Apache Pulsar Connector #

Flink provides an Apache Pulsar connector for reading and writing data from and to Pulsar topics with exactly-once guarantees.

Dependency #

You can use the connector with the Pulsar 2.8.1 or higher. Because the Pulsar connector supports Pulsar transactions, it is recommended to use the Pulsar 2.9.2 or higher. Details on Pulsar compatibility can be found in PIP-72.

<dependency>
    <groupId>org.apache.flink</groupId>
    <artifactId>flink-connector-pulsar</artifactId>
    <version>1.16.2</version>
</dependency>

In order to use the Pulsar connector in PyFlink jobs, the following dependencies are required:

Flink’s streaming connectors are not part of the binary distribution. See how to link with them for cluster execution here.

Pulsar Source #

This part describes the Pulsar source based on the new data source API.

Usage #

The Pulsar source provides a builder class for constructing a PulsarSource instance. The code snippet below builds a PulsarSource instance. It consumes messages from the earliest cursor of the topic “persistent://public/default/my-topic” in Exclusive subscription type (my-subscription) and deserializes the raw payload of the messages as strings.

PulsarSource<String> source = PulsarSource.builder()
    .setServiceUrl(serviceUrl)
    .setAdminUrl(adminUrl)
    .setStartCursor(StartCursor.earliest())
    .setTopics("my-topic")
    .setDeserializationSchema(PulsarDeserializationSchema.flinkSchema(new SimpleStringSchema()))
    .setSubscriptionName("my-subscription")
    .setSubscriptionType(SubscriptionType.Exclusive)
    .build();

env.fromSource(source, WatermarkStrategy.noWatermarks(), "Pulsar Source");

pulsar_source = PulsarSource.builder() \
    .set_service_url('pulsar://localhost:6650') \
    .set_admin_url('http://localhost:8080') \
    .set_start_cursor(StartCursor.earliest()) \
    .set_topics("my-topic") \
    .set_deserialization_schema(
        PulsarDeserializationSchema.flink_schema(SimpleStringSchema())) \
    .set_subscription_name('my-subscription') \
    .set_subscription_type(SubscriptionType.Exclusive) \
    .build()

env.from_source(source=pulsar_source,
                watermark_strategy=WatermarkStrategy.for_monotonous_timestamps(),
                source_name="pulsar source")

The following properties are required for building a PulsarSource:

• Pulsar service URL, configured by setServiceUrl(String)
• Pulsar service HTTP URL (also known as admin URL), configured by setAdminUrl(String)
• Pulsar subscription name, configured by setSubscriptionName(String)
• Topics / partitions to subscribe, see the following topic-partition subscription for more details.
• Deserializer to parse Pulsar messages, see the following deserializer for more details.

It is recommended to set the consumer name in Pulsar Source by setConsumerName(String). This sets a unique name for the Flink connector in the Pulsar statistic dashboard. You can use it to monitor the performance of your Flink connector and applications.

Topic-partition Subscription #

Pulsar source provide two ways of topic-partition subscription:

• Topic list, subscribing messages from all partitions in a list of topics. For example:

  Java

  PulsarSource.builder().setTopics("some-topic1", "some-topic2");
  
  // Partition 0 and 2 of topic "topic-a"
  PulsarSource.builder().setTopics("topic-a-partition-0", "topic-a-partition-2");
  

  Python

  PulsarSource.builder().set_topics(["some-topic1", "some-topic2"])
  
  # Partition 0 and 2 of topic "topic-a"
  PulsarSource.builder().set_topics(["topic-a-partition-0", "topic-a-partition-2"])
  

• Topic pattern, subscribing messages from all topics whose name matches the provided regular expression. For example:

  Java

  PulsarSource.builder().setTopicPattern("topic-*");
  

  Python

  PulsarSource.builder().set_topic_pattern("topic-*")
  

Flexible Topic Naming #

Since Pulsar 2.0, all topic names internally are in a form of {persistent|non-persistent}://tenant/namespace/topic. Now, for partitioned topics, you can use short names in many cases (for the sake of simplicity). The flexible naming system stems from the fact that there is now a default topic type, tenant, and namespace in a Pulsar cluster.

This table lists a mapping relationship between your input topic name and the translated topic name:

For non-persistent topics, you need to specify the entire topic name, as the default-based rules do not apply for non-partitioned topics. Thus, you cannot use a short name like non-persistent://my-topic and need to use non-persistent://public/default/my-topic instead.

Subscribing Pulsar Topic Partition #

Internally, Pulsar divides a partitioned topic as a set of non-partitioned topics according to the partition size.

For example, if a simple-string topic with 3 partitions is created under the sample tenant with the flink namespace. The topics on Pulsar would be:

You can directly consume messages from the topic partitions by using the non-partitioned topic names above. For example, use PulsarSource.builder().setTopics("sample/flink/simple-string-partition-1", "sample/flink/simple-string-partition-2") would consume the partitions 1 and 2 of the sample/flink/simple-string topic.

Setting Topic Patterns #

The Pulsar source extracts the topic type (persistent or non-persistent) from the provided topic pattern. For example, you can use the PulsarSource.builder().setTopicPattern("non-persistent://my-topic*") to specify a non-persistent topic. By default, a persistent topic is created if you do not specify the topic type in the regular expression.

You can use setTopicPattern("topic-*", RegexSubscriptionMode.AllTopics) to consume both persistent and non-persistent topics based on the topic pattern. The Pulsar source would filter the available topics by the RegexSubscriptionMode.

Deserializer #

A deserializer (PulsarDeserializationSchema) is for decoding Pulsar messages from bytes. You can configure the deserializer using setDeserializationSchema(PulsarDeserializationSchema). The PulsarDeserializationSchema defines how to deserialize a Pulsar Message<byte[]>.

If only the raw payload of a message (message data in bytes) is needed, you can use the predefined PulsarDeserializationSchema. Pulsar connector provides three implementation methods.

• Decode the message by using Pulsar’s Schema.

  // Primitive types
  PulsarDeserializationSchema.pulsarSchema(Schema);
  
  // Struct types (JSON, Protobuf, Avro, etc.)
  PulsarDeserializationSchema.pulsarSchema(Schema, Class);
  
  // KeyValue type
  PulsarDeserializationSchema.pulsarSchema(Schema, Class, Class);
  

• Decode the message by using Flink’s DeserializationSchema

  Java

  PulsarDeserializationSchema.flinkSchema(DeserializationSchema);
  

  Python

  PulsarDeserializationSchema.flink_schema(DeserializationSchema)
  

• Decode the message by using Flink’s TypeInformation

  Java

  PulsarDeserializationSchema.flinkTypeInfo(TypeInformation, ExecutionConfig);
  

  Python

  PulsarDeserializationSchema.flink_type_info(TypeInformation)
  

Pulsar Message<byte[]> contains some extra properties, such as message key, message publish time, message time, and application-defined key/value pairs etc. These properties could be defined in the Message<byte[]> interface.

If you want to deserialize the Pulsar message by these properties, you need to implement PulsarDeserializationSchema. Ensure that the TypeInformation from the PulsarDeserializationSchema.getProducedType() is correct. Flink uses this TypeInformation to pass the messages to downstream operators.

Pulsar Subscriptions #

A Pulsar subscription is a named configuration rule that determines how messages are delivered to Flink readers. The subscription name is required for consuming messages. Pulsar connector supports four subscription types:

• Exclusive
• Shared
• Failover
• Key_Shared

There is no difference between Exclusive and Failover in the Pulsar connector. When a Flink reader crashes, all (non-acknowledged and subsequent) messages are redelivered to the available Flink readers.

By default, if no subscription type is defined, Pulsar source uses the Shared subscription type.

// Shared subscription with name "my-shared"
PulsarSource.builder().setSubscriptionName("my-shared");

// Exclusive subscription with name "my-exclusive"
PulsarSource.builder().setSubscriptionName("my-exclusive").setSubscriptionType(SubscriptionType.Exclusive);

# Shared subscription with name "my-shared"
PulsarSource.builder().set_subscription_name("my-shared")

# Exclusive subscription with name "my-exclusive"
PulsarSource.builder().set_subscription_name("my-exclusive").set_subscription_type(SubscriptionType.Exclusive)

Key_Shared subscriptions #

All the Pulsar’s messages will be calculated with a key hash in Key_Shared subscription. The hash range must be 0 to 65535. We try to compute the key hash in the order of Message.getOrderingKey(), Message.getKey() or Message.getKeyBytes(). We will use "NO_KEY" str as the message key if none of these keys has been provided.

Pulsar’s Key_Shared subscription comes in two forms in Connector, the KeySharedMode.SPLIT and KeySharedMode.JOIN. Different KeySharedMode means different split assignment behaviors. If you only consume a subset of Pulsar’s key hash range, remember to use the KeySharedMode.JOIN which will subscribe all the range in only one reader. Otherwise, when the ranges can join into a full Pulsar key hash range (0~65535) you should use KeySharedMode.SPLIT mode for sharing the splits among all the backend readers.

In the KeySharedMode.SPLIT mode. The topic will be subscribed by multiple readers. But Pulsar has one limit in this situation. That is if a Message can’t find the corresponding reader by the key hash range. No messages will be delivered to the current readers, until there is a reader which can subscribe to such messages.

Define a RangeGenerator #

Ensure that you have provided a RangeGenerator implementation if you want to use the Key_Shared subscription type on the Pulsar connector. The RangeGenerator generates a set of key hash ranges so that a respective reader subtask only dispatches messages where the hash of the message key is contained in the specified range.

The Pulsar connector uses SplitRangeGenerator that divides the range by the Flink source parallelism if no RangeGenerator is provided in the Key_Shared subscription type.

Since the Pulsar didn’t expose the key hash range method. We have to provide an FixedKeysRangeGenerator for end-user. You can add the keys you want to consume, no need to calculate any hash ranges. The key’s hash isn’t specified to only one key, so the consuming results may contain the messages with different keys comparing the keys you have defined in this range generator. Remember to use flink’s DataStream.filter() method after the Pulsar source.

FixedKeysRangeGenerator.builder()
    .supportNullKey()
    .key("someKey")
    .keys(Arrays.asList("key1", "key2"))
    .build()

Starting Position #

The Pulsar source is able to consume messages starting from different positions by setting the setStartCursor(StartCursor) option. Built-in start cursors include:

• Start from the earliest available message in the topic.

  Java

  StartCursor.earliest();
  

  Python

  StartCursor.earliest()
  

• Start from the latest available message in the topic.

  Java

  StartCursor.latest();
  

  Python

  StartCursor.latest()
  

• Start from a specified message between the earliest and the latest. The Pulsar connector consumes from the latest available message if the message ID does not exist.

  The start message is included in consuming result.

  Java

  StartCursor.fromMessageId(MessageId);
  

  Python

  StartCursor.from_message_id(message_id)
  

• Start from a specified message between the earliest and the latest. The Pulsar connector consumes from the latest available message if the message ID doesn’t exist.

  Include or exclude the start message by using the second boolean parameter.

  Java

  StartCursor.fromMessageId(MessageId, boolean);
  

  Python

  StartCursor.from_message_id(message_id, boolean)
  

• Start from the specified message publish time by Message<byte[]>.getPublishTime(). This method is deprecated because the name is totally wrong which may cause confuse. You can use StartCursor.fromPublishTime(long) instead.

  Java

  StartCursor.fromMessageTime(long);
  

  Python

  StartCursor.from_message_time(int)
  

• Start from the specified message publish time by Message<byte[]>.getPublishTime().

  Java

  StartCursor.fromPublishTime(long);
  

  Python

  StartCursor.from_publish_time(int)
  

Each Pulsar message belongs to an ordered sequence on its topic. The sequence ID (MessageId) of the message is ordered in that sequence. The MessageId contains some extra information (the ledger, entry, partition) about how the message is stored, you can create a MessageId by using DefaultImplementation.newMessageId(long ledgerId, long entryId, int partitionIndex).

Boundedness #

The Pulsar source supports streaming and batch execution mode. By default, the PulsarSource is configured for unbounded data.

For unbounded data the Pulsar source never stops until a Flink job is stopped or failed. You can use the setUnboundedStopCursor(StopCursor) to set the Pulsar source to stop at a specific stop position.

You can use setBoundedStopCursor(StopCursor) to specify a stop position for bounded data.

Built-in stop cursors include:

• The Pulsar source never stops consuming messages.

  Java

  StopCursor.never();
  

  Python

  StopCursor.never()
  

• Stop at the latest available message when the Pulsar source starts consuming messages.

  Java

  StopCursor.latest();
  

  Python

  StopCursor.latest()
  

• Stop when the connector meets a given message, or stop at a message which is produced after this given message.

  Java

  StopCursor.atMessageId(MessageId);
  

  Python

  StopCursor.at_message_id(message_id)
  

• Stop but include the given message in the consuming result.

  Java

  StopCursor.afterMessageId(MessageId);
  

  Python

  StopCursor.after_message_id(message_id)
  

• Stop at the specified event time by Message<byte[]>.getEventTime(). The message with the given event time won’t be included in the consuming result.

  Java

  StopCursor.atEventTime(long);
  

  Python

  StopCursor.at_event_time(int)
  

• Stop after the specified event time by Message<byte[]>.getEventTime(). The message with the given event time will be included in the consuming result.

  Java

  StopCursor.afterEventTime(long);
  

  Python

  StopCursor.after_event_time(int)
  

• Stop at the specified publish time by Message<byte[]>.getPublishTime(). The message with the given publish time won’t be included in the consuming result.

  Java

  StopCursor.atPublishTime(long);
  

  Python

  StopCursor.at_publish_time(int)
  

• Stop after the specified publish time by Message<byte[]>.getPublishTime(). The message with the given publish time will be included in the consuming result.

  Java

  StopCursor.afterPublishTime(long);
  

  Python

  StopCursor.after_publish_time(int)
  

Source Configurable Options #

In addition to configuration options described above, you can set arbitrary options for PulsarClient, PulsarAdmin, Pulsar Consumer and PulsarSource by using setConfig(ConfigOption<T>, T), setConfig(Configuration) and setConfig(Properties).

PulsarClient Options #

The Pulsar connector uses the client API to create the Consumer instance. The Pulsar connector extracts most parts of Pulsar’s ClientConfigurationData, which is required for creating a PulsarClient, as Flink configuration options in PulsarOptions.

PulsarAdmin Options #

The admin API is used for querying topic metadata and for discovering the desired topics when the Pulsar connector uses topic-pattern subscription. It shares most part of the configuration options with the client API. The configuration options listed here are only used in the admin API. They are also defined in PulsarOptions.

Pulsar Consumer Options #

In general, Pulsar provides the Reader API and Consumer API for consuming messages in different scenarios. The Pulsar connector uses the Consumer API. It extracts most parts of Pulsar’s ConsumerConfigurationData as Flink configuration options in PulsarSourceOptions.

PulsarSource Options #

The configuration options below are mainly used for customizing the performance and message acknowledgement behavior. You can ignore them if you do not have any performance issues.

Dynamic Partition Discovery #

To handle scenarios like topic scaling-out or topic creation without restarting the Flink job, the Pulsar source periodically discover new partitions under a provided topic-partition subscription pattern. To enable partition discovery, you can set a non-negative value for the PulsarSourceOptions.PULSAR_PARTITION_DISCOVERY_INTERVAL_MS option:

// discover new partitions per 10 seconds
PulsarSource.builder()
    .setConfig(PulsarSourceOptions.PULSAR_PARTITION_DISCOVERY_INTERVAL_MS, 10000);

# discover new partitions per 10 seconds
PulsarSource.builder()
    .set_config("pulsar.source.partitionDiscoveryIntervalMs", 10000)

• Partition discovery is enabled by default. The Pulsar connector queries the topic metadata every 30 seconds.
• To disable partition discovery, you need to set a negative partition discovery interval.
• Partition discovery is disabled for bounded data even if you set this option with a non-negative value.

Event Time and Watermarks #

By default, the message uses the timestamp embedded in Pulsar Message<byte[]> as the event time. You can define your own WatermarkStrategy to extract the event time from the message, and emit the watermark downstream:

env.fromSource(pulsarSource, new CustomWatermarkStrategy(), "Pulsar Source With Custom Watermark Strategy");

env.from_source(pulsar_source, CustomWatermarkStrategy(), "Pulsar Source With Custom Watermark Strategy")

This documentation describes details about how to define a WatermarkStrategy.

Message Acknowledgement #

When a subscription is created, Pulsar retains all messages, even if the consumer is disconnected. The retained messages are discarded only when the connector acknowledges that all these messages are processed successfully. The Pulsar connector supports four subscription types, which makes the acknowledgement behaviors vary among different subscriptions.

Acknowledgement on Exclusive and Failover Subscription Types #

Exclusive and Failover subscription types support cumulative acknowledgment. In these subscription types, Flink only needs to acknowledge the latest successfully consumed message. All the message before the given message are marked with a consumed status.

The Pulsar source acknowledges the current consuming message when checkpoints are completed, to ensure the consistency between Flink’s checkpoint state and committed position on the Pulsar brokers.

If checkpointing is disabled, Pulsar source periodically acknowledges messages. You can use the PulsarSourceOptions.PULSAR_AUTO_COMMIT_CURSOR_INTERVAL option to set the acknowledgement period.

Pulsar source does NOT rely on committed positions for fault tolerance. Acknowledging messages is only for exposing the progress of consumers and monitoring on these two subscription types.

Acknowledgement on Shared and Key_Shared Subscription Types #

In Shared and Key_Shared subscription types, messages are acknowledged one by one. You can acknowledge a message in a transaction and commit it to Pulsar.

You should enable transaction in the Pulsar borker.conf file when using these two subscription types in connector:

transactionCoordinatorEnabled=true

The default timeout for Pulsar transactions is 3 hours. Make sure that that timeout is greater than checkpoint interval + maximum recovery time. A shorter checkpoint interval indicates a better consuming performance. You can use the PulsarSourceOptions.PULSAR_TRANSACTION_TIMEOUT_MILLIS option to change the transaction timeout.

If checkpointing is disabled or you can not enable the transaction on Pulsar broker, you should set PulsarSourceOptions.PULSAR_ENABLE_AUTO_ACKNOWLEDGE_MESSAGE to true. The message is immediately acknowledged after consuming. No consistency guarantees can be made in this scenario.

All acknowledgements in a transaction are recorded in the Pulsar broker side.

Pulsar Sink #

The Pulsar Sink supports writing records into one or more Pulsar topics or a specified list of Pulsar partitions.

This part describes the Pulsar sink based on the new data sink API.

If you still want to use the legacy SinkFunction or on Flink 1.14 or previous releases, just use the StreamNative’s pulsar-flink.

Usage #

The Pulsar Sink uses a builder class to construct the PulsarSink instance. This example writes a String record to a Pulsar topic with at-least-once delivery guarantee.

DataStream<String> stream = ...

PulsarSink<String> sink = PulsarSink.builder()
    .setServiceUrl(serviceUrl)
    .setAdminUrl(adminUrl)
    .setTopics("topic1")
    .setSerializationSchema(PulsarSerializationSchema.flinkSchema(new SimpleStringSchema()))
    .setDeliveryGuarantee(DeliveryGuarantee.AT_LEAST_ONCE)
    .build();

stream.sinkTo(sink);

stream = ...

pulsar_sink = PulsarSink.builder() \
    .set_service_url('pulsar://localhost:6650') \
    .set_admin_url('http://localhost:8080') \
    .set_topics("topic1") \
    .set_serialization_schema(PulsarSerializationSchema.flink_schema(SimpleStringSchema())) \
    .set_delivery_guarantee(DeliveryGuarantee.AT_LEAST_ONCE) \
    .build()

stream.sink_to(pulsar_sink)

The following properties are required for building PulsarSink:

• Pulsar service url, configured by setServiceUrl(String)
• Pulsar service http url (aka. admin url), configured by setAdminUrl(String)
• Topics / partitions to write, see writing targets for more details.
• Serializer to generate Pulsar messages, see serializer for more details.

It is recommended to set the producer name in Pulsar Source by setProducerName(String). This sets a unique name for the Flink connector in the Pulsar statistic dashboard. You can use it to monitor the performance of your Flink connector and applications.

Producing to topics #

Defining the topics for producing is similar to the topic-partition subscription in the Pulsar source. We support a mix-in style of topic setting. You can provide a list of topics, partitions, or both of them.

// Topic "some-topic1" and "some-topic2"
PulsarSink.builder().setTopics("some-topic1", "some-topic2")

// Partition 0 and 2 of topic "topic-a"
PulsarSink.builder().setTopics("topic-a-partition-0", "topic-a-partition-2")

// Partition 0 and 2 of topic "topic-a" and topic "some-topic2"
PulsarSink.builder().setTopics("topic-a-partition-0", "topic-a-partition-2", "some-topic2")

# Topic "some-topic1" and "some-topic2"
PulsarSink.builder().set_topics(["some-topic1", "some-topic2"])

# Partition 0 and 2 of topic "topic-a"
PulsarSink.builder().set_topics(["topic-a-partition-0", "topic-a-partition-2"])

# Partition 0 and 2 of topic "topic-a" and topic "some-topic2"
PulsarSink.builder().set_topics(["topic-a-partition-0", "topic-a-partition-2", "some-topic2"])

The topics you provide support auto partition discovery. We query the topic metadata from the Pulsar in a fixed interval. You can use the PulsarSinkOptions.PULSAR_TOPIC_METADATA_REFRESH_INTERVAL option to change the discovery interval option.

Configuring writing targets can be replaced by using a custom [TopicRouter] message routing. Configuring partitions on the Pulsar connector is explained in the flexible topic naming section.

If you build the Pulsar sink based on both the topic and its corresponding partitions, Pulsar sink merges them and only uses the topic.

For example, when using the PulsarSink.builder().setTopics("some-topic1", "some-topic1-partition-0") option to build the Pulsar sink, this is simplified to PulsarSink.builder().setTopics("some-topic1").

Serializer #

A serializer (PulsarSerializationSchema) is required for serializing the record instance into bytes. Similar to PulsarSource, Pulsar sink supports both Flink’s SerializationSchema and Pulsar’s Schema. Pulsar’s Schema.AUTO_PRODUCE_BYTES() is not supported in the Pulsar sink.

If you do not need the message key and other message properties in Pulsar’s Message interface, you can use the predefined PulsarSerializationSchema. The Pulsar sink provides two implementation methods.

• Encode the message by using Pulsar’s Schema.

  // Primitive types
  PulsarSerializationSchema.pulsarSchema(Schema)
  
  // Struct types (JSON, Protobuf, Avro, etc.)
  PulsarSerializationSchema.pulsarSchema(Schema, Class)
  
  // KeyValue type
  PulsarSerializationSchema.pulsarSchema(Schema, Class, Class)
  

• Encode the message by using Flink’s SerializationSchema

  Java

  PulsarSerializationSchema.flinkSchema(SerializationSchema)
  

  Python

  PulsarSerializationSchema.flink_schema(SimpleStringSchema())
  

Schema evolution can be enabled by users using PulsarSerializationSchema.pulsarSchema() and PulsarSinkBuilder.enableSchemaEvolution(). This means that any broker schema validation is in place.

Schema<SomePojo> schema = Schema.AVRO(SomePojo.class);
PulsarSerializationSchema<SomePojo> pulsarSchema = PulsarSerializationSchema.pulsarSchema(schema, SomePojo.class);

PulsarSink<String> sink = PulsarSink.builder()
    ...
    .setSerializationSchema(pulsarSchema)
    .enableSchemaEvolution()
    .build();

If you use Pulsar schema without enabling schema evolution, the target topic will have a Schema.BYTES schema. Consumers will need to handle the deserialization (if needed) themselves.

For example, if you set PulsarSerializationSchema.pulsarSchema(Schema.STRING) without enabling schema evolution, the schema stored in Pulsar topics is Schema.BYTES.

Message Routing #

Routing in Pulsar Sink is operated on the partition level. For a list of partitioned topics, the routing algorithm first collects all partitions from different topics, and then calculates routing within all the partitions. By default Pulsar Sink supports two router implementation.

• KeyHashTopicRouter: use the hashcode of the message’s key to decide the topic partition that messages are sent to.

  The message key is provided by PulsarSerializationSchema.key(IN, PulsarSinkContext) You need to implement this interface and extract the message key when you want to send the message with the same key to the same topic partition.

  If you do not provide the message key. A topic partition is randomly chosen from the topic list.

  The message key can be hashed in two ways: MessageKeyHash.JAVA_HASH and MessageKeyHash.MURMUR3_32_HASH. You can use the PulsarSinkOptions.PULSAR_MESSAGE_KEY_HASH option to choose the hash method.

• RoundRobinRouter: Round-robin among all the partitions.

  All messages are sent to the first partition, and switch to the next partition after sending a fixed number of messages. The batch size can be customized by the PulsarSinkOptions.PULSAR_BATCHING_MAX_MESSAGES option.

Let’s assume there are ten messages and two topics. Topic A has two partitions while topic B has three partitions. The batch size is set to five messages. In this case, topic A has 5 messages per partition which topic B does not receive any messages.

You can configure custom routers by using the TopicRouter interface. If you implement a TopicRouter, ensure that it is serializable. And you can return partitions which are not available in the pre-discovered partition list.

Thus, you do not need to specify topics using the PulsarSinkBuilder.setTopics option when you implement the custom topic router.

@PublicEvolving
public interface TopicRouter<IN> extends Serializable {

    String route(IN in, List<String> partitions, PulsarSinkContext context);

    default void open(SinkConfiguration sinkConfiguration) {
        // Nothing to do by default.
    }
}

Internally, a Pulsar partition is implemented as a topic. The Pulsar client provides APIs to hide this implementation detail and handles routing under the hood automatically. Pulsar Sink uses a lower client API to implement its own routing layer to support multiple topics routing.

For details, see partitioned topics.

Delivery Guarantee #

PulsarSink supports three delivery guarantee semantics.

• NONE: Data loss can happen even when the pipeline is running. Basically, we use a fire-and-forget strategy to send records to Pulsar topics in this mode. It means that this mode has the highest throughput.
• AT_LEAST_ONCE: No data loss happens, but data duplication can happen after a restart from checkpoint.
• EXACTLY_ONCE: No data loss happens. Each record is sent to the Pulsar broker only once. Pulsar Sink uses Pulsar transaction and two-phase commit (2PC) to ensure records are sent only once even after pipeline restarts.

Delayed message delivery #

Delayed message delivery enables you to delay the possibility to consume a message. With delayed message enabled, the Pulsar sink sends a message to the Pulsar topic immediately, but the message is delivered to a consumer once the specified delay is over.

Delayed message delivery only works in the Shared subscription type. In Exclusive and Failover subscription types, the delayed message is dispatched immediately.

You can configure the MessageDelayer to define when to send the message to the consumer. The default option is to never delay the message dispatching. You can use the MessageDelayer.fixed(Duration) option to Configure delaying all messages in a fixed duration. You can also implement the MessageDelayer interface to dispatch messages at different time.

The dispatch time should be calculated by the PulsarSinkContext.processTime().

Sink Configurable Options #

You can set options for PulsarClient, PulsarAdmin, Pulsar Producer and PulsarSink by using setConfig(ConfigOption<T>, T), setConfig(Configuration) and setConfig(Properties).

PulsarClient and PulsarAdmin Options #

For details, refer to PulsarAdmin options.

Pulsar Producer Options #

The Pulsar connector uses the Producer API to send messages. It extracts most parts of Pulsar’s ProducerConfigurationData as Flink configuration options in PulsarSinkOptions.

PulsarSink Options #

The configuration options below are mainly used for customizing the performance and message sending behavior. You can just leave them alone if you do not have any performance issues.

Sink Metrics #

This table lists supported metrics. The first 6 metrics are standard Pulsar Sink metrics as described in FLIP-33: Standardize Connector Metrics

• numBytesOut, numRecordsOut, numRecordsOutErrors are retrieved from Pulsar client metrics.

• currentSendTime tracks the time from when the producer calls sendAync() to the time when the message is acknowledged by the broker. This metric is not available in NONE delivery guarantee.

The Pulsar producer refreshes its stats every 60 seconds by default. The PulsarSink retrieves the Pulsar producer stats every 500ms. That means that numRecordsOut, numBytesOut, numAcksReceived, and numRecordsOutErrors are updated every 60 seconds. To increase the metrics refresh frequency, you can change the Pulsar producer stats refresh interval to a smaller value (minimum 1 second), as shown below.

builder.setConfig(PulsarOptions.PULSAR_STATS_INTERVAL_SECONDS, 1L);

builder.set_config("pulsar.client.statsIntervalSeconds", "1")

numBytesOutRate and numRecordsOutRate are calculated based on the numBytesOut and numRecordsOUt counter respectively. Flink internally uses a fixed 60 seconds window to calculate the rates.

Brief Design Rationale #

Pulsar sink follow the Sink API defined in FLIP-191.

Stateless SinkWriter #

In EXACTLY_ONCE mode, the Pulsar sink does not store transaction information in a checkpoint. That means that new transactions will be created after a restart. Therefore, any message in previous pending transactions is either aborted or timed out (They are never visible to the downstream Pulsar consumer). The Pulsar team is working to optimize the needed resources by unfinished pending transactions.

Pulsar Schema Evolution #

Pulsar Schema Evolution allows you to reuse the same Flink job after certain “allowed” data model changes, like adding or deleting a field in a AVRO-based Pojo class. Please note that you can specify Pulsar schema validation rules and define an auto schema update. For details, refer to Pulsar Schema Evolution.

Upgrading to the Latest Connector Version #

The generic upgrade steps are outlined in upgrading jobs and Flink versions guide. The Pulsar connector does not store any state on the Flink side. The Pulsar connector pushes and stores all the states on the Pulsar side. For Pulsar, you additionally need to know these limitations:

• Do not upgrade the Pulsar connector and Pulsar broker version at the same time.
• Always use a newer Pulsar client with Pulsar connector to consume messages from Pulsar.

Troubleshooting #

If you have a problem with Pulsar when using Flink, keep in mind that Flink only wraps PulsarClient or PulsarAdmin and your problem might be independent of Flink and sometimes can be solved by upgrading Pulsar brokers, reconfiguring Pulsar brokers or reconfiguring Pulsar connector in Flink.

Known Issues #

This section describes some known issues about the Pulsar connectors.

Unstable on Java 11 #

Pulsar connector has some known issues on Java 11. It is recommended to run Pulsar connector on Java 8.

No TransactionCoordinatorNotFound, but automatic reconnect #

Pulsar transactions are still in active development and are not stable. Pulsar 2.9.2 introduces a break change in transactions. If you use Pulsar 2.9.2 or higher with an older Pulsar client, you might get a TransactionCoordinatorNotFound exception.

You can use the latest pulsar-client-all release to resolve this issue.

Back to top