[SPARK-25871] Streaming WAL should not use hdfs erasure coding, regardless of FS defaults

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 2.4.0
• Fix Version/s: 3.0.0
• Component/s: DStreams
• Assignee: Imran Rashid
• Reporter: Imran Rashid
• Created: 29/Oct/18 16:10
• Updated: 16/Nov/18 21:47
• Resolved: 06/Nov/18 18:53

Description

The FileBasedWriteAheadLogWriter expects the output stream for the WAL to support hflush(), but hdfs erasure coded files do not support that.

https://hadoop.apache.org/docs/r3.0.0/hadoop-project-dist/hadoop-hdfs/HDFSErasureCoding.html#Limitations

otherwise you get exceptions like:

17/10/17 17:31:34 ERROR executor.Executor: Exception in task 0.2 in stage 6.0 (TID 85)
org.apache.spark.SparkException: Could not read data from write ahead log record FileBasedWriteAheadLogSegment(hdfs://quasar-yxckyb-1.vpc.cloudera.com:8020/tmp/__spark__a10be3a3-85ec-4d4f-8782-a4760df4cc4c/88657/checkpoints/receivedData/0/log-1508286672978-1508286732978,1321921,189000)
	at org.apache.spark.streaming.rdd.WriteAheadLogBackedBlockRDD.org$apache$spark$streaming$rdd$WriteAheadLogBackedBlockRDD$$getBlockFromWriteAheadLog$1(WriteAheadLogBackedBlockRDD.scala:145)
	at org.apache.spark.streaming.rdd.WriteAheadLogBackedBlockRDD$$anonfun$compute$1.apply(WriteAheadLogBackedBlockRDD.scala:173)
	at org.apache.spark.streaming.rdd.WriteAheadLogBackedBlockRDD$$anonfun$compute$1.apply(WriteAheadLogBackedBlockRDD.scala:173)
	at scala.Option.getOrElse(Option.scala:121)
	at org.apache.spark.streaming.rdd.WriteAheadLogBackedBlockRDD.compute(WriteAheadLogBackedBlockRDD.scala:173)
	at org.apache.spark.rdd.RDD.computeOrReadCheckpoint(RDD.scala:323)
	at org.apache.spark.rdd.RDD.iterator(RDD.scala:287)
	at org.apache.spark.rdd.MapPartitionsRDD.compute(MapPartitionsRDD.scala:38)
	at org.apache.spark.rdd.RDD.computeOrReadCheckpoint(RDD.scala:323)
	at org.apache.spark.rdd.RDD.iterator(RDD.scala:287)
	at org.apache.spark.rdd.MapPartitionsRDD.compute(MapPartitionsRDD.scala:38)
	at org.apache.spark.rdd.RDD.computeOrReadCheckpoint(RDD.scala:323)
	at org.apache.spark.rdd.RDD.iterator(RDD.scala:287)
	at org.apache.spark.scheduler.ResultTask.runTask(ResultTask.scala:87)
	at org.apache.spark.scheduler.Task.run(Task.scala:108)
	at org.apache.spark.executor.Executor$TaskRunner.run(Executor.scala:338)
	at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1142)
	at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:617)
	at java.lang.Thread.run(Thread.java:745)
Caused by: java.io.EOFException: Cannot seek after EOF
	at org.apache.hadoop.hdfs.DFSStripedInputStream.seek(DFSStripedInputStream.java:331)
	at org.apache.hadoop.fs.FSDataInputStream.seek(FSDataInputStream.java:65)
	at org.apache.spark.streaming.util.FileBasedWriteAheadLogRandomReader.read(FileBasedWriteAheadLogRandomReader.scala:37)
	at org.apache.spark.streaming.util.FileBasedWriteAheadLog.read(FileBasedWriteAheadLog.scala:120)
	at org.apache.spark.streaming.rdd.WriteAheadLogBackedBlockRDD.org$apache$spark$streaming$rdd$WriteAheadLogBackedBlockRDD$$getBlockFromWriteAheadLog$1(WriteAheadLogBackedBlockRDD.scala:142)
	... 18 more

HDFS allows you to force a file to be replicated, regardless of the FS defaults – we should do that for the WAL.

Issue Links

• causes: SPARK-26094 Streaming WAL should create parent dirs
• relates to: SPARK-25855 Don't use Erasure Coding for event log files
• links to: [Github] Pull Request #22882 (squito)