Spark read file from S3 using sc.textFile ("s3n://...)

Confirmed that this is related to the Spark build against Hadoop 2.60. Just installed Spark 1.4.0 "Pre built for Hadoop 2.4 and later" (instead of Hadoop 2.6). And the code now works OK.

sc.textFile("s3n://bucketname/Filename") now raises another error:

java.lang.IllegalArgumentException: AWS Access Key ID and Secret Access Key must be specified as the username or password (respectively) of a s3n URL, or by setting the fs.s3n.awsAccessKeyId or fs.s3n.awsSecretAccessKey properties (respectively).

The code below uses the S3 URL format to show that Spark can read S3 file. Using dev machine (no Hadoop libs).

scala> val lyrics = sc.textFile("s3n://MyAccessKeyID:MySecretKey@zpub01/SafeAndSound_Lyrics.txt")
lyrics: org.apache.spark.rdd.RDD[String] = MapPartitionsRDD[3] at textFile at <console>:21

scala> lyrics.count
res1: Long = 9

Even Better: the code above with AWS credentials inline in the S3N URI will break if the AWS Secret Key has a forward "/". Configuring AWS Credentials in SparkContext will fix it. Code works whether the S3 file is public or private.

sc.hadoopConfiguration.set("fs.s3n.awsAccessKeyId", "BLABLA")
sc.hadoopConfiguration.set("fs.s3n.awsSecretAccessKey", "....") // can contain "/"
val myRDD = sc.textFile("s3n://myBucket/MyFilePattern")
myRDD.count

Despite that this question has already an accepted answer, I think that the exact details of why this is happening are still missing. So I think there might be a place for one more answer.

If you add the required hadoop-aws dependency, your code should work.

Starting Hadoop 2.6.0, s3 FS connector has been moved to a separate library called hadoop-aws. There is also a Jira for that: Move s3-related FS connector code to hadoop-aws.

This means that any version of spark, that has been built against Hadoop 2.6.0 or newer will have to use another external dependency to be able to connect to the S3 File System.
Here is an sbt example that I have tried and is working as expected using Apache Spark 1.6.2 built against Hadoop 2.6.0:

libraryDependencies += "org.apache.hadoop" % "hadoop-aws" % "2.6.0"

In my case, I encountered some dependencies issues, so I resolved by adding exclusion:

libraryDependencies += "org.apache.hadoop" % "hadoop-aws" % "2.6.0" exclude("tomcat", "jasper-compiler") excludeAll ExclusionRule(organization = "javax.servlet")

On other related note, I have yet to try it, but that it is recommended to use "s3a" and not "s3n" filesystem starting Hadoop 2.6.0.

The third generation, s3a: filesystem. Designed to be a switch in replacement for s3n:, this filesystem binding supports larger files and promises higher performance.

You can add the --packages parameter with the appropriate jar: to your submission:

bin/spark-submit --packages com.amazonaws:aws-java-sdk-pom:1.10.34,org.apache.hadoop:hadoop-aws:2.6.0 code.py

I had to copy the jar files from a hadoop download into the $SPARK_HOME/jars directory. Using the --jars flag or the --packages flag for spark-submit didn't work.

Details:

• Spark 2.3.0
• Hadoop downloaded was 2.7.6
• Two jar files copied were from (hadoop dir)/share/hadoop/tools/lib/
  • aws-java-sdk-1.7.4.jar
  • hadoop-aws-2.7.6.jar

This is a sample spark code which can read the files present on s3

val hadoopConf = sparkContext.hadoopConfiguration
hadoopConf.set("fs.s3.impl", "org.apache.hadoop.fs.s3native.NativeS3FileSystem")
hadoopConf.set("fs.s3.awsAccessKeyId", s3Key)
hadoopConf.set("fs.s3.awsSecretAccessKey", s3Secret)
var jobInput = sparkContext.textFile("s3://" + s3_location)