Apache Spark
Last updated
Last updated
This guide describes how to configure to access Alluxio.
Applications using Spark 1.1 or later can access Alluxio through its HDFS-compatible interface. Using Alluxio as the data access layer, Spark applications can transparently access data in many different types of persistent storage services (e.g., AWS S3 buckets, Azure Object Store buckets, remote HDFS deployments and etc). Data can be actively fetched or transparently cached into Alluxio to speed up I/O performance especially when the Spark deployment is remote to the data. In addition, Alluxio can help simplify the architecture by decoupling compute and physical storage. When the data path in persistent under storage is hidden from Spark, changes to under storage can be independent from application logic; meanwhile, as a near-compute cache, Alluxio can still provide compute frameworks data-locality.
Java 8 Update 60 or higher (8u60+), 64-bit.
An Alluxio cluster is set up and is running. This guide assumes the persistent under storage is a local HDFS deployment. E.g., a line of alluxio.master.mount.table.root.ufs=hdfs://localhost:9000/alluxio/
is included in ${ALLUXIO_HOME}/conf/alluxio-site.properties
. Note that Alluxio supports many other under storage systems in addition to HDFS. To access data from any number of those systems is orthogonal to the focus of this guide but covered by .
Make sure that the Alluxio client jar is available. This Alluxio client jar file can be found at {{site.ALLUXIO_CLIENT_JAR_PATH}}
in the tarball distribution downloaded from Alluxio . Alternatively, advanced users can compile this client jar from the source code by following the .
The Alluxio client jar must be distributed across the all nodes where Spark drivers or executors are running. Place the client jar on the same local path (e.g. {{site.ALLUXIO_CLIENT_JAR_PATH}}
) on each node.
The Alluxio client jar must be in the classpath of all Spark drivers and executors in order for Spark applications to access Alluxio. Add the following line to spark/conf/spark-defaults.conf
on every node running Spark. Also, make sure the client jar is copied to every node running Spark.
This section shows how to use Alluxio as input and output sources for your Spark applications.
Copy local data to the Alluxio file system. Put the LICENSE
file into Alluxio, assuming you are in the Alluxio installation directory:
Run the following commands from spark-shell
, assuming the Alluxio Master is running on localhost
:
Alluxio transparently fetches data from the under storage system, given the exact path. For this section, HDFS is used as an example of a distributed under storage system.
Put a file Input_HDFS
into HDFS:
At this point, Alluxio does not know about this file since it was added to HDFS directly. Verify this by going to the web UI. Run the following commands from spark-shell
assuming Alluxio Master is running on localhost
:
When connecting to an HA-enabled Alluxio cluster using internal leader election, set the alluxio.master.rpc.addresses
property via the Java options in ${SPARK_HOME}/conf/spark-defaults.conf
so Spark applications know which Alluxio masters to connect to and how to identify the leader. For example:
Alternatively, add the property to the Hadoop configuration file ${SPARK_HOME}/conf/core-site.xml
:
Spark users can use pass JVM system properties to set Alluxio properties on to Spark jobs by adding "-Dproperty=value"
to spark.executor.extraJavaOptions
for Spark executors and spark.driver.extraJavaOptions
for Spark drivers. For example, to submit a Spark job with that uses the Alluxio CACHE_THROUGH
write type:
Alternatively, users may specify the HA authority directly in the URI without any configuration setup. For example, specify the master rpc addresses in the URI to connect to Alluxio configured for HA using internal leader election:
Storing RDDs in Alluxio memory is as simple as saving the RDD file to Alluxio. Two common ways to save RDDs as files in Alluxio are
saveAsTextFile
: writes the RDD as a text file, where each element is a line in the file,
saveAsObjectFile
: writes the RDD out to a file, by using Java serialization on each element.
The saved RDDs in Alluxio can be read again (from memory) by using sc.textFile
or sc.objectFile
respectively.
Storing Spark DataFrames in Alluxio memory is as simple as saving the DataFrame as a file to Alluxio. DataFrames are commonly written as parquet files, with df.write.parquet()
. After the parquet is written to Alluxio, it can be read from memory by using spark.read.parquet()
(or sqlContext.read.parquet()
for older versions of Spark).
Note: Alluxio workers use hostnames to represent network addresses to be consistent with HDFS. There is a workaround when launching Spark to achieve data locality. Users can explicitly specify hostnames by using the following script offered in Spark. Start the Spark Worker in each slave node with slave-hostname:
Note for older versions of Spark the script is called start-slave.sh
.
For example:
You can also set the SPARK_LOCAL_HOSTNAME
in $SPARK_HOME/conf/spark-env.sh
to achieve this. For example:
Either way, the Spark Worker addresses become hostnames and Locality Level becomes NODE_LOCAL
as shown in Spark WebUI below.
To maximize the amount of locality your Spark jobs attain, you should use as many executors as possible, hopefully at least one executor per node. It is recommended to co-locate Alluxio workers with the Spark executors.
When a Spark job is run on YARN, Spark launches its executors without taking data locality into account. Spark will then correctly take data locality into account when deciding how to distribute tasks to its executors.
For example, if host1
contains blockA
and a job using blockA
is launched on the YARN cluster with --num-executors=1
, Spark might place the only executor on host2
and have poor locality. However, if --num-executors=2
and executors are started on host1
and host2
, Spark will be smart enough to prioritize placing the job on host1
.
Class alluxio.hadoop.FileSystem not found
Issues with SparkSQL and Hive Metastorejava.io.IOException: No FileSystem for scheme: alluxio
Issue with Spark on YARNIf you use Spark on YARN with Alluxio and run into the exception java.io.IOException: No FileSystem for scheme: alluxio
, please add the following content to ${SPARK_HOME}/conf/core-site.xml
:
You may also open your browser and check . There should be an output directory /Output
which contains the doubled content of the input file Input
.
Open your browser and check . There should be an output directory Output_HDFS
which contains the doubled content of the input file Input_HDFS
. Also, the input file Input_HDFS
now will be 100% loaded in the Alluxio file system space.
Users can also configure Spark to connect to an Alluxio HA cluster using Zookeeper-based leader election. Refer to .
To customize Alluxio client-side properties for a Spark job, see .
Note that in client mode you need to set --driver-java-options "-Dalluxio.user.file.writetype.default=CACHE_THROUGH"
instead of --conf spark.driver.extraJavaOptions=-Dalluxio.user.file.writetype.default=CACHE_THROUGH
(see ).
If Spark configured using the instructions in , you can write URIs using the alluxio:///
scheme without specifying cluster information in the authority. This is because in HA mode, the address of leader Alluxio master will be served by the internal leader election or by the configured Zookeeper service.
Note that you must use semicolons rather than commas to separate different addresses to refer a URI of Alluxio in HA mode in Spark. Otherwise, the URI will be considered invalid by Spark. Please refer to the instructions in .
See the blog article "".
See the blog article "".
You may configure Spark's application logging for debugging purposes. The Spark documentation explains
If you are using YARN then there is a separate section which explains
If Spark task locality is ANY
while it should be NODE_LOCAL
, it is probably because Alluxio and Spark use different network address representations. One of them them may use hostname while another uses IP address. Refer to JIRA ticket for more details, where you can find solutions from the Spark community.
To run the spark-shell
with the Alluxio client, the Alluxio client jar will must be added to the classpath of the Spark driver and Spark executors, as . However, sometimes SparkSQL may fail to save tables to the Hive Metastore (location in Alluxio), with an error message similar to the following:
The recommended solution is to configure . In Spark 1.4.0 and later, Spark uses an isolated classloader to load java classes for accessing the Hive Metastore. The isolated classloader ignores certain packages and allows the main classloader to load "shared" classes (the Hadoop HDFS client is one of these "shared" classes). The Alluxio client should also be loaded by the main classloader, and you can append the alluxio
package to the configuration parameter spark.sql.hive.metastore.sharedPrefixes
to inform Spark to load Alluxio with the main classloader. For example, the parameter may be set in spark/conf/spark-defaults.conf
: