Spark Deployment Patterns
- Former user (Deleted)
The embedded Spark server in WSO2 EI Analytics Server can be used in several deployment modes, depending on your requirement.
| Mode | Description | When to use |
|---|---|---|
| Local (default) | In this mode, all of the Spark related work is done within a single node/JVM. | This is ideally suited for evaluation purposes and testing Spark queries in EI Analytics. |
| Cluster (recommended) | EI Analytics creates its own Spark cluster in the Carbon environment (using Hazelcast). This mode can be used with several high availability (HA) clustering patterns to handle failover scenarios. Additionally, in the Cluster mode, EI Analytics can be setup without a Spark application. This allows other components to use the EI Analytics cluster as an external Spark cluster. | For clustered production setups. |
| Client | In this mode, EI Analytics acts only as a Spark client pointing to a separate Spark master. | This is suited to scenarios where you want to submit EI analytics jobs to an external Spark cluster. |
The following topics list out the configuration instructions for the different deployment modes and also provide instructions on disabling Spark applications.
Local mode
This is the default mode for a typical EI Analytics instance. This mode enables users to evaluate Spark analytics in the EI Analytics Server. In this mode, a separate master or worker is not spawned. Instead, everything would run on a single JVM. Therefore, certain options like Spark Master UI and Spark Worker UI are not active.
Do the following to configure local mode.
- Ensure that Carbon clustering is disabled. To do this, open the
<EI-ANALYTICS_HOME>/repository/conf/axis2/axis2.xmlfile and setenable=”false”as shown below.
<clustering class="org.wso2.carbon.core.clustering.hazelcast.HazelcastClusteringAgent" enable="false"> - Set the Spark master to local. To do this, open the
<EI-ANALYTICS_HOME>/repository/conf/analytics/spark/spark-defaults.conffile and add the following entry (unless it already exists).
carbon.spark.master local[<number of cores>]
Cluster mode
Cluster mode is the recommended deployment pattern for EI Analytics in the production environment. Here, EI Analytics creates its own Spark cluster using the Carbon environment and Hazelcast. In this clustering approach, the Spark Standalone mode is used along with a custom implementation of the Standalone Recovery Mode API in Spark.
Since this mode uses a custom standalone recovery mode, the following configurations are passed into the server by default.
# Standalone Cluster Configs spark.deploy.recoveryMode CUSTOM spark.deploy.recoveryMode.factory org.wso2.carbon.analytics.spark.core.deploy.AnalyticsRecoveryModeFactory
- Enable Carbon clustering. To do this, in the
<EI-ANALYTICS_HOME>/repository/conf/axis2/axis2.xmlset enable=”true” for clustering as shown below.
<clustering class="org.wso2.carbon.core.clustering.hazelcast.HazelcastClusteringAgent" enable="true"> - In the
<EI-ANALYTICS_HOME>/repository/conf/analytics/spark/spark-defaults.conffile, set the number of masters in the cluster by adding the following entry.
carbon.spark.master.count <number of masters in the cluster>
While configuring this, make a note of the following.
- Ensure that the “
carbon.spark.master local” configuration remains unchanged. This acts as a flag to use Carbon clustering. - Each node can start as both a master and worker. So, in a two node cluster there would be two masters and two workers, one of the master nodes is active and the other is passive. You must specify the total number of masters in the cluster based on your requirement.
Client mode
Client mode is where EI Analytics submits all the Spark related jobs to an external Spark cluster. Since this uses an external Spark cluster, you must ensure that all the .jar files required by the Carbon Spark App are included in the SPARK_CLASSPATH of the Spark master and worker.
Do the following to configure client mode.
- In the
<EI-ANALYTICS_HOME>/repository/conf/analytics/spark/spark-defaults.conffile, add the following entry.
carbon.spark.master spark://<host1:port1, host2:port2, ...>
You can include a single master or a list of masters to ensure high-availability.
Disabling a Spark application
In addition to the above modes, you can also configure EI Analytics to startup without a Spark application. Up until the current Spark version (version 1.4.0), there can only be one active Sparkcontext inside a single JVM. Therefore, it is not possible to allow multiple Spark applications to be created in a single JVM. Furthermore, by default, applications submitted to the standalone mode cluster run in FIFO (first-in-first-out) order, and each application attempts to use all available nodes. The Carbon Spark application used for EI Analytics runs throughout the lifetime of the EI Analytics cluster. Therefore, even if you create a separate Spark application in a different JVM, it can only use the resources in the cluster when the Carbon Spark application is terminated.
In order to allow other clients to use the EI Analytics Spark cluster, there is an option provided to disable this Carbon Spark application. This can be done by setting a system variable in the server startup. See Disabling DAS components in the DAS documentation for more information.