If you have questions that are not resolved below, you can subscribe and post to the user mailing list. You can follow the instructions here.

Using PredictionIO

Q: How do I check to see if various dependencies, such as Elasticsearch and HBase, are running?

You can run $ pio status from the terminal and it will return the status of various components that PredictionIO depends on.

• You should see the following message if everything is OK:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19

$ pio status
PredictionIO
  Installed at: /home/vagrant/PredictionIO
  Version: 0.8.6

Apache Spark
  Installed at: /home/vagrant/PredictionIO/vendors/spark-1.2.0
  Version: 1.2.0 (meets minimum requirement of 1.2.0)

Storage Backend Connections
  Verifying Meta Data Backend
  Verifying Model Data Backend
  Verifying Event Data Backend
  Test write Event Store (App Id 0)
2015-02-03 18:52:38,904 INFO  hbase.HBLEvents - The table predictionio_eventdata:events_0 doesn't exist yet. Creating now...
2015-02-03 18:52:39,868 INFO  hbase.HBLEvents - Removing table predictionio_eventdata:events_0...

(sleeping 5 seconds for all messages to show up...)
Your system is all ready to go.

• If you see the following error message, it usually means Elasticsearch is not running properly:

1
2
3
4
5
6
7
8
9

  ...
Storage Backend Connections
  Verifying Meta Data Backend
  ...
Caused by: org.elasticsearch.client.transport.NoNodeAvailableException: None of the configured nodes are available: []
    at org.elasticsearch.client.transport.TransportClientNodesService.ensureNodesAreAvailable(TransportClientNodesService.java:298)
  ...

Unable to connect to all storage backend(s) successfully. Please refer to error message(s) above. Aborting.

You can check if there is any Elasticsearch process by running 'jps'.

Please see How to start Elasticsearch below.

• If you see the following error message, it usually means HBase is not running properly:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

Storage Backend Connections
  Verifying Meta Data Backend
  Verifying Model Data Backend
  Verifying Event Data Backend
2015-02-03 18:40:04,810 ERROR zookeeper.RecoverableZooKeeper - ZooKeeper exists failed after 1 attempts
2015-02-03 18:40:04,812 ERROR zookeeper.ZooKeeperWatcher - hconnection-0x1e4075ce, quorum=localhost:2181, baseZNode=/hbase Received unexpected KeeperException, re-throwing exception
org.apache.zookeeper.KeeperException$ConnectionLossException: KeeperErrorCode = ConnectionLoss for /hbase/hbaseid
...
2015-02-03 18:40:07,021 ERROR hbase.StorageClient - Failed to connect to HBase. Please check if HBase is running properly.
2015-02-03 18:40:07,026 ERROR storage.Storage$ - Error initializing storage client for source HBASE
2015-02-03 18:40:07,027 ERROR storage.Storage$ - Can't connect to ZooKeeper
java.util.NoSuchElementException: None.get
...

Unable to connect to all storage backend(s) successfully. Please refer to error message(s) above. Aborting.

You can check if there is any HBase-related process by running 'jps'.

Please see How to start HBase below.

Q: How to start Elasticsearch?

If you followed the instructions to install PredictionIO, Elasticsearch would have been installed at PredictionIO/vendors/elasticsearch-x.y.z/ where x.y.z is the version number. To start it, run:

1

$ ~/PredictionIO/vendors/elasticsearch-x.y.z/bin/elasticsearch

If you didn't use install script, please go to where Elasticsearch is installed to start it.

Q: How to start HBase ?

If you followed the instructions to install PredictionIO, the HBase is installed at ~/PredictionIO/vendors/hbase-x.y.z/ where x.y.z is the version number. To start it, run:

1

$ ~/PredictionIO/vendors/hbase-x.y.z/bin/start-hbase.sh

If you didn't use install script, please go to where HBase is installed to start it.

Problem with Event Server

Q: How do I increase the JVM heap size of the Event Server?

Add the JAVA_OPTS environmental variable to supply JVM options, e.g.

1

$ JAVA_OPTS=-Xmx16g bin/pio eventserver ...

Engine Training

Q: How to increase Spark driver program and worker executor memory size?

In general, the PredictionIO bin/pio scripts wraps around Spark's spark-submit script. You can specify a lot of Spark configurations (i.e. executor memory, cores, master url, etc.) with it. You can supply these as pass-through arguments at the end of bin/pio command.

If the engine training seems stuck, it's possible that the the executor doesn't have enough memory.

First, follow instruction here to start standalone Spark cluster and get the master URL. If you use the provided quick install script to install PredictionIO, the Spark is installed at PredictionIO/vendors/spark-1.2.0/ where you could run the Spark commands in sbin/ as described in the Spark documentation. Then use following train command to specify executor memory (default is only 512 MB) and driver memory.

For example, the follow command set the Spark master to spark://localhost:7077 (the default url of standalone cluster), set the driver memory to 16G and set the executor memory to 24G for pio train.

1

$ pio train -- --master spark://localhost:7077 --driver-memory 16G --executor-memory 24G

Q: How to resolve "Exception in thread "main" org.apache.spark.SparkException: Job aborted due to stage failure: Serialized task 165:35 was 110539813 bytes, which exceeds max allowed: spark.akka.frameSize (10485760 bytes) - reserved (204800 bytes). Consider increasing spark.akka.frameSize or using broadcast variables for large values."?

A likely reason is the local algorithm model is larger than the default frame size. You can specify a larger value as a pass-thru argument to spark-submit when you pio train. The following command increase the frameSize to 1024MB.

1

$ pio train -- --conf spark.akka.frameSize=1024

Deploy Engine

Q: How to increase heap space memory for "pio deploy"?

If you see the following error during pio deploy, it means there is not enough heap space memory.

1
2
3
4

...
[ERROR] [LocalFSModels] Java heap space
[ERROR] [OneForOneStrategy] None.get
...

To increase the heap space, specify the "-- --driver-memory " parameter in the command. For example, set the driver memory to 8G when deploy the engine:

1

$ pio deploy -- --driver-memory 8G

Building PredictionIO

Q: How to resolve "Error: Could not find or load main class org.apache.predictionio.tools.Console" after ./make_distribution.sh?

1
2

$ bin/pio app
Error: Could not find or load main class org.apache.predictionio.tools.Console

When PredictionIO bumps a version, it creates another JAR file with the new version number.

Delete everything but the latest pio-assembly-<VERSION>.jar in $PIO_HOME/assembly directory. For example:

1
2
3
4
5
6
7
8
9

PredictionIO$ cd assembly/
PredictionIO/assembly$ ls -al
total 197776
drwxr-xr-x  2 yipjustin yipjustin      4096 Nov 12 00:08 .
drwxr-xr-x 17 yipjustin yipjustin      4096 Nov 12 00:09 ..
-rw-r--r--  1 yipjustin yipjustin 101184982 Nov  5 06:05 pio-assembly-0.8.1-SNAPSHOT.jar
-rw-r--r--  1 yipjustin yipjustin 101324859 Nov 12 00:09 pio-assembly-0.8.2.jar

PredictionIO/assembly$ rm pio-assembly-0.8.1-SNAPSHOT.jar

Q: How to resolve ".......[error] (data/compile:compile) java.lang.AssertionError: assertion failed: java.lang.AutoCloseable" when ./make_distribution.sh?

PredictionIO only support Java 8 or later. Please make sure you have the correct Java version with the command:

1

$ javac -version

Engine Development

Q: What's the difference between P- and L- prefixed classes and functions?

PredictionIO v0.8 is built on the top of Spark, a massively scalable programming framework. A spark algorithm is different from conventional single machine algorithm in a way that spark algorithms use the RDD abstraction as its primary data type.

PredictionIO framework natively support both RDD-based algorithms and traditional single-machine algorithms. For controllers prefixed by "P" (i.e. PJavaDataSource, PJavaAlgorithm), their data include RDD abstraction; For "L" controllers, they are traditional single machine algorithms.

Running HBase

Q: How to resolve 'Exception in thread "main" java.lang.NullPointerException at org.apache.hadoop.net.DNS.reverseDns(DNS.java:92)'?

HBase relies on reverse DNS be set up properly to function. If your network configuration changes (such as working on a laptop with public WiFi hotspots), there could be a chance that reverse DNS does not function properly. You can install a DNS server on your own computer. Some users have reported that using Google Public DNS would also solve the problem.

Q: How to fix HBase issues after cleaning up a disk that was full?

You may receive error messages like write error: No space left on device when disk is full, and also receive error from pio status even after restarting PredictionIO services (due to an issue in ZooKeeper).

One quick fix is to delete the newest snapshot.xxxxx and log.xxxoo in the ZooKeeper data directory (e.g. $HBASE_HOME/zookeeper/zookeeper_0/version-2). Restart all services with pio-start-all, and use pio status to check whether your setup is good to go again.

If you still have problems connecting to the event server, take a look at the HBase dashboard to see if there are regions under transition. If so, try the following:

1. Try hbase hbck -repair and hbase hbck -repairHoles. If it solves the problem, you are all set. Otherwise, continue on.

2. Find out failing regions by hbase hbck.

   1 2 3 4 5 6 7

   ... Summary: Table pio_event:events_1 is inconsistent. Number of regions: 2 Deployed on: prediction.io,54829,1489213832255 ... 2 inconsistencies detected.

3. Shutdown HBase process and delete recovered.edits folders in the HBase data directory (e.g. $HBASE_HOME/hbase/data/pio_event/events_1 in this example) for failing regions.

4. Run hbase hbck -repairHoles and restart all PredictionIO services.