Posts

fixing spark classpath issues on CDH5 accessing Accumulo 1.7.2

We experienced some strange NoSuchMethorError while migrating a Accumulo based application from 1.6.0 to 1.7.2 running on CDH5. A couple of code changes where necessary moving from 1.6.0 to 1.7.2, but these were pretty straightforward (members visibility changed, some getters were introduced). Everything compiled fine, but when we executed the spark application on the cluster we got an exception that was pointing directly to a line we changed during the migration:

Exception in thread "main" java.lang.NoSuchMethodError: 
        com.syscrest.HelloworldDriver$Opts.getPrincipal()Ljava/lang/String;
        at com.syscrest.HelloworldDriver.main(HelloworldDriver.java:29)
        at sun.reflect.NativeMethodAccessorImpl.invoke0(Native Method)
        at sun.reflect.NativeMethodAccessorImpl.invoke(NativeMethodAccessorImpl.java:62)
        at sun.reflect.DelegatingMethodAccessorImpl.invoke(DelegatingMethodAccessorImpl.java:43)
        at java.lang.reflect.Method.invoke(Method.java:498)
        at org.apache.spark.deploy.SparkSubmit$.org$apache$spark$deploy$SparkSubmit$$runMain(SparkSubmit.scala:731)
        at org.apache.spark.deploy.SparkSubmit$.doRunMain$1(SparkSubmit.scala:181)
        at org.apache.spark.deploy.SparkSubmit$.submit(SparkSubmit.scala:206)
        at org.apache.spark.deploy.SparkSubmit$.main(SparkSubmit.scala:121)
        at org.apache.spark.deploy.SparkSubmit.main(SparkSubmit.scala)

We double checked our fat jar that we bundle the right version and then we checked the full classpath of the CDH5 spark service.

the distributed spark classpath on CDH5

We noticed that CDH5 (CDH 5.8.2 in our case) already bundles four accumulo jars:

/opt/cloudera/parcels/CDH/jars/accumulo-core-1.6.0.jar
/opt/cloudera/parcels/CDH/jars/accumulo-fate-1.6.0.jar
/opt/cloudera/parcels/CDH/jars/accumulo-start-1.6.0.jar
/opt/cloudera/parcels/CDH/jars/accumulo-trace-1.6.0.jar

And all jars in /opt/cloudera/parcels/cdh/jars are automatically inserted into the spark distributed classpath, as they are listed in /etc/spark/conf/classpath.txt.

spark.{driver,executor}.userClassPathFirst

Before tampering with the spark distributed classpath we tried to get it working using

spark-submit \
  --conf "spark.driver.userClassPathFirst=true" \
  --conf "spark.executor.userClassPathFirst=true" \
  ....

but this resulted in another mismatch of libraries on the classpath and did not solved our initial problem:

org.xerial.snappy.SnappyNative.maxCompressedLength(I)I
java.lang.UnsatisfiedLinkError: org.xerial.snappy.SnappyNative.maxCompressedLength(I)I
at org.xerial.snappy.SnappyNative.maxCompressedLength(Native Method)
at org.xerial.snappy.Snappy.maxCompressedLength(Snappy.java:316)
at org.xerial.snappy.SnappyOutputStream.(SnappyOutputStream.java:79)
at org.apache.spark.io.SnappyCompressionCodec.compressedOutputStream(CompressionCodec.scala:157)
at org.apache.spark.broadcast.TorrentBroadcast$$anonfun$4.apply(TorrentBroadcast.scala:199)
at org.apache.spark.broadcast.TorrentBroadcast$$anonfun$4.apply(TorrentBroadcast.scala:199)
at scala.Option.map(Option.scala:145)
at org.apache.spark.broadcast.TorrentBroadcast$.blockifyObject(TorrentBroadcast.scala:199)
at org.apache.spark.broadcast.TorrentBroadcast.writeBlocks(TorrentBroadcast.scala:101)
at org.apache.spark.broadcast.TorrentBroadcast.(TorrentBroadcast.scala:84)
at org.apache.spark.broadcast.TorrentBroadcastFactory.newBroadcast(TorrentBroadcastFactory.scala:34)
at org.apache.spark.broadcast.TorrentBroadcastFactory.newBroadcast(TorrentBroadcastFactory.scala:29)
at org.apache.spark.broadcast.BroadcastManager.newBroadcast(BroadcastManager.scala:62)
at org.apache.spark.SparkContext.broadcast(SparkContext.scala:1051)
at org.apache.spark.rdd.NewHadoopRDD.(NewHadoopRDD.scala:77)
at org.apache.spark.SparkContext.newAPIHadoopRDD(SparkContext.scala:878)
at org.apache.spark.api.java.JavaSparkContext.newAPIHadoopRDD(JavaSparkContext.scala:516)

So we tried to fix the distributed spark classpath directly. Unfortunately /etc/spark/conf/classpath.txt is not modifiable via cloudera manager, and we did not want to change it manually on all nodes.

But as /etc/spark/conf/classpath.txt is being read in /etc/spark/conf/spark-env.sh

# Set distribution classpath. This is only used in CDH 5.3 and later.
export SPARK_DIST_CLASSPATH=$(paste -sd: "$SELF/classpath.txt")

we were able to use Spark Service Advanced Configuration Snippet (Safety Valve) for spark-conf/spark-env.sh – Spark (Service-Wide) to tamper with SPARK_DIST_CLASSPATH.

In our case it was sufficient to add the correct accumulo jars (version 1.7.2 from the installed parcel) at the beginning of the list so they take precedence over the outdated ones:

SPARK_DIST_CLASSPATH=/opt/cloudera/parcels/ACCUMULO/lib/accumulo/lib/accumulo-core.jar:$SPARK_DIST_CLASSPATH
SPARK_DIST_CLASSPATH=/opt/cloudera/parcels/ACCUMULO/lib/accumulo/lib/accumulo-fate.jar:$SPARK_DIST_CLASSPATH
SPARK_DIST_CLASSPATH=/opt/cloudera/parcels/ACCUMULO/lib/accumulo/lib/accumulo-start.jar:$SPARK_DIST_CLASSPATH
SPARK_DIST_CLASSPATH=/opt/cloudera/parcels/ACCUMULO/lib/accumulo/lib/accumulo-trace.jar:$SPARK_DIST_CLASSPATH
export SPARK_DIST_CLASSPATH