hadoop - How do you use s3a with spark 2.1.0 on aws us-east-2?

Question

Background

I have been working on getting a flexible setup for myself to use spark on aws with docker swarm mode. The docker image I have been using is configured to use the latest spark, which at the time is 2.1.0 with Hadoop 2.7.3, and is available at jupyter/pyspark-notebook.

This is working, and I have been just going through to test out the various connectivity paths that I plan to use. The issue I came across is with the uncertainty around the correct way to interact with s3. I have followed the trail on how to provide the dependencies for spark to connect to data on aws s3 using the s3a protocol, vs s3n protocol.

I finally came across the hadoop aws guide and thought I was following how to provide the configuration. However, I was still receiving the 400 Bad Request error, as seen in this question that describes how to fix it by defining the endpoint, which I had already done.

I ended up being too far off the standard configuration by being on us-east-2, making me uncertain if I had a problem with the jar files. To eliminate the region issue, I set things back up on the regular us-east-1 region, and I was able to finally connect with s3a. So I have narrowed down the problem to the region, but thought I was doing everything required to operate on the other region.

Question

What is the correct way to use the configuration variables for hadoop in spark to use us-east-2?

Note: This example uses local execution mode to simplify things.

import os
import pyspark

I can see in the console for the notebook these download after creating the context, and adding these took me from being completely broken, to getting the Bad Request error.

os.environ['PYSPARK_SUBMIT_ARGS'] = '--packages com.amazonaws:aws-java-sdk:1.7.4,org.apache.hadoop:hadoop-aws:2.7.3 pyspark-shell'

conf = pyspark.SparkConf('local[1]')
sc = pyspark.SparkContext(conf=conf)
sql = pyspark.SQLContext(sc)

For the aws config, I tried both the below method and by just using the above conf, and doing conf.set(spark.hadoop.fs.<config_string>, <config_value>) pattern equivalent to what I do below, except doing it this was I set the values on conf before creating the spark context.

hadoop_conf = sc._jsc.hadoopConfiguration()

hadoop_conf.set("fs.s3.impl", "org.apache.hadoop.fs.s3a.S3AFileSystem")
hadoop_conf.set("fs.s3a.endpoint", "s3.us-east-2.amazonaws.com")
hadoop_conf.set("fs.s3a.access.key", access_id)
hadoop_conf.set("fs.s3a.secret.key", access_key)

One thing to note, is that I also tried an alternative endpoint for us-east-2 of s3-us-east-2.amazonaws.com.

I then read some parquet data off of s3.

df = sql.read.parquet('s3a://bucket-name/parquet-data-name')
df.limit(10).toPandas()

Again, after moving the EC2 instance to us-east-1, and commenting out the endpoint config, the above works for me. To me, it seems like endpoint config isn't being used for some reason.

Answer 1

us-east-2 is a V4 auth S3 instance so, as you attemped, the fs.s3a.endpoint value must be set.

if it's not being picked up then assume the config you are setting isn't the one being used to access the bucket. Know that Hadoop caches filesystem instances by URI, even when the config changes. The first attempt to access a filesystem fixes, the config, even when its lacking in auth details.

Some tactics

1. set the value is spark-defaults
2. using the config you've just created, try to explicitly load the filesystem via a call to Filesystem.get(new URI("s3a://bucket-name/parquet-data-name", myConf) will return the bucket with that config (unless it's already there). I don't know how to make that call in .py though.
3. set the property "fs.s3a.impl.disable.cache" to true to bypass the cache before the get command

Adding more more diagnostics on BadAuth errors, along with a wiki page, is a feature listed for S3A phase III. If you were to add it, along with a test, I can review it and get it in