I might be mistaken, but yes, even with the changes you mentioned you will not be able to access S3 if Spark is built against Hadoop 2.6+ unless you install additional libraries. The issue is explained in SPARK-7481 <https://issues.apache.org/jira/browse/SPARK-7481> and SPARK-7442 <https://issues.apache.org/jira/browse/SPARK-7442>.
On Fri, Nov 6, 2015 at 12:22 AM Christian <engr...@gmail.com> wrote: > Even with the changes I mentioned above? > On Thu, Nov 5, 2015 at 8:10 PM Nicholas Chammas < > nicholas.cham...@gmail.com> wrote: > >> Yep, I think if you try spark-1.5.1-hadoop-2.6 you will find that you >> cannot access S3, unfortunately. >> >> On Thu, Nov 5, 2015 at 3:53 PM Christian <engr...@gmail.com> wrote: >> >>> I created the cluster with the following: >>> >>> --hadoop-major-version=2 >>> --spark-version=1.4.1 >>> >>> from: spark-1.5.1-bin-hadoop1 >>> >>> Are you saying there might be different behavior if I download >>> spark-1.5.1-hadoop-2.6 and create my cluster? >>> >>> On Thu, Nov 5, 2015 at 1:28 PM, Christian <engr...@gmail.com> wrote: >>> >>>> Spark 1.5.1-hadoop1 >>>> >>>> On Thu, Nov 5, 2015 at 10:28 AM, Nicholas Chammas < >>>> nicholas.cham...@gmail.com> wrote: >>>> >>>>> > I am using both 1.4.1 and 1.5.1. >>>>> >>>>> That's the Spark version. I'm wondering what version of Hadoop your >>>>> Spark is built against. >>>>> >>>>> For example, when you download Spark >>>>> <http://spark.apache.org/downloads.html> you have to select from a >>>>> number of packages (under "Choose a package type"), and each is built >>>>> against a different version of Hadoop. When Spark is built against Hadoop >>>>> 2.6+, from my understanding, you need to install additional libraries >>>>> <https://issues.apache.org/jira/browse/SPARK-7481> to access S3. When >>>>> Spark is built against Hadoop 2.4 or earlier, you don't need to do this. >>>>> >>>>> I'm confirming that this is what is happening in your case. >>>>> >>>>> Nick >>>>> >>>>> On Thu, Nov 5, 2015 at 12:17 PM Christian <engr...@gmail.com> wrote: >>>>> >>>>>> I am using both 1.4.1 and 1.5.1. In the end, we used 1.5.1 because of >>>>>> the new feature for instance-profile which greatly helps with this as >>>>>> well. >>>>>> Without the instance-profile, we got it working by copying a >>>>>> .aws/credentials file up to each node. We could easily automate that >>>>>> through the templates. >>>>>> >>>>>> I don't need any additional libraries. We just need to change the >>>>>> core-site.xml >>>>>> >>>>>> -Christian >>>>>> >>>>>> On Thu, Nov 5, 2015 at 9:35 AM, Nicholas Chammas < >>>>>> nicholas.cham...@gmail.com> wrote: >>>>>> >>>>>>> Thanks for sharing this, Christian. >>>>>>> >>>>>>> What build of Spark are you using? If I understand correctly, if you >>>>>>> are using Spark built against Hadoop 2.6+ then additional configs alone >>>>>>> won't help because additional libraries also need to be installed >>>>>>> <https://issues.apache.org/jira/browse/SPARK-7481>. >>>>>>> >>>>>>> Nick >>>>>>> >>>>>>> On Thu, Nov 5, 2015 at 11:25 AM Christian <engr...@gmail.com> wrote: >>>>>>> >>>>>>>> We ended up reading and writing to S3 a ton in our Spark jobs. >>>>>>>> For this to work, we ended up having to add s3a, and s3 key/secret >>>>>>>> pairs. We also had to add fs.hdfs.impl to get these things to work. >>>>>>>> >>>>>>>> I thought maybe I'd share what we did and it might be worth adding >>>>>>>> these to the spark conf for out of the box functionality with S3. >>>>>>>> >>>>>>>> We created: >>>>>>>> >>>>>>>> ec2/deploy.generic/root/spark-ec2/templates/root/spark/conf/core-site.xml >>>>>>>> >>>>>>>> We changed the contents form the original, adding in the following: >>>>>>>> >>>>>>>> <property> >>>>>>>> <name>fs.file.impl</name> >>>>>>>> <value>org.apache.hadoop.fs.LocalFileSystem</value> >>>>>>>> </property> >>>>>>>> >>>>>>>> <property> >>>>>>>> <name>fs.hdfs.impl</name> >>>>>>>> <value>org.apache.hadoop.hdfs.DistributedFileSystem</value> >>>>>>>> </property> >>>>>>>> >>>>>>>> <property> >>>>>>>> <name>fs.s3.impl</name> >>>>>>>> <value>org.apache.hadoop.fs.s3native.NativeS3FileSystem</value> >>>>>>>> </property> >>>>>>>> >>>>>>>> <property> >>>>>>>> <name>fs.s3.awsAccessKeyId</name> >>>>>>>> <value>{{aws_access_key_id}}</value> >>>>>>>> </property> >>>>>>>> >>>>>>>> <property> >>>>>>>> <name>fs.s3.awsSecretAccessKey</name> >>>>>>>> <value>{{aws_secret_access_key}}</value> >>>>>>>> </property> >>>>>>>> >>>>>>>> <property> >>>>>>>> <name>fs.s3n.awsAccessKeyId</name> >>>>>>>> <value>{{aws_access_key_id}}</value> >>>>>>>> </property> >>>>>>>> >>>>>>>> <property> >>>>>>>> <name>fs.s3n.awsSecretAccessKey</name> >>>>>>>> <value>{{aws_secret_access_key}}</value> >>>>>>>> </property> >>>>>>>> >>>>>>>> <property> >>>>>>>> <name>fs.s3a.awsAccessKeyId</name> >>>>>>>> <value>{{aws_access_key_id}}</value> >>>>>>>> </property> >>>>>>>> >>>>>>>> <property> >>>>>>>> <name>fs.s3a.awsSecretAccessKey</name> >>>>>>>> <value>{{aws_secret_access_key}}</value> >>>>>>>> </property> >>>>>>>> >>>>>>>> This change makes spark on ec2 work out of the box for us. It took >>>>>>>> us several days to figure this out. It works for 1.4.1 and 1.5.1 on >>>>>>>> Hadoop >>>>>>>> version 2. >>>>>>>> >>>>>>>> Best Regards, >>>>>>>> Christian >>>>>>>> >>>>>>> >>>>>> >>>> >>>