No FileSystem for scheme: cos - pyspark

Question

I'm trying to connect to IBM Cloud Object Storage from IBM Data Science Experience:
access_key = 'XXX'
secret_key = 'XXX'
bucket = 'mybucket'
host = 'lon.ibmselect.objstor.com'
service = 'mycos'
sqlCxt = SQLContext(sc)
hconf = sc._jsc.hadoopConfiguration()
hconf.set('fs.cos.myCos.access.key', access_key)
hconf.set('fs.cos.myCos.endpoint', 'http://' + host)
hconf.set('fs.cose.myCos.secret.key', secret_key)
hconf.set('fs.cos.service.v2.signer.type', 'false')
obj = 'mydata.tsv.gz'
rdd = sc.textFile('cos://{0}.{1}/{2}'.format(bucket, service, obj))
print(rdd.count())
This returns:
Py4JJavaError: An error occurred while calling z:org.apache.spark.api.python.PythonRDD.collectAndServe.
: java.io.IOException: No FileSystem for scheme: cos
I'm guessing I need to use the 'cos' scheme based on the stocator docs. However, the error suggests stocator isn't available or is an old version?
Any ideas?
Update 1:
I have also tried the following:
sqlCxt = SQLContext(sc)
hconf = sc._jsc.hadoopConfiguration()
hconf.set('fs.cos.impl', 'com.ibm.stocator.fs.ObjectStoreFileSystem')
hconf.set('fs.stocator.scheme.list', 'cos')
hconf.set('fs.stocator.cos.impl', 'com.ibm.stocator.fs.cos.COSAPIClient')
hconf.set('fs.stocator.cos.scheme', 'cos')
hconf.set('fs.cos.mycos.access.key', access_key)
hconf.set('fs.cos.mycos.endpoint', 'http://' + host)
hconf.set('fs.cos.mycos.secret.key', secret_key)
hconf.set('fs.cos.service.v2.signer.type', 'false')
service = 'mycos'
obj = 'mydata.tsv.gz'
rdd = sc.textFile('cos://{0}.{1}/{2}'.format(bucket, service, obj))
print(rdd.count())
However, this time the response was:
Py4JJavaError: An error occurred while calling z:org.apache.spark.api.python.PythonRDD.collectAndServe.
: java.io.IOException: No object store for: cos
at com.ibm.stocator.fs.ObjectStoreVisitor.getStoreClient(ObjectStoreVisitor.java:121)
...
Caused by: java.lang.ClassNotFoundException: com.ibm.stocator.fs.cos.COSAPIClient

Answer 1

The latest version of Stocator (v1.0.9) that supports fs.cos scheme is not yet deployed on Spark aaService (It will be soon). Please use the stocator scheme "fs.s3d" to connect to your COS.
Example:
endpoint = 'endpointXXX'
access_key = 'XXX'
secret_key = 'XXX'
prefix = "fs.s3d.service"
hconf = sc._jsc.hadoopConfiguration()
hconf.set(prefix + ".endpoint", endpoint)
hconf.set(prefix + ".access.key", access_key)
hconf.set(prefix + ".secret.key", secret_key)
bucket = 'mybucket'
obj = 'mydata.tsv.gz'
rdd = sc.textFile('s3d://{0}.service/{1}'.format(bucket, obj))
rdd.count()
Alternatively, you can use ibmos2spark. The lib is already installed on our service. Example:
import ibmos2spark
credentials = {
'endpoint': 'endpointXXXX',
'access_key': 'XXXX',
'secret_key': 'XXXX'
}
configuration_name = 'os_configs' # any string you want
cos = ibmos2spark.CloudObjectStorage(sc, credentials, configuration_name)
bucket = 'mybucket'
obj = 'mydata.tsv.gz'
rdd = sc.textFile(cos.url(obj, bucket))
rdd.count()

Answer 2

Stocator is on the classpath for Spark 2.0 and 2.1 kernels, but the cos scheme is not configured. You can access the config by executing the following in a Python notebook:
!cat $SPARK_CONF_DIR/core-site.xml
Look for the property fs.stocator.scheme.list. What I currently see is:
<property>
<name>fs.stocator.scheme.list</name>
<value>swift2d,swift,s3d</value>
</property>
I recommend that you raise a feature request against DSX to support the cos scheme.

Answer 3

It looks like cos driver is not properly initialized. Try this configuration:
hconf.set('fs.cos.impl', 'com.ibm.stocator.fs.ObjectStoreFileSystem')
hconf.set('fs.stocator.scheme.list', 'cos')
hconf.set('fs.stocator.cos.impl', 'com.ibm.stocator.fs.cos.COSAPIClient')
hconf.set('fs.stocator.cos.scheme', 'cos')
hconf.set('fs.cos.mycos.access.key', access_key)
hconf.set('fs.cos.mycos.endpoint', 'http://' + host)
hconf.set('fs.cos.mycos.secret.key', secret_key)
hconf.set('fs.cos.service.v2.signer.type', 'false')
UPDATE 1:
You also need to ensure stocator classes are on the classpath. You can use packages system by exceuting pyspark in the following way:
./bin/pyspark --packages com.ibm.stocator:stocator:1.0.24
This works with swift2d and cos scheme.
UPDATE 2:
Just follow Stocator documentation (https://github.com/CODAIT/stocator). It contains all details how to install it, what branch to use, etc.

Answer 4

I found the same issue, and to solve it I just changed environment:
Within IBM Watson Studio, if you start a a Jupyter notebook in an environment without a pre-configured spark cluster, than you get that error. Installing PySpark is not enough.
Instead, if you start a notebook with the Spark cluster available, you will be just fine.

Answer 5

You have to set .config("spark.hadoop.fs.stocator.scheme.list", "cos") along with some others fs.cos... configurations.
Here's an end-to-end snippet code example that works (tested with pyspark==2.3.2 and Python 3.7.3):
from pyspark.sql import SparkSession
stocator_jar = '/path/to/stocator-1.1.2-SNAPSHOT-IBM-SDK.jar'
cos_instance_name = '<myCosIntanceName>'
bucket_name = '<bucketName>'
s3_region = '<region>'
cos_iam_api_key = '*******'
iam_servicce_id = 'crn:v1:bluemix:public:iam-identity::<****************>'
spark_builder = (
SparkSession
.builder
.appName('test_app'))
spark_builder.config('spark.driver.extraClassPath', stocator_jar)
spark_builder.config('spark.executor.extraClassPath', stocator_jar)
spark_builder.config(f"fs.cos.{cos_instance_name}.iam.api.key", cos_iam_api_key)
spark_builder.config(f"fs.cos.{cos_instance_name}.endpoint", f"s3.{s3_region}.cloud-object-storage.appdomain.cloud")
spark_builder.config(f"fs.cos.{cos_instance_name}.iam.service.id", iam_servicce_id)
spark_builder.config("spark.hadoop.fs.stocator.scheme.list", "cos")
spark_builder.config("spark.hadoop.fs.cos.impl", "com.ibm.stocator.fs.ObjectStoreFileSystem")
spark_builder.config("fs.stocator.cos.impl", "com.ibm.stocator.fs.cos.COSAPIClient")
spark_builder.config("fs.stocator.cos.scheme", "cos")
spark_sess = spark_builder.getOrCreate()
dataset = spark_sess.range(1, 10)
dataset = dataset.withColumnRenamed('id', 'user_idx')
dataset.repartition(1).write.csv(
f'cos://{bucket_name}.{cos_instance_name}/test.csv',
mode='overwrite',
header=True)
spark_sess.stop()
print('done!')