuery.

For detailed design of this class, see https://s.apache.org/hll-in-beam.

Examples

Example 1: Create long-type sketch for a PCollection<Long> and specify precision

 PCollection<Long> input = ...;
 int p = ...;
 PCollection<byte[]> sketch = input.apply(HllCount.Init.forLongs().withPrecision(p).globally());
 

Example 2: Create bytes-type sketch for a PCollection<KV<String, byte[]>>

 PCollection<KV<String, byte[]>> input = ...;
 PCollection<KV<String, byte[]>> sketch = input.apply(HllCount.Init.forBytes().perKey());
 

Example 3: Merge existing sketches in a PCollection<byte[]> into a new one

 PCollection<byte[]> sketches = ...;
 PCollection<byte[]> mergedSketch = sketches.apply(HllCount.MergePartial.globally());
 

Example 4: Estimates the count of distinct elements in a PCollection<String>

 PCollection<String> input = ...;
 PCollection<Long> countDistinct =
     input.apply(HllCount.Init.forStrings().globally()).apply(HllCount.Extract.globally());
 

Note: Currently HllCount does not work on FnAPI workers. See Jira ticket [BEAM-7879].

Nested Class Summary

Nested Classes

Modifier and Type	Class and Description
static class	HllCount.Extract Provides PTransforms to extract the estimated count of distinct elements (as Longs) from each HLL++ sketch.
static class	HllCount.Init Provides PTransforms to aggregate inputs into HLL++ sketches.
static class	HllCount.MergePartial Provides PTransforms to merge HLL++ sketches into a new sketch.

Field Summary

Fields

Modifier and Type	Field and Description
static int	DEFAULT_PRECISION The default precision value used in HllCount.Init.Builder.withPrecision(int) is 15.
static int	MAXIMUM_PRECISION The maximum precision value you can set in HllCount.Init.Builder.withPrecision(int) is 24.
static int	MINIMUM_PRECISION The minimum precision value you can set in HllCount.Init.Builder.withPrecision(int) is 10.

Method Summary

Modifier and Type	Method and Description
static byte[]	getSketchFromByteBuffer(java.nio.ByteBuffer bf) Converts the passed-in sketch from ByteBuffer to byte[], mapping null ByteBuffers (representing empty sketches) to empty byte[]s.

Methods inherited from class java.lang.Object

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

Field Detail

MINIMUM_PRECISION

public static final int MINIMUM_PRECISION

The minimum precision value you can set in HllCount.Init.Builder.withPrecision(int) is 10.

See Also:

Constant Field Values

MAXIMUM_PRECISION

public static final int MAXIMUM_PRECISION

The maximum precision value you can set in HllCount.Init.Builder.withPrecision(int) is 24.

See Also:

Constant Field Values

DEFAULT_PRECISION

public static final int DEFAULT_PRECISION

The default precision value used in HllCount.Init.Builder.withPrecision(int) is 15.

See Also:

Constant Field Values

Method Detail

getSketchFromByteBuffer

public static byte[] getSketchFromByteBuffer(@Nullable
                                             java.nio.ByteBuffer bf)

Converts the passed-in sketch from ByteBuffer to byte[], mapping null ByteBuffers (representing empty sketches) to empty byte[]s.

Utility method to convert sketches materialized with ZetaSQL/BigQuery to valid inputs for Beam HllCount transforms.