Spark Jobs API

This is a set of endpoints for configuring and running Spark jobs.

Spark job subtypes

For the Spark job type, the available subtypes are listed below.

ALS Recommender

Train a collaborative filtering matrix decomposition recommender using SparkML’s Alternating Least Squares (ALS) to batch-compute user recommendations and item similarities.

Aggregation

Define an aggregation job to be executed by Fusion Spark.

Co-occurrence Similarity

Compute a mutual-information item similarity model.

Random Forest Classifier Training

Train a random forest classifier for text classification.

Script

Run a custom Scala script as a Fusion Job.

Matrix Decomposition-Based Query-Query Similarity Job

Train a collaborative filtering matrix decomposition recommender using SparkML’s Alternating Least Squares (ALS) to batch-compute query-query similarities.

Bisecting KMeans Clustering Job

Train a bisecting KMeans clustering model.

Logistic Regression Classifier Training Job

Trains a regularized logistic regression model for text classification.

Item Similarity Recommender

Compute user recommendations based on pre-computed item similarity model.

Spark Configuration Properties

Fusion passes all configuration properties with prefix "spark." to the Spark master, Spark worker and each Spark application, both for aggregation jobs and custom-scripted processing.

These properties are stored in Fusion’s ZooKeeper and can be updated via requests to Fusion endpoint api/apollo/configurations which will update the stored value without restarting the service, therefore existing jobs and SparkContexts will not be affected. The Fusion endpoint api/apollo/configurations returns all configured properties for that installation. You can examine spark default configurations in a Unix shell using the utilities curl and grep. Here is an example which checks a local Fusion installation running on port 8764:

curl -u username:password http://localhost:8764/api/apollo/configurations | grep '"spark.'

  "spark.executor.memory" : "2g",
  "spark.task.maxFailures" : "10",
  "spark.worker.cleanup.appDataTtl" : "7200",
  "spark.worker.cleanup.enabled" : "true",
  "spark.worker.memory" : "2g",

The default SparkContext that Fusion uses for aggregation jobs can be assigned a fraction of cluster resources (executor memory and/or available CPU cores). This allows other applications (such as scripted jobs, or shell sessions) to use the remaining cluster resources even when some aggregation jobs are running. Fusion 2.3 also permits dynamic allocation for all applications. This can be overriden per application. In practice, this means that even when there’s an already running SparkContext with a relatively long idle time (eg. 10 minutes) but there are no active jobs that use it, its resources (CPU cores and executor memory) will be released for use by other applications.

For scripted Spark jobs, users can specify per-job configuration overrides as a set of key / value pairs in a "sparkConfig" property element of a script job configuration, which takes precedence over values stored in ZooKeeper. The following is an example of a scripted job with a "sparkConfig" section:

{
  "id": "scripted_job_example",
  "script": "val rdd = sc.textFile(\"/foo.txt\")\nrdd.count\n",
  "sparkConfig": {
    "spark.cores.max": 2,
    "spark.executor.memory": "1g"
  }
}

The following table lists those Spark configuration properties that Fusion overrides or uses in order to determine applications' resource allocations.

Property Description

spark.master.url

By default, left unset. This property is only specified when using an external Spark cluster; when Fusion is using its own standalone Spark cluster, this property isn’t set.

spark.cores.max

The maximum number of cores across the cluster assigned to the application. If not specified, there is no limit. The default is unset, i.e., an unlimited number of cores.

spark.executor.memory

Amount of memory assigned to each application’s executor. The default is 2G.

spark.scheduler.mode

Controls how tasks are assigned to available resources. Can be either 'FIFO' or 'FAIR'. Default value is 'FAIR'.

spark.dynamicAllocation.enabled

Boolean - whether or not to enable dynamic allocation of executors. Default value is 'TRUE'.

spark.shuffle.service.enabled

Boolean - whether or not to enable internal shuffle service for standalone Spark cluster. Default value is 'TRUE'.

spark.dynamicAllocation.executorIdleTimeout

Number of seconds after which idle executors are removed. Default value is '60s'.

spark.dynamicAllocation.minExecutors

Number of executors to leave running even when idle. Default value is 0.

spark.eventLog.enabled

Boolean - whether or not event log is enabled. Event log stores job details and can be accessed after application finishes. Default value is 'TRUE'.

spark.eventLog.dir

Directory which stores event logs. Default location is $FUSION_HOME/var/spark-eventlog.

spark.eventLog.compress

Boolean - whether or not to compress event log data. Default value is 'TRUE'.

spark.logConf

Boolean - whether or not to log effective SparkConf of new SparkContext-s. Default value is 'TRUE'.

spark.deploy.recoveryMode

Default value is 'ZOOKEEPER'

spark.deploy.zookeeper.url

ZooKeeper connect string. Default value is $FUSION_ZK

spark.deploy.zookeeper.dir

ZooKeeper path, default value is /lucid/spark

spark.worker.cleanup.enabled

Boolean - whether or not to periodically cleanup worker data. Default value is 'TRUE'.

spark.worker.cleanup.appDataTtl

Time-to-live in seconds. Default value is 86400 (24h).

spark.deploy.retainedApplications

The maximum number of applications to show in the UI. Default value is 50.

spark.deploy.retainedDrivers

The maximum number of drivers. Default value is 50.

spark.worker.timeout

The maximum timeout in seconds allowed before a worker is considered lost. The default value is 30.

spark.worker.memory

The maximum total heap allocated to all executors running on this worker. Defaults to value of the executor memory heap.

Fusion Configuration Properties

Property Description

fusion.spark.master.port

Spark master job submission port. Default value is 8766.

fusion.spark.master.ui.port

Spark master UI port. Default value is 8767.

fusion.spark.idleTime

Maximum idle time in seconds, after which the application (ie. SparkContext) is shut down. Default value is 300.

fusion.spark.executor.memory.min

Minimum executor memory in MB. Default value 450Mb, which is sufficient to let Fusion components in application task’s to initialize themselves

fusion.spark.executor.memory.fraction

A float number in range (0.0, 1.0] indicating what portion of spark.executor.memory to allocate to this application. Default value is 1.0.

fusion.spark.cores.fraction

A float number in range (0.0, 1.0] indicating what portion of spark.cores.max to allocate to this application. Default value is 1.0.