Follow the suggestions in these topics to help resolve common issues with running transformations with the Adaptive Execution Layer.
- Steps Cannot Run in Parallel
- Table Input Step Fails
- User ID Below Minimum Allowed
- Hadoop Version Conflict
- Hadoop Libraries are Missing
- Spark Libraries Conflict with Hadoop Libraries
- Message Size Exceeded
- Running Sub-Transformations on Spark
Steps Cannot Run in Parallel
If you are using the Spark engine to run a transformation with a step that cannot run in parallel, it generates errors in the log.
Some steps cannot run in parallel (on multiple nodes in a cluster), and will produce unexpected results. However, these steps can run as a coalesced dataset on a single node in a cluster. To enable a step to run as a coalesced dataset, add the step ID as a property value in the configuration file for using the Spark engine.
Get the Step ID
Each PDI step has a step ID, a globally unique identifier of the step. Use either of the following two methods to get the ID of a step:
Method 1: Retrieve the ID from the log
You can retrieve a step ID though the PDI client with the following steps:
In the PDI client, create a new transformation and add the step to the transformation. For example, if you needed to know the ID for the Select values step, you would add that step to the new transformation.
- Set the log level to debug.
- Execute the transformation using the Spark engine. The step ID will display in the Logging tab of the Execution Results pane. For example, the log will display
Selected the SelectValues step to run in parallel as a GenericSparkOperation,where SelectValues is the step ID.
Method 2: Retrieve the ID from the PDI plugin registry
If you are a developer, you can retrieve the step ID from the PDI plugin registry as described in Building Transformations Dynamically.
If you have created your own PDI transformation step plugin, the step ID is one of the annotation attributes that the developer supplies.
Add the Step ID to the Configuration File
The configuration file,
org.pentaho.pdi.engine.spark.cfg, contains the
forceCoalesceSteps property. The property is a pipe-delimited listing of all the IDs for the steps that should run with a coalesced dataset. Pentaho supplies a default set to which you can add IDs for steps that generate errors.
Perform the following steps to add another step ID to the configuration file:
- Navigate to the
data-integration/system/karaf/etcfolder and open the
- Append your step ID to the forceCoalesceSteps property value list, using a pipe character separator between the step IDs.
- Save and close the file.
Table Input Step Fails
If you run a transform using the Table Input step with a large database, the step will not complete. Use one of the following methods to resolve the issue:
Method 1: Load the data to HDFS before running the transform
- Run a different transformation using the Pentaho engine to move the data to the HDFS cluster.
- Then use HDFS Input to run the transformation using the Spark engine.
Method 2: Increase the driver side memory configuration
- Navigate to the
config/folder and open the
- Increase the value of the
sparkDriverMemoryparameter, then save and close the file.
User ID Below Minimum Allowed
If you are using the Spark engine in a secured cluster and an error about minimum user ID occurs, the user ID of the proxy user is below the minimum user ID required by the cluster. See Cloudera documentation for details.
To resolve, change the ID of the proxy user to be higher than the minimum user ID specified for the cluster.
Hadoop Version Conflict
On an HDP cluster, if you receive the following message, your Hadoop library is in conflict and the AEL Daemon along with the PDI client might stop working:
command hdp-select is not found, please manually *export HDP_VERSION* in spark-env.sh or current environment
To resolve the issue, you must export the HDP_VERSION variable using a command like the following example:
The HDP version number should match the HDP version number of the distribution on the cluster. You can check your HDP version with the hdp-select status hadoop-client command.
Hadoop Libraries are Missing
If you use the Spark libraries packaged with Cloudera and Hortonworks’ distributions, you must add the Hadoop libraries to the classpath with the
SPARK_DIST_CLASSPATH environment variable because these distributions are not packaged with the Hadoop libraries.
Add the CLASSPATH
The following command will add the libraries to the classpath:
export SPARK_DIST_CLASSPATH=$(hadoop classpath)
You can add this command to the
daemon.sh file so you do not have run this command every time you the start the AEL daemon.
Set SPARKHOME Variable
If you are using the Spark client from the Cloudera or Hortonworks Hadoop distributions, you may also receive the following log error:
Exception in thread "main" java.lang.NoSuchFieldError: TOKEN_KIND
If you received this log error, you must also complete the following steps for your Hadoop distribution:
- Download the Spark client for your Hadoop cluster distribution (Cloudera or Hortonworks).
- Navigate to the
adaptive-execution/configdirectory and open the
- Set the
sparkHome location to where Spark 2 is located on your machine.
Example for Cloudera:
Example for Hortonworks:
Spark Libraries Conflict with Hadoop Libraries
In some cases, library versions contained in JARs from PDI, Spark, Hadoop, AEL, and/or Kettle plugins may conflict with one another, causing general problems where Spark libraries conflict with Hadoop libraries and potentially creating AEL-specific problems. To read more about this issue, including how to address it, see the article AEL and Spark Library Conflicts on the Pentaho Community Wiki.
Message Size Exceeded
If you are using the Spark engine to run an AEL transformation and an error is generated indicating a decoded message was too big for the output buffer, you need to increase the maximum size (2 MB by default) of the message buffers for your AEL environment.
Perform the following steps to increase the message buffer limit:
- Stop the AEL daemon.
- Navigate to the .../data-integration/adaptive-execution/config directory and open the application.properties file using a text editor.
- Enter the following incoming WebSocket message buffer properties, setting the same value for each property:
Property Value daemon.websocket.maxMessageBufferSize
The maximum size (in bytes) for the message buffer on the AEL daemon. For example, to allocate a 4 MB limit, set the memory value as shown here:
The maximum size (in bytes) for the message buffer on the AEL Spark driver. For example, to allocate a 4 MB limit, set the memory value as shown here:
- Save and close the file.
- Restart the AEL daemon.
When the AEL daemon submits the AEL Spark driver application, it passes the driver’s maximum message buffer size value as part of the submit; then, when the driver application is started, it receives the maximum buffer size value sent by the daemon.
Running Sub-Transformations on Spark
To use a sub-transformation with Spark, you must use the Transformation Executor step. Transformations with these steps are not supported in Spark:
- Simple Mapping (sub-transformation)
- Mapping (sub-transformation)