Cloudmesh Elastic Map Reduce
cloudmesh-emr allows cloudmesh users to easily iteract with Amazon's Elastic Map Reduce services with a particular focus on using Apache Spark. Users are able to manage the entire lifecycle of Apache Spark clusters from the commandline, including uploading data and files to the cluster, launching the analysis, and checking in on the status of tasks and clusters. The goal of cloudmesh-emr is to abstract away a lot of the details of launching and managing a cluster by including commonsense defaults while allowing the user the flexibility to specify different options, if needed.
cloudmesh-emr can either be installed via pip or, for advanced users wanting to use the development versions, via git:
Install via pip
$ pip install cloudmesh-cmd5
$ pip install cloudmesh-sys
$ pip install cloudmesh-cloud
$ pip install cloudmesh-emr
$ git clone https://github.com/cloudmesh/cloudmesh-emr.git
$ cd cloudmesh-emr
$ pip install -e .
The only external library required is boto3, which should automatically be installed by pip. Once the package is installed, the plugin requires the following keys to be present in the cloudmesh.yaml file available to Cloudmesh's configuration utility:
The first two values are provided when you create a user on your AWS account's IAM management console. The third one is user provided and must be a valid AWS region (e.g. us-west-1, eu-west-3, etc.). A full listing of AWS regions is available here. Once these keys are provided, cloudmesh-emr will be able to authenticate with the correct account and region and begin managing your EMR resources.
Finally, before beginning use of cloudmesh-emr, one must start up the local MongoDB database as cloudmesh-emr stores the results of each query into the cloudmesh collection. You can start MongoDB by using the following command:
$ cms admin mongo start
If Mongo has not been installed previously, you can install it via cloudmesh by utilizing the admin command:
$ cms admin mongo install
cloudmesh-emr provides a simple way for users to manage AWS EMR clusters.
The list subcommand provides high level details of clusters, instances, and steps (tasks). It allows the user to quickly examine all available clusters and nodes and to monitor their status. Keep in mind that cloudmesh only has access to the resources associated with the AWS Access ID and Secret Key provided in the configuration file.
$ cms emr list clusters [--status=STATUS...] [--format=FORMAT]
$ cms emr list instances CLUSTERID [--status=STATUS...] [--type=TYPE...] [--format=FORMAT]
$ cms emr list steps CLUSTERID [--state=STATE...] [--format=FORMAT]
Valid options for the
--format argument are:
table option will print a formatted table to the
dict option will print a raw dictionary. The
yaml options will save the output to the specified file format.
Valid options for the
--status flag when listing clusters are:
error. This option may be repeated multiple times to list a union of clusters
in a particular status. For example,
--status=start --status=boot will list
all clusters that are either booting or starting. If the
all flag is present,
all clusters will be listed regardless of other flags present.
When viewing instances, the
--status flag allows for:
down. Like above, the
all option will
supersede all other status flags included in the command. Otherwise, multiple
status flags may be included in order to view the union of those statuses.
--type flag in the list instances command accepts:
task. A master node manages core and tasks nodes. Task nodes are
unique in that they only provide worker threads to Spark and do not count as
individual data nodes. See the description
Finally, for the
--state flag when viewing tasks, the following options are
CLUSTERID is the AWS cluster ID, usually provided in the format similar to
The describe command queries AWS for key characteristics of a specified cluster. The command is used as follows:
$ cms emr describe CLUSTERID [--format=FORMAT]
The command will provide the cluster's state, location, type, hours, and what
applications are installed on teh cluster (typically Hadoop and Spark for EMR).
As with the list command, the
CLUSTERID is the one created by AWS to identify
your cluster. The
--format flag allows you to print the result table to the
terminal; a CSV, YAML, or JSON file; or to the terminal as a dictionary (useful
for piping to other commands).
The start command is the first step when beginning an analysis. It spins up a Spark cluster using AWS's EMR service and configures it to run PySpark applications from an S3 bucket. The command the initiate startup is:
$ cms emr start NAME [--master=MASTER] [--node=NODE] [--count=COUNT]
Once cloudmesh has initiated cluster startup, it'll return the cluster ID given to it from AWS. While you can name a cluster to keep track of them, all interactions must be done using the cluster ID rather than the name. The options for the command are:
--master=MASTER tells cloudmesh what instance type to use for the master node.
For a Spark cluster, there is only one master node. A list of available node
types are available here. By
default, cloudmesh will spin up one m3.xlarge instance for use as the master
node. Typically, this can be downscaled slightly when compared to the worker
nodes as it is solely responsible for coordinating nodes and does not
participate in the actual processing of data.
--node=NODE tells cloudmesh what instance type to use for the worker nodes. By
default, cloudmesh will spin up m3.xlarge servers to use as the worker nodes.
Since these nodes drive the analysis being run, it is worthwhile to consider
upscaling these nodes, depending on budget and predicted workload (e.g. a server
that costs twice as much is worthwhile if it does the work in a quarter of the
time). Since these nodes are using Spark, which keeps data in memory, it is
better to use nodes with lots of high bandwidth memory rather than nodes with
lots of disk space. Additional vCores are also worthwhile to extra compute
power. Imporant: Do not use a server type with 1 vCore. Spark uses one core
for the node's manager thread and then creates threads to occupy the other
cores. If the server only provides one core, then there will be no free cores to
create worker threads - you end up with a useless server that has one thread for
management and zero threads to actually do the work, which will stall work being
--count=COUNT flag tells how many servers to provision from AWS.
With 1 node being reserved for the master node,
COUNT-1 nodes will be created
as worker nodes. Note that
--count=1 will create one master node and 0 worker
nodes - meaning there will be no servers to do the actual work of the analysis.
The upload function provides the user with the ability to upload programs and data to an S3 bucket for use in the analysis. This is provided as a convenience and it is recommended that the user utilize the more formal methods of uploading files and data to S3 via the cloudmesh-storage. To use this function, simply provide the local filename, the destination S3 bucket associated with the account, and the filename to store the file as in the S3 bucket:
$ cms emr upload FILE BUCKET BUCKETNAME
Programs uploaded to S3 can be invoked via the run command while data becomes directly accessible to the Spark cluster by utilizing 'S3://' in the path to the file.
In unusual scenarios, support files may need to be copied from S3 to the servers themselves. This may improve performance as local IO is faster than reading from S3 - especially for repeated accesses. To copy a file from an S3 bucket to local, use:
$ cms emr copy BUCKET BUCKETNAME
By default, the file will be copied from the S3 bucket given and downloaded into the master node's /home/hadoop/ folder.
The run command is the primary command of cloudmesh-emr. It calls
on the master node of the given cluster and points it to the program file on an
S3 bucket. Essentially, it is equivalent to typing in
spark-submit S3://BUCKET/BUCKETNAME on the master node. You can invoke this command using:
$ cms emr run CLUSTERID BUCKET BUCKETNAME
As with the above,
CLUSTERID refers to the AWS cluster ID and the
BUCKETNAME arguments point to an S3 bucket and a file in that bucket that
contain the program to be run. The program then runs as any normal Spark
application and can use any and all Python and Spark functions normally
available. For convenience, AWS allows users to use the 'S3://' file path to
directly access S3 buckets - saving users from having to manually handle
extracting results from the servers to a separate storage solution.
While programs are running, their status can be checked on via the
As the name suggests, the stop command will initiate a shutdown of all instances associated with a cluster ID. The command is invoked as:
$ cms emr stop CLUSTERID
Once a cluster has initiated shutdown, all instances are released, attached storage is removed, and all pending or running tasks are stopped. This process is irreversible. The cluster ID will remain associated with your account for approximately 30 days after the cluster is terminated, after which is will be removed.
Cloudmesh-emr also provides an OpenAPI specification that allows users to run the plugin as a REST service. The server can be started by utilizing cloudmesh-openapi. To invoke the server, navigate to the OpenAPI folder in the cloudmesh-emr directory and use:
$ cms openapi server start ./emr.yaml
to start up the server on the default address and port. From there, users can access cloudmesh-emr's functionality via REST service calls. The functions are available as:
All of the options described in the previous section are available in the OpenAPI specification as arguments. For example, to list all worker instances associated with a cluster, the following URL can be visited:
The only thing to consider when using the OpenAPI server is that
behave slightly differently than anticipated. When utilizing that part of the
file argument will be relative to the server, not the client
making the request.