cloudera / livy   0.3.0

Livy is an open source REST interface for interacting with Apache Spark from anywhere. It supports executing snippets of code or programs in a Spark context that runs locally or in Apache Hadoop YARN.

• Interactive Scala, Python and R shells
• Batch submissions in Scala, Java, Python
• Multiple users can share the same server (impersonation support)
• Can be used for submitting jobs from anywhere with REST
• Does not require any code change to your programs

Pull requests are welcomed! But before you begin, please check out the Wiki.

To build Livy, you will need:

Debian/Ubuntu:

• mvn (from maven package or maven3 tarball)
• openjdk-7-jdk (or Oracle Java7 jdk)
• Python 2.6+
• R 3.x

Redhat/CentOS:

• mvn (from maven package or maven3 tarball)
• java-1.7.0-openjdk (or Oracle Java7 jdk)
• Python 2.6+
• R 3.x

MacOS:

• Xcode command line tools
• Oracle's JDK 1.7+
• Maven (Homebrew)
• Python 2.6+
• R 3.x

Required python packages for building Livy:

• cloudpickle
• requests
• requests-kerberos
• flake8
• flaky
• pytest

To run Livy, you will also need a Spark installation. You can get Spark releases at https://spark.apache.org/downloads.html.

Livy requires at least Spark 1.6 and supports both Scala 2.10 and 2.11 builds of Spark, Livy will automatically pick repl dependencies through detecting the Scala version of Spark.

Livy also supports Spark 2.0+ for both interactive and batch submission, you could seamlessly switch to different versions of Spark through SPARK_HOME configuration, without needing to rebuild Livy.

Livy is built using Apache Maven. To check out and build Livy, run:

git clone https://github.com/cloudera/livy.git
cd livy
mvn package

By default Livy is built against Apache Spark 1.6.2, but the version of Spark used when running Livy does not need to match the version used to build Livy. Livy internally uses reflection to mitigate the gaps between different Spark versions, also Livy package itself does not contain a Spark distribution, so it will work with any supported version of Spark (Spark 1.6+) without needing to rebuild against specific version of Spark.

In order to run Livy with local sessions, first export these variables:

export SPARK_HOME=/usr/lib/spark
export HADOOP_CONF_DIR=/etc/hadoop/conf

Then start the server with:

./bin/livy-server

Livy uses the Spark configuration under SPARK_HOME by default. You can override the Spark configuration by setting the SPARK_CONF_DIR environment variable before starting Livy.

It is strongly recommended to configure Spark to submit applications in YARN cluster mode. That makes sure that user sessions have their resources properly accounted for in the YARN cluster, and that the host running the Livy server doesn't become overloaded when multiple user sessions are running.

Livy uses a few configuration files under configuration the directory, which by default is the conf directory under the Livy installation. An alternative configuration directory can be provided by setting the LIVY_CONF_DIR environment variable when starting Livy.

The configuration files used by Livy are:

• livy.conf: contains the server configuration. The Livy distribution ships with a default configuration file listing available configuration keys and their default values.
• spark-blacklist.conf: list Spark configuration options that users are not allowed to override. These options will be restricted to either their default values, or the values set in the Spark configuration used by Livy.
• log4j.properties: configuration for Livy logging. Defines log levels and where log messages will be written to. The default configuration will print log messages to stderr.

A few things changed between since Livy 0.1 that require manual intervention when upgrading.

• Sessions that were active when the Livy 0.1 server was stopped may need to be killed manually. Use the tools from your cluster manager to achieve that (for example, the yarn command line tool).
• The configuration file has been renamed from livy-defaults.conf to livy.conf.
• A few configuration values do not have any effect anymore. Notably:
  • livy.server.session.factory: this config option has been replaced by the Spark configuration under SPARK_HOME. If you wish to use a different Spark configuration for Livy, you can set SPARK_CONF_DIR in Livy's environment. To define the default file system root for sessions, set HADOOP_CONF_DIR to point at the Hadoop configuration to use. The default Hadoop file system will be used.
  • livy.yarn.jar: this config has been replaced by separate configs listing specific archives for different Livy features. Refer to the default livy.conf file shipped with Livy for instructions.
  • livy.server.spark-submit: replaced by the SPARK_HOME environment variable.

Livy provides a programmatic Java/Scala and Python API that allows applications to run code inside Spark without having to maintain a local Spark context. Here shows how to use the Java API.

Add the Cloudera repository to your application's POM:

<repositories>
  <repository>
    <id>cloudera.repo</id>
    <url>https://repository.cloudera.com/artifactory/cloudera-repos</url>
    <name>Cloudera Repositories</name>
    <snapshots>
      <enabled>false</enabled>
    </snapshots>
  </repository>
</repositories>

And add the Livy client dependency:

<dependency>
  <groupId>com.cloudera.livy</groupId>
  <artifactId>livy-client-http</artifactId>
  <version>0.2.0</version>
</dependency>

To be able to compile code that uses Spark APIs, also add the correspondent Spark dependencies.

To run Spark jobs within your applications, extend com.cloudera.livy.Job and implement the functionality you need. Here's an example job that calculates an approximate value for Pi:

import java.util.*;

import org.apache.spark.api.java.*;
import org.apache.spark.api.java.function.*;

import com.cloudera.livy.*;

public class PiJob implements Job<Double>, Function<Integer, Integer>,
  Function2<Integer, Integer, Integer> {

  private final int samples;

  public PiJob(int samples) {
    this.samples = samples;
  }

  @Override
  public Double call(JobContext ctx) throws Exception {
    List<Integer> sampleList = new ArrayList<Integer>();
    for (int i = 0; i < samples; i++) {
      sampleList.add(i + 1);
    }

    return 4.0d * ctx.sc().parallelize(sampleList).map(this).reduce(this) / samples;
  }

  @Override
  public Integer call(Integer v1) {
    double x = Math.random();
    double y = Math.random();
    return (x*x + y*y < 1) ? 1 : 0;
  }

  @Override
  public Integer call(Integer v1, Integer v2) {
    return v1 + v2;
  }

}

To submit this code using Livy, create a LivyClient instance and upload your application code to the Spark context. Here's an example of code that submits the above job and prints the computed value:

LivyClient client = new LivyClientBuilder()
  .setURI(new URI(livyUrl))
  .build();

try {
  System.err.printf("Uploading %s to the Spark context...\n", piJar);
  client.uploadJar(new File(piJar)).get();

  System.err.printf("Running PiJob with %d samples...\n", samples);
  double pi = client.submit(new PiJob(samples)).get();

  System.out.println("Pi is roughly: " + pi);
} finally {
  client.stop(true);
}

To learn about all the functionality available to applications, read the javadoc documentation for the classes under the api module.

Here's a step-by-step example of interacting with Livy in Python with the Requests library. By default Livy runs on port 8998 (which can be changed with the livy.server.port config option). We’ll start off with a Spark session that takes Scala code:

sudo pip install requests

import json, pprint, requests, textwrap
host = 'http://localhost:8998'
data = {'kind': 'spark'}
headers = {'Content-Type': 'application/json'}
r = requests.post(host + '/sessions', data=json.dumps(data), headers=headers)
r.json()

{u'state': u'starting', u'id': 0, u'kind': u'spark'}

Once the session has completed starting up, it transitions to the idle state:

session_url = host + r.headers['location']
r = requests.get(session_url, headers=headers)
r.json()

{u'state': u'idle', u'id': 0, u'kind': u'spark'}

Now we can execute Scala by passing in a simple JSON command:

statements_url = session_url + '/statements'
data = {'code': '1 + 1'}
r = requests.post(statements_url, data=json.dumps(data), headers=headers)
r.json()

{u'output': None, u'state': u'running', u'id': 0}

If a statement takes longer than a few milliseconds to execute, Livy returns early and provides a statement URL that can be polled until it is complete:

statement_url = host + r.headers['location']
r = requests.get(statement_url, headers=headers)
pprint.pprint(r.json())

{u'id': 0,
  u'output': {u'data': {u'text/plain': u'res0: Int = 2'},
              u'execution_count': 0,
              u'status': u'ok'},
  u'state': u'available'}

That was a pretty simple example. More interesting is using Spark to estimate Pi. This is from the Spark Examples:

data = {
  'code': textwrap.dedent("""
    val NUM_SAMPLES = 100000;
    val count = sc.parallelize(1 to NUM_SAMPLES).map { i =>
      val x = Math.random();
      val y = Math.random();
      if (x*x + y*y < 1) 1 else 0
    }.reduce(_ + _);
    println(\"Pi is roughly \" + 4.0 * count / NUM_SAMPLES)
    """)
}

r = requests.post(statements_url, data=json.dumps(data), headers=headers)
pprint.pprint(r.json())

statement_url = host + r.headers['location']
r = requests.get(statement_url, headers=headers)
pprint.pprint(r.json())

{u'id': 1,
 u'output': {u'data': {u'text/plain': u'Pi is roughly 3.14004\nNUM_SAMPLES: Int = 100000\ncount: Int = 78501'},
             u'execution_count': 1,
             u'status': u'ok'},
 u'state': u'available'}

Finally, close the session:

session_url = 'http://localhost:8998/sessions/0'
requests.delete(session_url, headers=headers)

<Response [204]>

PySpark has the same API, just with a different initial request:

data = {'kind': 'pyspark'}
r = requests.post(host + '/sessions', data=json.dumps(data), headers=headers)
r.json()

{u'id': 1, u'state': u'idle'}

The Pi example from before then can be run as:

data = {
  'code': textwrap.dedent("""
    import random
    NUM_SAMPLES = 100000
    def sample(p):
      x, y = random.random(), random.random()
      return 1 if x*x + y*y < 1 else 0

    count = sc.parallelize(xrange(0, NUM_SAMPLES)).map(sample).reduce(lambda a, b: a + b)
    print "Pi is roughly %f" % (4.0 * count / NUM_SAMPLES)
    """)
}

r = requests.post(statements_url, data=json.dumps(data), headers=headers)
pprint.pprint(r.json())

{u'id': 12,
u'output': {u'data': {u'text/plain': u'Pi is roughly 3.136000'},
            u'execution_count': 12,
            u'status': u'ok'},
u'state': u'running'}

SparkR has the same API:

data = {'kind': 'sparkr'}
r = requests.post(host + '/sessions', data=json.dumps(data), headers=headers)
r.json()

{u'id': 1, u'state': u'idle'}

The Pi example from before then can be run as:

data = {
  'code': textwrap.dedent("""
    n <- 100000
    piFunc <- function(elem) {
      rands <- runif(n = 2, min = -1, max = 1)
      val <- ifelse((rands[1]^2 + rands[2]^2) < 1, 1.0, 0.0)
      val
    }
    piFuncVec <- function(elems) {
      message(length(elems))
      rands1 <- runif(n = length(elems), min = -1, max = 1)
      rands2 <- runif(n = length(elems), min = -1, max = 1)
      val <- ifelse((rands1^2 + rands2^2) < 1, 1.0, 0.0)
      sum(val)
    }
    rdd <- parallelize(sc, 1:n, slices)
    count <- reduce(lapplyPartition(rdd, piFuncVec), sum)
    cat("Pi is roughly", 4.0 * count / n, "\n")
    """)
}

r = requests.post(statements_url, data=json.dumps(data), headers=headers)
pprint.pprint(r.json())

{u'id': 12,
 u'output': {u'data': {u'text/plain': u'Pi is roughly 3.136000'},
             u'execution_count': 12,
             u'status': u'ok'},
 u'state': u'running'}

• User group: http://groups.google.com/a/cloudera.org/group/livy-user

• Dev group: http://groups.google.com/a/cloudera.org/group/livy-dev

• Dev slack: https://livy-dev.slack.com.

  To join: http://livy-slack-invite.azurewebsites.net. Invite token: I'm not a bot.

• JIRA: https://issues.cloudera.org/browse/LIVY

• Pull requests: https://github.com/cloudera/livy/pulls

Returns all the active interactive sessions.

Creates a new interactive Scala, Python, or R shell in the cluster.

The created Session.

Returns the session information.

The Session.

Returns the state of session

Kills the Session job.

Gets the log lines from this session.

Returns all the statements in a session.

Runs a statement in a session.

The statement object.

Returns a specified statement in a session.

The statement object.

Cancel the specified statement in this session.

Returns all the active batch sessions.

The created Batch object.

Returns the batch session information.

The Batch.

Returns the state of batch session

Kills the Batch job.

Gets the log lines from this batch.

A session represents an interactive shell.

To change the Python executable the session uses, Livy reads the path from environment variable PYSPARK_PYTHON (Same as pyspark).

Like pyspark, if Livy is running in local mode, just set the environment variable. If the session is running in yarn-cluster mode, please set spark.yarn.appMasterEnv.PYSPARK_PYTHON in SparkConf so the environment variable is passed to the driver.

To change the Python executable the session uses, Livy reads the path from environment variable PYSPARK3_PYTHON.

Like pyspark, if Livy is running in local mode, just set the environment variable. If the session is running in yarn-cluster mode, please set spark.yarn.appMasterEnv.PYSPARK3_PYTHON in SparkConf so the environment variable is passed to the driver.

A statement represents the result of an execution statement.

Apache License, Version 2.0 http://www.apache.org/licenses/LICENSE-2.0