sagecreator 0.1.1

Package to orchestrate architecture in AWS

SageCreator

SageCreator is a package meant to simplify cluster setup for Machine Learning in AWS.

It does all the heavy lifting to get cluster up and running in a matter of minutes on any of the AWS instance type(s).

It uses spot instances by default which can significantly reduce total cost of running the cluster.

If spot instances are not available or the specified spot instance price is too low, it falls back to on-demand instances.

You can access Jupyter notebook that can run your code against provisioned server(s). See Jupyter access for more info.

Installation

Install and update using pip:

$ pip install sagecreator

Python 3 is required and it is highly recommended to install and run the package in virtualenv. Supported in Python >= 3.5.0.

Prerequisites

AWS Account

To provision the cluster you need an AWS Account and an IAM user with:

• Access Key ID

• Secret Access Key

User should either be in Administrators group as described in IAM user tutorial, or create a custom IAM policy

Execution

After the installation, configure the tool by specifying configuration parameters:

$ sage configure
Access key id: <AWS Access Key ID>
Secret access key: <AWS Secret Access Key>
Company: <Name of your organization>
Owner: <Name of your team>
Key pair name: <Name of the key pair> (Optional - if NOT provided it will be created with a new private key)
Private key file: <Absolute path to private key file> (required only if Key pair name was provided)

Company, Owner, Service are required - those are used as tags for each instance in the cluster.

Key pair name, Private key file are optional - if provided, given ‘key pair name’ / ‘private key file’ will be used to provision the cluster.

---

Provision the cluster.

Provision step can take up to 20 minutes depending on the network connection, cluster size, and instance type.

$ sage provision
Service: <Name of your service>
Instance type [t3.small]: <Instance type> (Optional, defaults to t3.small)
Spot instance price [0.1]: <Spot instance price> (Optional, defaults to $0.1 per instance)
Cluster size [1]: <Cluster size> (Optional, defaults to 1 node)

Important:

The tool provides NO guarantee that the instance(s) will be provisioned at specified Spot instance price.

If specified price is lower than the current AWS spot instance price then On-demand instance(s) will be provisioned instead.

Thus, it is up to the user to ensure that specified price is high enough for the request to be fulfilled.

---

Display path of the cluster configuration file.

Though not necessary it is possible to manually edit that file with customizations prior to running provision step.

$ sage pwd

---

Terminate cluster. This operation terminates all cluster nodes matching tags tuple of Company, Owner, Service.

$ sage terminate
Service: <Name of your service to terminate>

Jupyter access

Once provisioning step is done and the cluster is up you can access jupyter notebook in your browser at http://localhost:9000.

We have provided a sample notebook to execute. It trains the model on Fashion MNIST dataset using CNN in Keras.

Under the hood

The logic that orchestrates the cluster is written in Ansible

Custom IAM policy

Alternatively, instead of assigning user to Administrators group which has access to all AWS services (as described in IAM user), you can create separate Group named Provisioners with more restrictive policy:

{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Action": "ec2:*",
      "Effect": "Allow",
      "Resource": "*"
    },
    {
      "Action": "rds:*",
      "Effect": "Allow",
      "Resource": "*"
    },
    {
      "Action": "route53:*",
      "Effect": "Allow",
      "Resource": "*"
    }
  ]
}

Then assign the user to the Provisioners group which has access to a subset of AWS services that are sufficient to orchestrate the cluster.

SSH access

If Key pair name / Private key file were NOT provided when configuring the cluster then default key pair is created and a new private key is stored locally.

In order to ssh into the servers point ssh to the correct (private key) file:

$ ssh -i <path to private key file> ubuntu@<host>

where path to private key file is ../venv/lib/python3.X/site-packages/sagebase/.ssh/pkey.pem