Puddle Python Client Demo

This notebook demonstrates how to use the Puddle Python client. It covers the following topics:

  • How to obtain and install the Puddle Python Client.

  • How to create a new system.

  • How to connect to a Driverless AI running on provisioned system.

  • How to stop the system.

  • How to terminate the system.

Prerequisites

This notebook was created using Python 3.6. Other versions may work with Puddle, but Driverless AI requires Python 3.6.

To download the Puddle Python Client, go to the Puddle Web UI and select Download Python Client from the user’s menu in the top right-hand corner. This downloads the Puddle wheel. Sni%CC%81mka%20obrazovky%202019-10-10%20o%2014.13.59.png

This wheel can be installed using:

pip install puddle_client-*-py3-none-any.whl

API keys can be obtained from Puddle Web UI. Select the API Keys from the user’s menu in the top right-hand corner. Sni%CC%81mka%20obrazovky%202019-10-10%20o%2014.17.01.png

The last requirement is the Driverless AI Python Client. The Driverless AI Python Client is not required in the beginning of this demo, but it is needed in order to create experiments. To download the Driverless AI Python Client, a running instance of Driverless AI is required. If no such instance is available, wait until Puddle provisions a new system. This system can then be used to download the client. Once Driverless AI is available, access the UI of Driverless AI and then on the top menu, select the RESOURCES > PYTHON CLIENT link. This downloads the h2oai_client wheel. Please make sure that you are using the correct combination of Driverless AI Python Client version and Driverless AI version.

resources.png

This wheel can be installed using:

pip install h2oai_client-*-py3-none-any.whl

Creating a new system

At first we need to import some Puddle modules.

[ ]:
import os
import time

from puddle.client import SystemsClient
from puddle.client.cloud import CLOUD_AWS, CLOUD_AZURE
from puddle.client.product import PRODUCT_DAI
from puddle.client.models.tag_data import TagData

Let’s check that all of the required environment variables are set. Puddle needs the following environment variables:

  • PUDDLE_SERVER_ADDRESS - URL where Puddle is accessible.

  • PUDDLE_API_KEY_ID - API Key ID which can be obtained if you go to the Puddle UI, then you should click the user menu in top right corner and then you should select the API Keys option.

  • PUDDLE_API_SECRET_KEY - API Secret Key which can be obtained if you go to the Puddle UI, then you should click the user menu in top right corner and then you should select the API Keys option.

If you are missing any of these, please do export <VARIABLE>=<VALUE> and reload the environment if required.

[ ]:
print("PUDDLE_SERVER_ADDRESS =", os.getenv("PUDDLE_SERVER_ADDRESS", None))
print("PUDDLE_API_KEY_ID =", os.getenv("PUDDLE_API_KEY_ID", None))
print("PUDDLE_API_SECRET_KEY =",os.getenv("PUDDLE_API_SECRET_KEY", None))

Now an instance of SystemsClient should be created. In this example we’ll be using Amazon AWS. If you want to use Microsoft Azure instead, then please replace CLOUD_AWS with CLOUD_AZURE. SystemsClient will automatically read the environment variables (which were discussed earlier) to obtain backend URL and credentials. Backend address can be overriden by setting the server_address parameter in the constructor.

[ ]:
systems_client = SystemsClient(cloud=CLOUD_AWS)

The created instance of SystemsClient is now ready to use. Let’s proceed by creating a new system.

This demo shows the easiest way that a new system can be created. There are also other parameters. If you want to explore them, please review the documentation.

The following cell will create a system named python-client-demo with Driverless AI 1.7.0 backed by the CPU Small instance type. Your Puddle instance might not have CPU Small or Driverless AI 1.7.0. Please adjust these variables to fit your Puddle instance. In this example, we are also setting the tag dept (department). We are using value eng (engineering) for this tag. Again, if your Puddle does not have this tag, please adjust the list accordingly, or omit it completly.

[ ]:
system = systems_client.create(
    name="python-client-demo",
    product=PRODUCT_DAI,
    version="1.7.0",
    instance_type_name="CPU Small",
    tags=[TagData("dept", "eng")],
)
print(system)

We need to wait for the system to be ready. This could take up to 15 minutes, but it’s usually completed sooner.

[ ]:
systems_client.wait_for_system_started(system.id)

Let’s observe the system once again. It should be up and running now.

[ ]:
system = systems_client.get(system.id)
print(system)

Running an Experiment

Now we need to import some Driverless AI modules. Then we will be able to load a dataset and run an experiment. Please make sure that the correct version of Driverless AI Client is installed. If not please refer to Prerequisites section of this demo.

[ ]:
from h2oai_client import Client

Let’s connect to Driverless AI. We’ll use system from Puddle to provide address and credentials.

[ ]:
h2oai = Client(address=system.default_url, username=system.username, password=system.password)

For purpose of this demo, we will be using the IRIS dataset.

[ ]:
train = h2oai.create_dataset_from_s3_sync("https://s3.amazonaws.com/h2o-public-test-data/smalldata/extdata/iris_wheader.csv")

Everything is ready now, and we can start an experiment.

[ ]:
experiment = h2oai.start_experiment_sync(dataset_key=train.key,
                                         target_col="class",
                                         is_classification=True,
                                         accuracy=1,
                                         time=1,
                                         interpretability=1,
                                         scorer="AUC",
                                         enable_gpus=False,
                                         seed=1234)

Stopping the System

Once the experiment is finished, we should deallocate (stop) the system. This does not destroy anything. The system can be restarted again and all data will be there.

[ ]:
systems_client.stop(system.id)
systems_client.wait_for_system_stopped(system.id)
system = systems_client.get(system.id)
print(system)

For purpose of this demo we don’t need to restart the system, but if you need to restart it, you can use the following snippet:

systems_client.start(system.id)
systems_client.wait_for_system_started(system.id)
system = systems_client.get(system.id)
print(system)

Terminating the System

For purpose of this demo we don’t need to do anything else with this system, so we can terminate it. This operation cannot be reverted. Everything is gone forever, unless you saved it elsewhere.

[ ]:
systems_client.terminate(system.id)
systems_client.wait_for_system_terminated(system.id)
system = systems_client.get(system.id)
print(system)