Spark 2 Application Errors & Solutions

Exception - 

Exception in thread "broadcast-exchange-0" java.lang.OutOfMemoryError: Not enough memory to build and broadcast

This is Driver Exception and can be solved by 

• setting spark.sql.autoBroadcastJoinThreshold to -1
• Or, increasing --driver-memory

Exception - 

Container  is running beyond physical memory limits.

Current usage: X GB of Y GB physical memory used; X GB of Y GB virtual memory used. Killing container

YARN killed container as it was exceeding memory limits.

• Increase 
• --driver-memory
• --executor-memory
 

Exception -

ERROR Executor: Exception in task 600 in stage X.X (TID 12345)

java.lang.OutOfMemoryError: GC overhead limit exceeded

This means that Executor JVM was spending more time in Garbage collection than actual execution. 

• This JVM feature can be disabled by adding -XX:-UseGCOverheadLimit
• Increasing Executor memory may help --executor-memory
• Make data more distributed so that it is not skewed to one executor.
• Might use parallel GC -XX:+UseParallelGC or -XX:+UseConcMarkSweepGC

Exception -

org.apache.spark.shuffle.FetchFailedException: failed to allocate 65536 byte(s) of direct
memory (used: 1073699840, max: 1073741824)
at org.apache.spark.storage.ShuffleBlockFetcherIterator.throwFetchFailedException(ShuffleBlockFetcherIterator.scala:442)

This means that executor went out of memory .

• Can increase executor memory --executor-memory
• Make data more distributed so that it is not skewed to one executor.
• Increase shuffle partitions  --spark.sql.shuffle.partitions

Exception -

ExecutorLostFailure (executor 525 exited unrelated to the running tasks) Reason: Container container_1495825717937_0056_01_000916 on host: 10.0.0.14 was preempted.

This means you were running above YARN queue capacity assigned to your Job. 

• Ask for more YARN resources, or schedule Job when resources are available to you.

Exception - 

WARN TaskSetManager: Lost task 49.2 in stage 6.0 (TID xxx, xxx.xxx.xxx.compute.internal): ExecutorLostFailure (executor 16 exited caused by one of the running tasks) Reason: Container killed by YARN for exceeding memory limits. 10.4 GB of 10.4 GB physical memory used. Consider boosting spark.yarn.executor.memoryOverhead.

This means that executor is running out of memory as assigned by YARN

• Can increase spark.yarn.executor.memoryOverhead
• can increase --executor-memory
• Can try reducing number of cores for an executor --executor-cores

Exception- 

org.apache.spark.SparkException: Job aborted due to stage failure: Total size of serialized results of 16 tasks (1048.5 MB) is bigger than spark.driver.maxResultSize (1024.0 MB)

This means that total size of results is greater than the Spark Driver Max Result Size value. This not necessarily means that you are doing a collect causing results to be accumulated on driver. It may be the case that your Job is huge and resulting in large number of tasks, as tasks are serialized to executors by driver. 

• Consider boosting spark.driver.maxResultSize
• Or, may be break your job in to multiple sub jobs.

Exception - 

Caused by: org.apache.spark.shuffle.FetchFailedException: Too large frame: 5454002341

When the size of the shuffle data blocks exceeds the limit of 2 GB, which spark can handle, 

• Identify and re-partition the dataframe 
• increase parallelism spark.sql.shuffle.partitions

Exception - 

Caused by: java.lang.RuntimeException: Unsupported data type NullType.

at scala.sys.package$.error(package.scala:27)

This might be resulted because of SQL in place. In some cases, it is required to select NULL value(s). For example : Select NULL as col1 from Table1. Spark will not be able to determine data type for NULL value. Thus, our job fails with above error.

• please try to update the query and cast the NULL to appropriate data type . For example – cast (NULL as String) as col1

Exception - 

Caused by: org.apache.spark.sql.AnalysisException: Cannot overwrite table XXX that is also being read from;

Refer - http://querydb.blogspot.com/2019/07/read-from-hive-table-and-write-back-to.html

Exception - 

java.lang.NoClassDefFoundError: Could not initialize class org.xerial.snappy.Snappy

        at org.apache.parquet.hadoop.codec.SnappyDecompressor.decompress(SnappyDecompressor.java:62)

OR 

Caused by: java.lang.UnsatisfiedLinkError: /tmp/snappy-1.1.2-d5273c94-b734-4a61-b631-b68a9e859151-libsnappyjava.so: /tmp/snappy-1.1.2-d5273c94-b734-4a61-b631-b68a9e859151-libsnappyjava.so: failed to map segment from shared object: Operation not permitted

        at java.lang.ClassLoader$NativeLibrary.load(Native Method)

It is because that /tmp doesn't have execute permissions.

• Set temporary directory: --conf "spark.driver.extraJavaOptions=-Djava.io.tmpdir=/a/b/ctmp" --conf "spark.executor.extraJavaOptions=-Djava.io.tmpdir=/a/b/ctmp"

Exception-

org.apache.spark.sql.AnalysisException: Detected cartesian product for INNER join between logical plans

If source data has a static partition value than Spark will analyze execution plan thinking that it is a case wherein it should be a Cross join instead of Inner join.

• set property - "set spark.sql.crossJoin.enabled=true;"