org.apache.beam.sdk.io.aws2.kinesis

Class KinesisIO

java.lang.Object

org.apache.beam.sdk.io.aws2.kinesis.KinesisIO

@Experimental(value=SOURCE_SINK)
public final class KinesisIO
extends java.lang.Object

IO to read from Kinesis streams.

Reading from Kinesis

Example usages:

 p.apply(KinesisIO.read()
     .withStreamName("streamName")
     .withInitialPositionInStream(InitialPositionInStream.LATEST)
  .apply( ... ) // other transformations
 

At a minimum you have to provide:

• the name of the stream to read
• the position in the stream where to start reading, e.g. InitialPositionInStream.LATEST, InitialPositionInStream.TRIM_HORIZON, or alternatively, using an arbitrary point in time with KinesisIO.Read.withInitialTimestampInStream(Instant).

Watermarks

Kinesis IO uses arrival time for watermarks by default. To use processing time instead, use KinesisIO.Read.withProcessingTimeWatermarkPolicy():

 p.apply(KinesisIO.read()
    .withStreamName("streamName")
    .withInitialPositionInStream(InitialPositionInStream.LATEST)
    .withProcessingTimeWatermarkPolicy())
 

It is also possible to specify a custom watermark policy to control watermark computation using KinesisIO.Read.withCustomWatermarkPolicy(WatermarkPolicyFactory). This requires implementing WatermarkPolicy with a corresponding WatermarkPolicyFactory.

Throttling

By default Kinesis IO will poll the Kinesis getRecords() API as fast as possible as long as records are returned. The RateLimitPolicyFactory.DefaultRateLimiter will start throttling once getRecords() returns an empty response or if API calls get throttled by AWS.

A RateLimitPolicy is always applied to each shard individually.

You may provide a custom rate limit policy using KinesisIO.Read.withCustomRateLimitPolicy(RateLimitPolicyFactory). This requires implementing RateLimitPolicy with a corresponding RateLimitPolicyFactory.

Writing to Kinesis

Example usages:

 PCollection<KV<String, byte[]>> data = ...;

 data.apply(KinesisIO.write()
     .withStreamName("streamName")
     .withPartitionKey(KV::getKey)
     .withSerializer(KV::getValue);
 

Note: Usage of KV is just for illustration purposes here.

At a minimum you have to provide:

• the name of the Kinesis stream to write to,
• a KinesisPartitioner to distribute records across shards of the stream
• and a function to serialize your data to bytes on the stream

Though, generally, it's recommended to configure client retries using ClientConfiguration, see below.

Partitioning of writes

Choosing the right partitioning strategy by means of a KinesisPartitioner is one of the key considerations when writing to Kinesis. Typically, you should aime to evenly distribute data across all shards of the stream.

Partition keys are used as input to a hash function that maps the partition key and associated data to a specific shard. If the cardinality of your partition keys is of the same order of magnitude as the number of shards in the stream, the hash function will likely not distribute your keys evenly among shards. This may result in heavily skewed shards with some shards not utilized at all.

If you require finer control over the distribution of records, override KinesisPartitioner.getExplicitHashKey(Object) according to your needs.

Aggregation of records

To better leverage Kinesis API limits and to improve producer throughput, the writer aggregates multiple users records into an aggregated KPL record.

However, only records with the same effective hash key are aggregated, in which the effective hash key is either the explicit hash key if defined, or otherwise the hashed partition key.

Record aggregation can be explicitly disabled using KinesisIO.Write.withRecordAggregationDisabled().

Configuration of AWS clients

AWS clients for all AWS IOs can be configured using AwsOptions, e.g. --awsRegion=us-west-1. AwsOptions contain reasonable defaults based on default providers for Region and AwsCredentialsProvider.

If you require more advanced configuration, you may change the ClientBuilderFactory using AwsOptions.setClientBuilderFactory(Class).

Configuration for a specific IO can be overwritten using withClientConfiguration(), which also allows to configure the retry behavior for the respective IO.

Retries

Retries for failed requests can be configured using ClientConfiguration.Builder#retry(Consumer) and are handled by the AWS SDK unless there's a partial success (batch requests). The SDK uses a backoff strategy with equal jitter for computing the delay before the next retry.

Note: Once retries are exhausted the error is surfaced to the runner which may then opt to retry the current partition in entirety or abort if the max number of retries of the runner is reached.

Nested Class Summary

Nested Classes

Modifier and Type	Class and Description
static class	KinesisIO.Read Implementation of read().
static class	KinesisIO.RecordAggregation Configuration of Kinesis record aggregation.
static class	KinesisIO.Write<T> Implementation of write().

Constructor Summary

Constructors

Constructor and Description
KinesisIO()

Method Summary

Modifier and Type	Method and Description
static KinesisIO.Read	read() Returns a new KinesisIO.Read transform for reading from Kinesis.
static <T> KinesisIO.Write<T>	write() Returns a new KinesisIO.Write transform for writing to Kinesis.

Methods inherited from class java.lang.Object

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

Constructor Detail

KinesisIO

public KinesisIO()

Method Detail

read

public static KinesisIO.Read read()

Returns a new KinesisIO.Read transform for reading from Kinesis.

write

public static <T> KinesisIO.Write<T> write()

Returns a new KinesisIO.Write transform for writing to Kinesis.