Unified data processing with Apache Pulsar and Apache Spark.
- Java 8 or later
- Spark 3.4.0 or later
- Pulsar 2.10.2 or later
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}}
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}}
...
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:
-
Local maven repository
-
Maven central repository
-
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.
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)]
For more information on how to use other language bindings for Spark Structured Streaming, see Structured Streaming Programming Guide.
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)]
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
, __eventTime
and __messageProperties
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)
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()
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()
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.
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` | A service HTTP URL of your Pulsar cluster | No | None | Streaming and Batch | The Pulsar `serviceHttpUrl` configuration. Only needed when `maxBytesPerTrigger` is specified |
`maxBytesPerTrigger` | A long value in unit of number of bytes | No | None | Streaming and Batch | A soft limit of the maximum number of bytes we want to process per microbatch. If this is specified, `admin.url` also needs to be specified. |
`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. |
`pollTimeoutMs` | 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:
|
No |
|
Streaming and batch queries |
Note:
|
`endingOffsets` | The following are valid values:
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 |
|
`failOnDataLoss` | The following are valid values: true or false | No | true | Streaming query |
This may cause a false alarm. You can set it to 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.admin.*` | Pulsar Admin configurations | No | None | Streaming and Batch | Admin configurations. Example: "pulsar.admin.tlsAllowInsecureConnection".
Please check Pulsar Admin 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 |
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)]
- 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 settingallowDifferentTopicSchemas
totrue
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 thevalue
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 > |
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)
If you want to build a Spark-Pulsar connector reading data from Pulsar and writing results to Pulsar, follow the steps below.
- Checkout the source code.
$ git clone https://github.com/streamnative/pulsar-spark.git
$ cd pulsar-spark
- Install Docker.
Pulsar-spark connector is using Testcontainers for integration tests. In order to run the integration tests, make sure you have installed Docker.
- Set a Scala version.
Change
scala.version
andscala.binary.version
inpom.xml
.Scala version should be consistent with the Scala version of Spark you use.
- 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.
- 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.