[EDP] Add an engine for a Spark standalone deployment¶
The Spark provisioning plugin allows the creation of Spark standalone clusters in Sahara. Sahara EDP should support running Spark applications on those clusters.
The Spark plugin uses the standalone deployment mode for Spark (as opposed to Spark on YARN or Mesos). An EDP implementation must be created that supports the basic EDP functions using the facilities provided by the operating system and the Spark standalone deployment.
The basic EDP functions are:
The Sahara job manager has recently been refactored to allow provisioning plugins to select or provide an EDP engine based on the cluster and job_type. The plugin may return a custom EDP job engine object or choose a default engine provided by Sahara.
A default job engine for Spark standalone clusters can be added to Sahara that implements the basic EDP functions.
Note, there are no public APIs in Spark for job status or cancellation beyond facilities that might be available through a SparkContext object instantiated in a Scala program. However, it is possible to provide basic functionality without developing Scala programs.
Engine selection criteria
The Spark provisioning plugin must determine if the default Spark EDP engine may be used to run a particular job on a particular cluster. The following conditions must be true to use the engine
The default engine will use spark-submit for running jobs, therefore a cluster must have at least Spark version 1.0.0 to use the engine.
The job type must be Java (initially)
Remote commands via ssh
All operations should be implemented using ssh to run remote commands, as opposed to writing a custom agent and client. Furthermore, any long running commands should be run asynchronously.
The run_job() function will be implemented using the spark-submit script provided by Spark.
spark-submit must be run using client deploy mode
The spark-submit command will be executed via ssh. Ssh must return immediately and must return a process id (PID) that can be used for checking status or cancellation. This implies that the process is run in the background.
SIGINT must not be ignored by the process running spark-submit. Care needs to be taken here, since the default behavior of a process backgrounded from bash is to ignore SIGINT. (This can be handled by running spark-submit as a subprocess from a wrapper which first restores SIGINT, and launching the wrapper from ssh. In this case the wrapper must be sure to propagate SIGINT to the child).
spark-submit requires that the main application jar be in local storage on the node where it is run. Supporting jars may be in local storage or hdfs. For simplicity, all jars will be uploaded to local storage.
spark-submit should be run from a subdirectory on the master node in a well-known location. The subdirectory naming should incorporate the job name and job execution id from Sahara to make locating the directory easy. Program files, output, and logs should be written to this directory.
The exit status returned from spark-submit must be written to a file in the working directory.
stderr and stdout from spark-submit should be redirected and saved in the working directory. This will help debugging as well as preserve results for apps like SparkPi which write the result to stdout.
Sahara will allow the user to specify arguments to the Spark application. Any input and output data sources will be specified as path arguments to the Spark app.
Job status can be determined by monitoring the PID returned by run_job() via ps and reading the file containing the exit status
The initial job status is PENDING (the same for all Sahara jobs)
If the PID is present, the job status is RUNNING
If the PID is absent, check the exit status file - If the exit status is 0, the job status is SUCCEEDED - If the exit status is -2 or 130, the job status is KILLED (by SIGINT) - For any other exist status, the job status is DONEWITHERROR
If the job fails in Sahara (ie, because of an exception), the job status will be FAILED
A Spark application may be canceled by sending SIGINT to the process running spark-submit.
cancel_job() should send SIGINT to the PID returned by run_job(). If the PID is the process id of a wrapper, the wrapper must ensure that SIGINT is propagated to the child
If the command to send SIGINT is successful (ie, kill returns 0), cancel_job() should call get_job_status() to update the job status
The Ooyala job server is an alternative for implementing Spark EDP, but it’s a project of its own outside of OpenStack and introduces another dependency. It would have to be installed by the Spark provisioning plugin, and Sahara contributors would have to understand it thoroughly.
Other than Ooyala, there does not seem to be any existing client or API for handling job submission, monitoring, and cancellation.
Data model impact¶
There is no data model impact, but a few fields will be reused.
The oozie_job_id will store an id that allows the running application to be operated on. The name of this field should be generalized at some point in the future.
The job_execution.extra field may be used to store additional information necessary to allow operations on the running application
REST API impact¶
Other end user impact¶
None. Initially Spark jobs (jars) can be run using the Java job type. At some point a specific Spark job type will be added (this will be covered in a separate specification).
Sahara-dashboard / Horizon impact¶
implement default spark engine selection in spark provisioning plugin
implement launch wrapper
implement unit tests
Unit tests will be added for the changes in Sahara.
Integration tests for Spark standalone clusters will be added in another blueprint and specification.
The Elastic Data Processing section of the User Guide should talk about the ability to run Spark jobs and any restrictions.