Apache Spark can work with multiple data sources that include object stores like Amazon S3, OpenStack Swift, IBM SoftLayer, and more. To access an object store, Apache Spark uses Hadoop modules that contain drivers to the various object stores.
Apache Spark needs only small set of object store functionalities. Specifically, Apache Spark requires object listing, objects creation, read objects, and getting data partitions. Hadoop drivers, however, must be compliant with the Hadoop eco system. This means they support many more operations, such as shell operations on directories, including move, copy, rename, etc. (these are not native object store operations). Moreover, Hadoop Map Reduce Client is designed to work with file systems and not object stores. The temp files and folders it uses for every write operation are renamed, copied, and deleted. This leads to dozens of useless requests targeted at the object store. It’s clear that Hadoop is designed to work with file systems and not object stores.
Stocator is implicitly designed for the object stores, it has very a different architecture from the existing Hadoop driver. It doesn’t depends on the Hadoop modules and interacts directly with object stores.
Stocator is a generic connector, that may contain various implementations for object stores. It initially provided with complete Swift driver, based on the JOSS package. Stocator can be very easily extended with more object store implementations, like support for Amazon S3.
-
Hadoop eco system compliant. Implements Hadoop FileSystem interface
-
No need to change or recompile Apache Spark
-
Doesn’t create any temporary folders or files for write operations. Each Spark's task generates only one object in the object store. Preserves existing Hadoop fault tolerance model.
-
Object's name may contain "/" and thus simulate directory structures
-
Containers / buckets are automatically created
-
(Swift Driver) Uses streaming for object uploads, without knowing object size. This is unique to Swift driver and removes the need to store object locally prior uploading it.
-
(Swift Driver) Tested to read / write large objects (over 50GB)
-
Supports Swiftauth, Keystone V2, Keystone V3 Password Scope Authentication
-
Tested to work with vanilla Swift cluster, SoftLayer object store, IBM Bluemix Object service
-
Supports speculate mode
-
(Swift Driver) Supports public containers. For example
sc.textFile("swift2d://dal05.objectstorage.softlayer.net/v1/AUTH_ID/CONT/data.csv")
Checkout the master branch https://github.com/SparkTC/stocator.git
-
Change directory to
stocator -
Execute
mvn install -
If you want to build a jar with all the dependecies, please execute
mvn clean package -Pall-in-one
Stocator allows to access Swift via unique schema swift2d://.
The configuration template located under conf/core-site.xml.template.
Stocator verifies that
mapreduce.fileoutputcommitter.marksuccessfuljobs=true
The default value of mapreduce.fileoutputcommitter.marksuccessfuljobs is true, therefore this key may not exists at all in the Apache Spark's configuration
Add the dependence to Stocator in conf/core-site.xml
<property>
<name>fs.swift2d.impl</name>
<value>com.ibm.stocator.fs.ObjectStoreFileSystem</value>
</property>
The following is the list of the configuration keys
| Key | Info | Default value |
|---|---|---|
| fs.swift2d.service.SERVICE_NAME.auth.url | Mandatory | |
| fs.swift2d.service.SERVICE_NAME.public | Optional. Values: true, false | false |
| fs.swift2d.service.SERVICE_NAME.tenant | Mandatory | |
| fs.swift2d.service.SERVICE_NAME.password | Mandatory | |
| fs.swift2d.service.SERVICE_NAME.username | Mandatory | |
| fs.swift2d.service.SERVICE_NAME.block.size | Block size in MB | 128MB |
| fs.swift2d.service.SERVICE_NAME.region | Mandatory for Keystone | |
| fs.swift2d.service.SERVICE_NAME.auth.method | Optional. Values: keystone, swiftauth, keystoneV3 | keystoneV3 |
Below is the internal Keystone V3 mapping
| Driver configuration key | Keystone V3 key |
|---|---|
| fs.swift2d.service.SERVICE_NAME.username | user id |
| fs.swift2d.service.SERVICE_NAME.tenant | project id |
<property>
<name>fs.swift2d.service.SERVICE_NAME.auth.url</name>
<value>http://IP:PORT/v2.0/tokens</value>
</property>
<property>
<name>fs.swift2d.service.SERVICE_NAME.public</name>
<value>true</value>
</property>
<property>
<name>fs.swift2d.service.SERVICE_NAME.tenant</name>
<value>TENANT</value>
</property>
<property>
<name>fs.swift2d.service.SERVICE_NAME.password</name>
<value>PASSWORD</value>
</property>
<property>
<name>fs.swift2d.service.SERVICE_NAME.username</name>
<value>USERNAME</value>
</property>
<property>
<name>fs.swift2d.service.SERVICE_NAME.auth.method</name>
<value>keystone</value>
</property>
<property>
<name>fs.swift2d.service.dal05.auth.url</name>
<value>https://dal05.objectstorage.softlayer.net/auth/v1.0/</value>
</property>
<property>
<name>fs.swift2d.service.dal05.public</name>
<value>true</value>
</property>
<property>
<name>fs.swift2d.service.dal05.tenant</name>
<value>TENANT</value>
</property>
<property>
<name>fs.swift2d.service.dal05.password</name>
<value>API KEY</value>
</property>
<property>
<name>fs.swift2d.service.dal05.username</name>
<value>USERNAME</value>
</property>
<property>
<name>fs.swift2d.service.dal05.auth.method</name>
<value>swiftauth</value>
</property>
In order to properly connect to an IBM Bluemix object store service, you need to open that service in the IBM Bluemix dashboard and inspect the service credentials and update the properties below with the correspondent values :
<property>
<name>fs.swift2d.service.SERVICE_NAME.auth.url</name>
<value>https://identity.open.softlayer.com/v3/auth/tokens</value>
</property>
<property>
<name>fs.swift2d.service.SERVICE_NAME.public</name>
<value>true</value>
</property>
<property>
<name>fs.swift2d.service.SERVICE_NAME.tenant</name>
<value>PROJECTID</value>
</property>
<property>
<name>fs.swift2d.service.SERVICE_NAME.password</name>
<value>PASSWORD</value>
</property>
<property>
<name>fs.swift2d.service.SERVICE_NAME.username</name>
<value>USERID</value>
</property>
<property>
<name>fs.swift2d.service.SERVICE_NAME.auth.method</name>
<value>keystoneV3</value>
</property>
<property>
<name>fs.swift2d.service.SERVICE_NAME.region</name>
<value>dallas</value>
</property>
Stocator uses Apache httpcomponents.httpclient.version version 4.5.2 to access object stores based on Swift API. Below is the optional configuration that can be provided to Stocator and used internally to configure HttpClient.
| Configuration key | Default | Info |
|---|---|---|
| fs.stocator.MaxPerRoute | 25 | maximal connections per IP route |
| fs.stocator.MaxTotal | 50 | maximal concurrent connections |
| fs.stocator.SoTimeout | 1000 | low level socket timeout in milliseconds |
| fs.stocator.executionCount | 100 | number of retries for certain HTTP issues |
| fs.stocator.ReqConnectTimeout | 5000 | Request level connect timeout. Determines the timeout in milliseconds until a connection is established |
| fs.stocator.ReqConnectionRequestTimeout | 5000 | Request level connection timeout. Returns the timeout in milliseconds used when requesting a connection from the connection manager |
| fs.stocator.ReqSocketTimeout | 5000 | Defines the socket timeout (SO_TIMEOUT) in milliseconds, which is the timeout for waiting for data or, put differently, a maximum period inactivity between two consecutive data packets). |
| fs.stocator.joss.synchronize.time | false | Will disable JOSS to synchronize time with the server. Setting this value to 'false' will badly impact on temp url. However this will reduce HEAD on account which might be problematic if the user doesn't has access rights to HEAD an account |
It's possible to provide configuration keys in run time, without keeping them in core-sites.xml. Just use SparkContext variable with
sc.hadoopConfiguration.set("KEY","VALUE")
It is possible to execute Apache Spark with the new driver, without compiling Apache Spark.
Directory stocator/target contains standalone jar stocator-1.0.6-jar-with-dependencies.jar.
Run Apache Spark with
./bin/spark-shell --jars stocator-1.0.6-jar-with-dependencies.jar
Both main pom.xml and core/pom.xml should be modified.
add to the <properties> of the main pom.xml
<stocator.version>1.0.6</stocator.version>
add stocator dependency to the main pom.xml
<dependency>
<groupId>com.ibm.stocator</groupId>
<artifactId>stocator</artifactId>
<version>${stocator.version}</version>
<scope>${hadoop.deps.scope}</scope>
</dependency>
modify core/pom.xml to include stocator
<dependency>
<groupId>com.ibm.stocator</groupId>
<artifactId>stocator</artifactId>
</dependency>
Compile Apache Spark with Haddop support
mvn -Phadoop-2.6 -Dhadoop.version=2.6.0 -DskipTests package
val data = Array(1, 2, 3, 4, 5, 6, 7, 8)
val distData = sc.parallelize(data)
distData.saveAsTextFile("swift2d://newcontainer.SERVICE_NAME/one1.txt")
Listing container newcontainer directly with a REST client will display
one1.txt
one1.txt/_SUCCESS
one1.txt/part-00000-taskid
one1.txt/part-00001-taskid
one1.txt/part-00002-taskid
one1.txt/part-00003-taskid
one1.txt/part-00004-taskid
one1.txt/part-00005-taskid
one1.txt/part-00006-taskid
one1.txt/part-00007-taskid
Follow
https://github.com/ehiggs/spark-terasort
You can run Terasort as follows (remember to change the word "SERVICE_NAME" with your Swift SERVICE_NAME name) Step 1:
export MAVEN_OPTS="-Xmx2g -XX:MaxPermSize=512M -XX:ReservedCodeCacheSize=512m"
Step 2:
./bin/spark-submit --driver-memory 2g --class com.github.ehiggs.spark.terasort.TeraGen /spark-terasort/target/spark-terasort-1.0-SNAPSHOT-jar-with-dependencies.jar 1g swift2d://teradata.SERVICE_NAME/terasort_in
Step 3:
./bin/spark-submit --driver-memory 2g --class com.github.ehiggs.spark.terasort.TeraSort /target/spark-terasort-1.0-SNAPSHOT-jar-with-dependencies.jar 1g swift2d://teradata.SERVICE_NAME/terasort_in swift2d://teradata.SERVICE_NAME/terasort_out
Step 4:
./bin/spark-submit --driver-memory 2g --class com.github.ehiggs.spark.terasort.TeraValidate /target/spark-terasort-1.0-SNAPSHOT-jar-with-dependencies.jar swift2d://teradata.SERVICE_NAME/terasort_out swift2d://teradata.SERVICE_NAME/terasort_validate
Copy
src/test/resources/core-site.xml.tempate to src/test/resources/core-site.xml
Edit src/test/resources/core-site.xml and configure Swift access details.
Functional tests will use container from fs.swift2d.test.uri. To use different container, change drivertest to different name. Container need not to be exists in advance and will be created automatically.
If you like to work on code, you can easily setup Eclipse project via
mvn eclipse:eclipse
and import it into Eclipse workspace.
Please follow the development guide, prior you submit pull requests.
To ease the debugging process, Please modify conf/log4j.properties and add
log4j.logger.com.ibm.stocator=ALL
We ask that you include a line similar to the following as part of your pull request comments: “DCO 1.1 Signed-off-by: Random J Developer“. “DCO” stands for “Developer Certificate of Origin,” and refers to the same text used in the Linux Kernel community. By adding this simple comment, you tell the community that you wrote the code you are contributing, or you have the right to pass on the code that you are contributing.
Join Stocator mailing list by sending email to [email protected].
Use [email protected] to post questions.
Please follow our wiki for more details. More information about Stocator can be find at
- Stocator on developerWorks Open
- Fast Lane for Connecting Object Stores to Apache Spark
- Exabytes, Elephants, Objects and Apache Spark
This research was supported by IOStack, an H2020 project of the EU Commission