org.apache.beam.sdk.io.kafka

Class KafkaIO.Write<K,V>

java.lang.Object

org.apache.beam.sdk.transforms.PTransform<PCollection<KV<K,V>>,PDone>

org.apache.beam.sdk.io.kafka.KafkaIO.Write<K,V>

All Implemented Interfaces:

java.io.Serializable, HasDisplayData

Enclosing class:

KafkaIO

public abstract static class KafkaIO.Write<K,V>
extends PTransform<PCollection<KV<K,V>>,PDone>

A PTransform to write to a Kafka topic. See KafkaIO for more information on usage and configuration.

See Also:

Serialized Form

Field Summary

Fields inherited from class org.apache.beam.sdk.transforms.PTransform

name

Constructor Summary

Constructors

Constructor and Description
Write()

Method Summary

Modifier and Type	Method and Description
PDone	expand(PCollection<KV<K,V>> input) Override this method to specify how this PTransform should be expanded on the given InputT.
void	populateDisplayData(DisplayData.Builder builder) Register display data for the given transform or component.
KafkaIO.Write<K,V>	updateProducerProperties(java.util.Map<java.lang.String,java.lang.Object> configUpdates) Adds the given producer properties, overriding old values of properties with the same key.
void	validate(PipelineOptions options) Called before running the Pipeline to verify this transform is fully and correctly specified.
PTransform<PCollection<V>,PDone>	values() Writes just the values to Kafka.
KafkaIO.Write<K,V>	withBootstrapServers(java.lang.String bootstrapServers) Returns a new KafkaIO.Write transform with Kafka producer pointing to bootstrapServers.
KafkaIO.Write<K,V>	withConsumerFactoryFn(SerializableFunction<java.util.Map<java.lang.String,java.lang.Object>,? extends org.apache.kafka.clients.consumer.Consumer<?,?>> consumerFactoryFn) When exactly-once semantics are enabled (see withEOS(int, String)), the sink needs to fetch previously stored state with Kafka topic.
KafkaIO.Write<K,V>	withEOS(int numShards, java.lang.String sinkGroupId) Provides exactly-once semantics while writing to Kafka, which enables applications with end-to-end exactly-once guarantees on top of exactly-once semantics within Beam pipelines.
KafkaIO.Write<K,V>	withInputTimestamp() The timestamp for each record being published is set to timestamp of the element in the pipeline.
KafkaIO.Write<K,V>	withKeySerializer(java.lang.Class<? extends org.apache.kafka.common.serialization.Serializer<K>> keySerializer) Sets a Serializer for serializing key (if any) to bytes.
KafkaIO.Write<K,V>	withProducerFactoryFn(SerializableFunction<java.util.Map<java.lang.String,java.lang.Object>,org.apache.kafka.clients.producer.Producer<K,V>> producerFactoryFn) Sets a custom function to create Kafka producer.
KafkaIO.Write<K,V>	withPublishTimestampFunction(KafkaPublishTimestampFunction<KV<K,V>> timestampFunction) A function to provide timestamp for records being published.
KafkaIO.Write<K,V>	withTopic(java.lang.String topic) Sets the Kafka topic to write to.
KafkaIO.Write<K,V>	withValueSerializer(java.lang.Class<? extends org.apache.kafka.common.serialization.Serializer<V>> valueSerializer) Sets a Serializer for serializing value to bytes.

Methods inherited from class org.apache.beam.sdk.transforms.PTransform

getAdditionalInputs, getDefaultOutputCoder, getDefaultOutputCoder, getDefaultOutputCoder, getKindString, getName, toString

Methods inherited from class java.lang.Object

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

Constructor Detail

Write

public Write()

Method Detail

withBootstrapServers

public KafkaIO.Write<K,V> withBootstrapServers(java.lang.String bootstrapServers)

Returns a new KafkaIO.Write transform with Kafka producer pointing to bootstrapServers.

withTopic

public KafkaIO.Write<K,V> withTopic(java.lang.String topic)

Sets the Kafka topic to write to.

withKeySerializer

public KafkaIO.Write<K,V> withKeySerializer(java.lang.Class<? extends org.apache.kafka.common.serialization.Serializer<K>> keySerializer)

Sets a Serializer for serializing key (if any) to bytes.

A key is optional while writing to Kafka. Note when a key is set, its hash is used to determine partition in Kafka (see ProducerRecord for more details).

withValueSerializer

public KafkaIO.Write<K,V> withValueSerializer(java.lang.Class<? extends org.apache.kafka.common.serialization.Serializer<V>> valueSerializer)

Sets a Serializer for serializing value to bytes.

updateProducerProperties

public KafkaIO.Write<K,V> updateProducerProperties(java.util.Map<java.lang.String,java.lang.Object> configUpdates)

Adds the given producer properties, overriding old values of properties with the same key.

withProducerFactoryFn

public KafkaIO.Write<K,V> withProducerFactoryFn(SerializableFunction<java.util.Map<java.lang.String,java.lang.Object>,org.apache.kafka.clients.producer.Producer<K,V>> producerFactoryFn)

Sets a custom function to create Kafka producer. Primarily used for tests. Default is KafkaProducer

withInputTimestamp

public KafkaIO.Write<K,V> withInputTimestamp()

The timestamp for each record being published is set to timestamp of the element in the pipeline. This is equivalent to withPublishTimestampFunction((e, ts) -> ts).
NOTE: Kafka's retention policies are based on message timestamps. If the pipeline is processing messages from the past, they might be deleted immediately by Kafka after being published if the timestamps are older than Kafka cluster's log.retention.hours.

withPublishTimestampFunction

public KafkaIO.Write<K,V> withPublishTimestampFunction(KafkaPublishTimestampFunction<KV<K,V>> timestampFunction)

A function to provide timestamp for records being published.
NOTE: Kafka's retention policies are based on message timestamps. If the pipeline is processing messages from the past, they might be deleted immediately by Kafka after being published if the timestamps are older than Kafka cluster's log.retention.hours.

withEOS

public KafkaIO.Write<K,V> withEOS(int numShards,
                                  java.lang.String sinkGroupId)

Provides exactly-once semantics while writing to Kafka, which enables applications with end-to-end exactly-once guarantees on top of exactly-once semantics within Beam pipelines. It ensures that records written to sink are committed on Kafka exactly once, even in the case of retries during pipeline execution even when some processing is retried. Retries typically occur when workers restart (as in failure recovery), or when the work is redistributed (as in an autoscaling event).

Beam runners typically provide exactly-once semantics for results of a pipeline, but not for side effects from user code in transform. If a transform such as Kafka sink writes to an external system, those writes might occur more than once. When EOS is enabled here, the sink transform ties checkpointing semantics in compatible Beam runners and transactions in Kafka (version 0.11+) to ensure a record is written only once. As the implementation relies on runners checkpoint semantics, not all the runners are compatible. The sink throws an exception during initialization if the runner is not whitelisted. Flink runner is one of the runners whose checkpoint semantics are not compatible with current implementation (hope to provide a solution in near future). Dataflow runner and Spark runners are whitelisted as compatible.

Note on performance: Exactly-once sink involves two shuffles of the records. In addition to cost of shuffling the records among workers, the records go through 2 serialization-deserialization cycles. Depending on volume and cost of serialization, the CPU cost might be noticeable. The CPU cost can be reduced by writing byte arrays (i.e. serializing them to byte before writing to Kafka sink).

Parameters:

numShards - Sets sink parallelism. The state metadata stored on Kafka is stored across this many virtual partitions using sinkGroupId. A good rule of thumb is to set this to be around number of partitions in Kafka topic.

sinkGroupId - The group id used to store small amount of state as metadata on Kafka. It is similar to consumer group id used with a KafkaConsumer. Each job should use a unique group id so that restarts/updates of job preserve the state to ensure exactly-once semantics. The state is committed atomically with sink transactions on Kafka. See KafkaProducer.sendOffsetsToTransaction(Map, String) for more information. The sink performs multiple sanity checks during initialization to catch common mistakes so that it does not end up using state that does not seem to be written by the same job.

withConsumerFactoryFn

public KafkaIO.Write<K,V> withConsumerFactoryFn(SerializableFunction<java.util.Map<java.lang.String,java.lang.Object>,? extends org.apache.kafka.clients.consumer.Consumer<?,?>> consumerFactoryFn)

When exactly-once semantics are enabled (see withEOS(int, String)), the sink needs to fetch previously stored state with Kafka topic. Fetching the metadata requires a consumer. Similar to KafkaIO.Read.withConsumerFactoryFn(SerializableFunction), a factory function can be supplied if required in a specific case. The default is KafkaConsumer.

values

public PTransform<PCollection<V>,PDone> values()

Writes just the values to Kafka. This is useful for writing collections of values rather thank KVs.

expand

public PDone expand(PCollection<KV<K,V>> input)

Description copied from class: PTransform

Override this method to specify how this PTransform should be expanded on the given InputT.

NOTE: This method should not be called directly. Instead apply the PTransform should be applied to the InputT using the apply method.

Composite transforms, which are defined in terms of other transforms, should return the output of one of the composed transforms. Non-composite transforms, which do not apply any transforms internally, should return a new unbound output and register evaluators (via backend-specific registration methods).

Specified by:

expand in class PTransform<PCollection<KV<K,V>>,PDone>

validate

public void validate(PipelineOptions options)

Description copied from class: PTransform

Called before running the Pipeline to verify this transform is fully and correctly specified.

By default, does nothing.

Overrides:

validate in class PTransform<PCollection<KV<K,V>>,PDone>

populateDisplayData

public void populateDisplayData(DisplayData.Builder builder)

Description copied from class: PTransform

Register display data for the given transform or component.

populateDisplayData(DisplayData.Builder) is invoked by Pipeline runners to collect display data via DisplayData.from(HasDisplayData). Implementations may call super.populateDisplayData(builder) in order to register display data in the current namespace, but should otherwise use subcomponent.populateDisplayData(builder) to use the namespace of the subcomponent.

By default, does not register any display data. Implementors may override this method to provide their own display data.

Specified by:

populateDisplayData in interface HasDisplayData

Overrides:

populateDisplayData in class PTransform<PCollection<KV<K,V>>,PDone>

Parameters:

builder - The builder to populate with display data.

See Also:

HasDisplayData