I have deployed a lambda function which uses sparknlp, as a docker container. For working with sparknlp I need spark context. So, In my sparknlp code, I start with
sc = pyspark.SparkContext().getOrCreate()
I tested my lambda on local and it worked fine.
On aws I got this error :
java gateway process exited before sending its port number
even though JAVA_HOME was properly set.
I found out in the source code :
https://github.com/apache/spark/blob/master/python/pyspark/java_gateway.py
in the launch_gateway method that it tries to create a temporary file and if that file is not created it raises the above error. (line 105)
Lambda won't allow write access to file system, so the file cannot be created.
So I am trying to pass gateway_port and gateway_secret as environment variables.
I have kept the PYSPARK_GATEWAY_PORT=25333 which is the default value.
I am not able to figure out how to get the PYSPARK_GATEWAY_SECRET.
Which is why getting the error :
KeyError: 'PYSPARK_GATEWAY_SECRET'
Related
I can confirm the 3-replica cluster of h2o inside K3s is correctly deployed, as executing in the Python3 interpreter h2o.init(ip="x.x.x.x") works as expected. I followed the instructions noted here: https://www.h2o.ai/blog/running-h2o-cluster-on-a-kubernetes-cluster/
Nevertheless, I had to modify the service.yaml and comment out the line which says clusterIP: None, as K3s was complaining about something related to its inability to set the clusterIP to None. But even though, I can certify it is working correctly, and I am able to use an external IP to connect to the cluster.
If I try to load the dataset using the h2o cluster inside the K3s cluster using the exact same steps as described here http://docs.h2o.ai/h2o/latest-stable/h2o-docs/automl.html, this is the output that I get:
>>> train = h2o.import_file("https://s3.amazonaws.com/erin-data/higgs/higgs_train_10k.csv")
...
h2o.exceptions.H2OResponseError: Server error java.lang.IllegalArgumentException:
Error: Key not loaded: Key<Frame> https://s3.amazonaws.com/erin-data/higgs/higgs_train_10k.csv
Request: POST /3/ParseSetup
data: {'check_header': '0', 'source_frames': '["https://s3.amazonaws.com/erin-data/higgs/higgs_train_10k.csv"]'}
The same error occurs if I use the h2o.upoad_file("x.csv") method.
There is a clue about what may be happening here: Key not loaded: Key<Frame> while POSTing source frame through ParseSetup in H2O API call but I am not using curl, and I can not find any parameter that could help me overcome this issue: http://docs.h2o.ai/h2o/latest-stable/h2o-py/docs/h2o.html?highlight=import_file#h2o.import_file
I need to use the Python client inside the same K3s cluster due to different technical reasons, so I am not able to kick off nor Flow nor Firebug to know what may be happening.
I can confirm it is working correctly when I simply issue a h2o.init(), using the local Java instance.
UPDATE 1:
I have tried in different K3s clusters without success. I changed the service.yaml to a NodePort, and now this is the error traceback:
>>> train = h2o.import_file("https://s3.amazonaws.com/erin-data/higgs/higgs_train_10k.csv")
...
h2o.exceptions.H2OResponseError: Server error java.lang.IllegalArgumentException:
Error: Job is missing
Request: GET /3/Jobs/$03010a2a016132d4ffffffff$_a2366be93ec99a78d7bc161de8c54d67
UPDATE 2:
I have tried using different services (NodePort, LoadBalancer, ClusterIP) and none of them work. I also have tried using Minikube with the official image, and with a custom image made by me, without success. I suspect this is something related to either h2o itself, or the clustering between pods. I will keep digging and let's think there will be some gold in it.
UPDATE 3:
I also found out that the post about running H2O in Docker is really outdated https://www.h2o.ai/blog/h2o-docker/ nor is working the Dockerfile present at GitHub (I changed it to uncomment the ENTRYPOINT section without success): https://github.com/h2oai/h2o-3/blob/master/Dockerfile
Even though, I tried with the custom image I built for h2o-k8s and it is working seamlessly in pure Docker. I am wondering why it is still not working in K8s...
UPDATE 4:
I have tried modifying the environment variable called H2O_KUBERNETES_SERVICE_DNS without success.
In the meantime, the cluster started to be unavailable, that is, the readinessProbe's would not successfully complete. No matter what I change now, it does not work.
I spinned up a K3d cluster in local to see what happened, and surprisingly, the readinessProbe's were not failing, using v3.30.0.6. But now I started testing it with R instead of Python. I am glad I tried, because I may have pinpointed what was wrong. There is a version mismatch between the client and the server. So I updated accordingly the image to v3.30.0.1.
But now again, the readinessProbe is not working in my k3d cluster, so I am unable to test it.
It seems it is working now. R client version 3.30.0.1 with server version 3.30.0.1. Also tried with Python version 3.30.0.7 and server version 3.30.0.7 and it started working. Marvelous. The problem was caused by a version mismatch between the client and the server, as the python client was updated to 3.30.0.7 while the latest server for docker was 3.30.0.6.
I'm trying to access small text file stored directly on dbfs using local file API.
I'm getting the following error.
No such file or directory
My code:
val filename = "/dbfs/test/test.txt"
for (line <- Source.fromFile(filename).getLines()) {
println(line)
}
At the same time I can access this file without any problems using dbutils or load it to RDD via spark context.
I've tried specifying the path starting with dbfs:/ or /dbfs/ or just with the test folder name, both in Scala and Python, getting the same error each time. I'm running the code from the notebook. Is it some problem with the cluster configuration?
Check if your cluster has Credentials Passthrough enabled. If so, local file Api is not available.
https://docs.azuredatabricks.net/data/databricks-file-system.html#local-file-apis
I have an EC2 instance running pyspark and I'm able to connect to it (ssh) and run interactive code within a Jupyter Notebook.
I have a S3 bucket with a csv file that I want to read, when I attempt to read it with:
spark = SparkSession.builder.appName('Basics').getOrCreate()
df = spark.read.csv('https://s3.us-east-2.amazonaws.com/bucketname/filename.csv')
Which throws a long Python error message and then something related to:
Py4JJavaError: An error occurred while calling o131.csv.
Specify S3 path along with access key and secret key as following:
's3n://<AWS_ACCESS_KEY_ID>:<AWS_SECRET_ACCESS_KEY>#my.bucket/folder/input_data.csv'
Access key-related information can be introduced in the typical username + password manner for URLs. As a rule, the access protocol should be s3a, the successor to s3n (see Technically what is the difference between s3n, s3a and s3?). Putting this together, you get
spark.read.csv("s3a://<AWS_ACCESS_KEY_ID>:<AWS_SECRET_ACCESS_KEY>#bucketname/filename.csv")
As an aside, some Spark execution environments, e.g., Databricks, allow S3 buckets to be mounted as part of the file system. You can do the same when you build a cluster using something like s3fs.
I'm using the boto3 package to connect from outside an s3 cluster (i.e. the script is currently not being run within the AWS 'cloud', but from my MBP connecting to the relevant cluster). My code:
s3 = boto3.resource(
"s3",
aws_access_key_id=self.settings['CREDENTIALS']['aws_access_key_id'],
aws_secret_access_key=self.settings['CREDENTIALS']['aws_secret_access_key'],
)
bucket = s3.Bucket(self.settings['S3']['bucket_test'])
for bucket_in_all in boto3.resource('s3').buckets.all():
if bucket_in_all.name == self.settings['S3']['bucket_test']:
print ("Bucket {} verified".format(self.settings['S3']['bucket_test']))
Now I'm receiving this error message:
botocore.exceptions.ClientError: An error occurred (SignatureDoesNotMatch) when calling the ListBuckets operation
I'm aware of the sequence of how the aws credentials are checked, and tried different permutations of my environment variables and ~/.aws/credentials, and know that the credentials as per my .py script should override, however I'm still seeing this SignatureDoesNotMatch error message. Any ideas where I may be going wrong? I've also tried:
# Create a session
session = boto3.session.Session(
aws_access_key_id=self.settings['CREDENTIALS']['aws_access_key_id'],
aws_secret_access_key=self.settings['CREDENTIALS']['aws_secret_access_key'],
aws_session_token=self.settings['CREDENTIALS']['session_token'],
region_name=self.settings['CREDENTIALS']['region_name']
)
s3 = boto3.resource('s3')
for bucket in s3.buckets.all():
print(bucket.name)
...however I also see the same error traceback.
Actually, this was partly answered by #John Rotenstein and #bdcloud nevertheless I need to be more specific...
The following code in my case was not necessary and causing the error message:
# Create a session
session = boto3.session.Session(
aws_access_key_id=self.settings['CREDENTIALS']['aws_access_key_id'],
aws_secret_access_key=self.settings['CREDENTIALS']['aws_secret_access_key'],
aws_session_token=self.settings['CREDENTIALS']['session_token'],
region_name=self.settings['CREDENTIALS']['region_name']
)
The credentials now stored in self.settings mirror the ~/.aws/credentials. Weirdly (and like last week where the reverse happened), I now have access. It could be that a simple reboot of my laptop meant that my new credentials (since I updated these again yesterday) in ~/.aws/credentials were then 'accepted'.
I am trying to load custom properties in my spark application using :-
command-runner.jar,spark-submit,--deploy-mode,cluster,--properties-file,s3://spark-config-test/myprops.conf,--num-executors,5,--executor-cores,2,--class,com.amazon.Main,#{input.directoryPath}/SWALiveOrderModelSpark-1.0-super.jar
However, I am getting the following exception:-
Exception in thread "main" java.lang.IllegalArgumentException: Invalid
properties file 's3://spark-config-test/myprops.conf''. at
org.apache.spark.launcher.CommandBuilderUtils.checkArgument(CommandBuilderUtils.java:241)
at
org.apache.spark.launcher.AbstractCommandBuilder.loadPropertiesFile(AbstractCommandBuilder.java:284)
at
org.apache.spark.launcher.AbstractCommandBuilder.getEffectiveConfig(AbstractCommandBuilder.java:264)
at
org.apache.spark.launcher.SparkSubmitCommandBuilder.buildSparkSubmitCommand(SparkSubmitCommandBuilder.java:233)
at org
Is this the correct way to load file from S3?
You can't load a properties file directly from S3. Instead you will need to download the properties file to your master node somewhere, then submit the spark job referencing the local path on that node. You can do the download by using command-runner.jar to run the aws cli utility.