pulsar-spark

Unified data processing with Apache Pulsar and Apache Spark.

Prerequisites

• Java 8 or later
• Spark 3.2.2 or later
• Pulsar 2.10.2 or later

Preparations

Link

Client library

For Scala/Java applications using SBT/Maven project definitions, link your application with the following artifact:

    groupId = io.streamnative.connectors
    artifactId = pulsar-spark-connector_{{SCALA_BINARY_VERSION}}
    version = {{PULSAR_SPARK_VERSION}}

Deploy

Client library

As with any Spark applications, spark-submit is used to launch your application.
pulsar-spark-connector_{{SCALA_BINARY_VERSION}} and its dependencies can be directly added to spark-submit using --packages.

Example

$ ./bin/spark-submit 
  --packages io.streamnative.connectors:pulsar-spark-connector_{{SCALA_BINARY_VERSION}}:{{PULSAR_SPARK_VERSION}}
  ...

CLI

For experimenting on spark-shell (or pyspark for Python), you can also use --packages to add pulsar-spark-connector_{{SCALA_BINARY_VERSION}} and its dependencies directly.

Example

$ ./bin/spark-shell 
  --packages io.streamnative.connectors:pulsar-spark-connector_{{SCALA_BINARY_VERSION}}:{{PULSAR_SPARK_VERSION}}
  ...

When locating an artifact or library, --packages option checks the following repositories in order:

1. Local maven repository

2. Maven central repository

3. Other repositories specified by --repositories

The format for the coordinates should be groupId:artifactId:version.

For more information about submitting applications with external dependencies, see Application Submission Guide.

Usage

Read data from Pulsar

Create a Pulsar source for streaming queries

The following examples are in Scala.

// Subscribe to 1 topic
val df = spark
  .readStream
  .format("pulsar")
  .option("service.url", "pulsar://localhost:6650")
  .option("topic", "topic1")
  .load()
df.selectExpr("CAST(__key AS STRING)", "CAST(value AS STRING)")
  .as[(String, String)]

// Subscribe to multiple topics
val df = spark
  .readStream
  .format("pulsar")
  .option("service.url", "pulsar://localhost:6650")
  .option("topics", "topic1,topic2")
  .load()
df.selectExpr("CAST(__key AS STRING)", "CAST(value AS STRING)")
  .as[(String, String)]

// Subscribe to a topic pattern
val df = spark
  .readStream
  .format("pulsar")
  .option("service.url", "pulsar://localhost:6650")
  .option("topicsPattern", "topic.*")
  .load()
df.selectExpr("CAST(__key AS STRING)", "CAST(value AS STRING)")
  .as[(String, String)]

Tip

For more information on how to use other language bindings for Spark Structured Streaming, see Structured Streaming Programming Guide.

Create a Pulsar source for batch queries

If you have a use case that is better suited to batch processing, you can create a Dataset/DataFrame for a defined range of offsets.

The following examples are in Scala.

// Subscribe to 1 topic defaults to the earliest and latest offsets
val df = spark
  .read
  .format("pulsar")
  .option("service.url", "pulsar://localhost:6650")
  .option("topic", "topic1")
  .load()
df.selectExpr("CAST(__key AS STRING)", "CAST(value AS STRING)")
  .as[(String, String)]

// Subscribe to multiple topics, specifying explicit Pulsar offsets
import org.apache.spark.sql.pulsar.JsonUtils._
val startingOffsets = topicOffsets(Map("topic1" -> messageId1, "topic2" -> messageId2))
val endingOffsets = topicOffsets(...)
val df = spark
  .read
  .format("pulsar")
  .option("service.url", "pulsar://localhost:6650")
  .option("topics", "topic1,topic2")
  .option("startingOffsets", startingOffsets)
  .option("endingOffsets", endingOffsets)
  .load()
df.selectExpr("CAST(__key AS STRING)", "CAST(value AS STRING)")
  .as[(String, String)]

// Subscribe to a pattern, at the earliest and latest offsets
val df = spark
  .read
  .format("pulsar")
  .option("service.url", "pulsar://localhost:6650")
  .option("topicsPattern", "topic.*")
  .option("startingOffsets", "earliest")
  .option("endingOffsets", "latest")
  .load()
df.selectExpr("CAST(__key AS STRING)", "CAST(value AS STRING)")
  .as[(String, String)]

Write data to Pulsar

The DataFrame written to Pulsar can have arbitrary schema, since each record in DataFrame is transformed as one message sent to Pulsar, fields of DataFrame are divided into two groups: __key and __eventTime fields are encoded as metadata of Pulsar message; other fields are grouped and encoded using AVRO and put in value():

producer.newMessage().key(__key).value(avro_encoded_fields).eventTime(__eventTime)

Create a Pulsar sink for streaming queries

The following examples are in Scala.

// Write key-value data from a DataFrame to a specific Pulsar topic specified in an option
val ds = df
  .selectExpr("CAST(__key AS STRING)", "CAST(value AS STRING)")
  .writeStream
  .format("pulsar")
  .option("service.url", "pulsar://localhost:6650")
  .option("topic", "topic1")
  .start()

// Write key-value data from a DataFrame to Pulsar using a topic specified in the data
val ds = df
  .selectExpr("__topic", "CAST(__key AS STRING)", "CAST(value AS STRING)")
  .writeStream
  .format("pulsar")
  .option("service.url", "pulsar://localhost:6650")
  .start()

Write the output of batch queries to Pulsar

The following examples are in Scala.

// Write key-value data from a DataFrame to a specific Pulsar topic specified in an option
df.selectExpr("CAST(__key AS STRING)", "CAST(value AS STRING)")
  .write
  .format("pulsar")
  .option("service.url", "pulsar://localhost:6650")
  .option("topic", "topic1")
  .save()

// Write key-value data from a DataFrame to Pulsar using a topic specified in the data
df.selectExpr("__topic", "CAST(__key AS STRING)", "CAST(value AS STRING)")
  .write
  .format("pulsar")
  .option("service.url", "pulsar://localhost:6650")
  .save()

Limitations

Currently, we provide at-least-once semantic. Consequently, when writing either streaming queries or batch queries to Pulsar, some records may be duplicated. A possible solution to remove duplicates when reading the written data could be to introduce a primary (unique) key that can be used to perform de-duplication when reading.

Configurations

Option	Value	Required	Default	QueryType	Description
`service.url`	The Pulsar `serviceUrl` String	Yes	None	Streaming and Batch	The Pulsar `serviceUrl` configuration for Pulsar service. Example: "pulsar://localhost:6650".
`admin.url` (Deprecated)	A service HTTP URL of your Pulsar cluster	No	None	Streaming and Batch	The Pulsar `serviceHttpUrl` configuration.
`predefinedSubscription`	A Subscription name string	No	None	Streaming and Batch	The predefined subscription name used by the connector to track spark application progress.
`subscriptionPrefix`	A subscription prefix string	No	None	Streaming and Batch	A prefix used by the connector to generate a random subscription to track spark application progress.
`topic`	A topic name string	Yes	None	Streaming and Batch	The topic to be consumed. Only one of `topic`, `topics` or `topicsPattern` options can be specified for Pulsar source.
`topics`	A comma-separated list of topics	Yes	None	Streaming and Batch	The topic list to be consumed. Only one of `topic`, `topics` or `topicsPattern` options can be specified for Pulsar source.
`topicsPattern`	A Java regex string	Yes	None	Streaming and Batch	The pattern used to subscribe to topic(s). Only one of `topic`, `topics` or `topicsPattern` options can be specified for Pulsar source.
`poolTimeoutMs`	A number string in unit of milliseconds	No	"120000"	Streaming and Batch	The timeout for reading messages from Pulsar. Example: `6000`.
`waitingForNonExistedTopic`	The following are valid values: true or false	No	"false"	Streaming and Batch	Whether the connector should wait until the desired topics are created. By default, the connector will not wait for the topic
`startingOffsets`	The following are valid values: "earliest"(streaming and batch queries) "latest" (streaming query) A JSON string Example """ {"topic-1":[8,11,16,101,24,1,32,1],"topic-5":[8,15,16,105,24,5,32,5]} """	No	"earliest"（batch query) "latest"（streaming query)	Streaming and batch queries	startingOffsets option controls where a reader reads data from. "earliest": lacks a valid offset, the reader reads all the data in the partition, starting from the very beginning. "latest": lacks a valid offset, the reader reads from the newest records written after the reader starts running. A JSON string: specifies a starting offset for each Topic. You can use org.apache.spark.sql.pulsar.JsonUtils.topicOffsets(Map[String, MessageId]) to convert a message offset to a JSON string. Note: For batch query, "latest" is not allowed, either implicitly specified or use MessageId.latest ([8,-1,-1,-1,-1,-1,-1,-1,-1,127,16,-1,-1,-1,-1,-1,-1,-1,-1,127]) in JSON. For streaming query, "latest" only applies when a new query is started, and the resuming will always pick up from where the query left off. Newly discovered partitions during a query will start at "earliest".
`endingOffsets`	The following are valid values: "latest" (batch query) A JSON string Example {"topic-1":[8,12,16,102,24,2,32,2],"topic-5":[8,16,16,106,24,6,32,6]}	No	"latest"	Batch query	endingOffsets option controls where a reader stops reading data. "latest": the reader stops reading data at the latest record. A JSON string: specifies an ending offset for each topic. Note: MessageId.earliest ([8,-1,-1,-1,-1,-1,-1,-1,-1,-1,1,16,-1,-1,-1,-1,-1,-1,-1,-1,-1,1]) is not allowed.
`failOnDataLoss`	The following are valid values: true or false	No	true	Streaming query	failOnDataLoss option controls whether to fail a query when data is lost (for example, topics are deleted, or messages are deleted because of retention policy). This may cause a false alarm. You can set it to false when it doesn't work as you expected. A batch query always fails if it fails to read any data from the provided offsets due to data loss.
`allowDifferentTopicSchemas`	Boolean value	No	`false`	Streaming query	If multiple topics with different schemas are read, using this parameter automatic schema-based topic value deserialization can be turned off. In that way, topics with different schemas can be read in the same pipeline - which is then responsible for deserializing the raw values based on some schema. Since only the raw values are returned when this is `true`, Pulsar topic schema(s) are not taken into account during operation.
`pulsar.client.*`	Pulsar Client configurations	No	None	Streaming and Batch	Client configurations. Example: "pulsar.client.authPluginClassName". Please check Pulsar Client Configuration for more details
`pulsar.reader.*`	Pulsar Reader configurations	No	None	Streaming and Batch	Reader configurations. Example: "pulsar.reader.subscriptionName". Please check Pulsar Reader Configuration for more details
`pulsar.producer.*`	Pulsar Producer configurations	No	None	Streaming and Batch	Producer configurations. Example: "pulsar.producer.blockIfQueueFull". Please check Pulsar Producer Configuration for more details

Authentication

Should the Pulsar cluster require authentication, credentials can be set in the following way.

The following examples are in Scala.

// Secure connection with authentication, using the same credentials on the
// Pulsar client and admin interface (if not given explicitly, the client configuration
// is used for admin as well).
val df = spark
  .readStream
  .format("pulsar")
  .option("service.url", "pulsar://localhost:6650")
  .option("pulsar.client.authPluginClassName","org.apache.pulsar.client.impl.auth.AuthenticationToken")
  .option("pulsar.client.authParams","token:<valid client JWT token>")
  .option("topicsPattern", "sensitiveTopic")
  .load()
df.selectExpr("CAST(__key AS STRING)", "CAST(value AS STRING)")
  .as[(String, String)]

// Secure connection with client TLS enabled.
// Note that the certificate file has to be present at the specified
// path on every machine of the cluster!
val df = spark
  .readStream
  .format("pulsar")
  .option("service.url", "pulsar+ssl://localhost:6651")
  .option("pulsar.admin.authPluginClassName","org.apache.pulsar.client.impl.auth.AuthenticationToken")
  .option("pulsar.admin.authParams","token:<valid admin JWT token>")
  .option("pulsar.client.authPluginClassName","org.apache.pulsar.client.impl.auth.AuthenticationToken")
  .option("pulsar.client.authParams","token:<valid client JWT token>")
  .option("pulsar.client.tlsTrustCertsFilePath","/path/to/tls/cert/cert.pem")
  .option("pulsar.client.tlsAllowInsecureConnection","false")
  .option("pulsar.client.tlsHostnameVerificationenable","true")
  .option("topicsPattern", "sensitiveTopic")
  .load()
df.selectExpr("CAST(__key AS STRING)", "CAST(value AS STRING)")
  .as[(String, String)]

Schema of Pulsar source

• For topics without schema or with primitive schema in Pulsar, messages' payload is loaded to a value column with the corresponding type with Pulsar schema.
• For topics with Avro or JSON schema, their field names and field types are kept in the result rows.
• If the topicsPattern matches for topics which have different schemas, then setting allowDifferentTopicSchemas to true will allow the connector to read this content in a raw form. In this case it is the responsibility of the pipeline to apply the schema on this content, which is loaded to the value column.

Besides, each row in the source has the following metadata fields as well.

Column	Type
`__key`	Binary
`__topic`	String
`__messageId`	Binary
`__publishTime`	Timestamp
`__eventTime`	Timestamp
`__messageProperties`	Map < String, String >

Example

The topic of AVRO schema s in Pulsar is as below:

  case class Foo(i: Int, f: Float, bar: Bar)
  case class Bar(b: Boolean, s: String)
  val s = Schema.AVRO(Foo.getClass)

has the following schema as a DataFrame/DataSet in Spark:

root
 |-- i: integer (nullable = false)
 |-- f: float (nullable = false)
 |-- bar: struct (nullable = true)
 |    |-- b: boolean (nullable = false)
 |    |-- s: string (nullable = true)
 |-- __key: binary (nullable = true)
 |-- __topic: string (nullable = true)
 |-- __messageId: binary (nullable = true)
 |-- __publishTime: timestamp (nullable = true)
 |-- __messageProperties: map (nullable = true)
 |    |-- key: string
 |    |-- value: string (valueContainsNull = true)

For Pulsar topic with Schema.DOUBLE, it's schema as a DataFrame is:

root
|-- value: double (nullable = false)
|-- __key: binary (nullable = true)
|-- __topic: string (nullable = true)
|-- __messageId: binary (nullable = true)
|-- __publishTime: timestamp (nullable = true)
|-- __eventTime: timestamp (nullable = true)
|-- __messageProperties: map (nullable = true)
|    |-- key: string
|    |-- value: string (valueContainsNull = true)

Build Spark Pulsar Connector

If you want to build a Spark-Pulsar connector reading data from Pulsar and writing results to Pulsar, follow the steps below.

1. Checkout the source code.

$ git clone https://github.com/streamnative/pulsar-spark.git
$ cd pulsar-spark

2. Install Docker.

Pulsar-spark connector is using Testcontainers for integration tests. In order to run the integration tests, make sure you have installed Docker.

3. Set a Scala version.

Change scala.version and scala.binary.version in pom.xml.

Note

Scala version should be consistent with the Scala version of Spark you use.

4. Build the project.

$ mvn clean install -DskipTests

If you get the following error during compilation, try running Maven with Java 8:

[ERROR] [Error] : Source option 6 is no longer supported. Use 7 or later.
[ERROR] [Error] : Target option 6 is no longer supported. Use 7 or later.

5. Run the tests.

$ mvn clean install

Note: by configuring scalatest-maven-plugin in the usual ways, individual tests can be executed, if that is needed:

mvn -Dsuites=org.apache.spark.sql.pulsar.CachedPulsarClientSuite clean install

This might be handy if test execution is slower, or you get a java.io.IOException: Too many open files exception during full suite run.

Once the installation is finished, there is a fat jar generated under both local maven repo and target directory.

License