Ctrl+K

Kubeflow

Kubeflow#

Important

The kubeflow integration requires soopervisor>=0.7

Important

The Kubeflow tutorial is in beta! Got questions or found issues? Reach out to us on Slack.

This tutorial shows you how to export a Ploomber pipeline to Kubeflow.

If you encounter any issues with this tutorial, let us know.

Note

This tutorial uses cloud storage (S3 or Google Cloud Storage). In addition, it runs on the local cluster local storage for faster data fetch.

Pre-requisites#

Note

When installing Kubeflow, you must use a strong enough VM to meet the basic requirements. This tutorial assumes the Kubeflow is configured and you’re running within this cluster. Another option is to run the tutorial locally and upload the final ploomber_pipeline.yaml to Kubeflow.

Instructions#

First, let’s install ploomber:

pip install ploomber

Let’s now pull some sample code:

# get example
ploomber examples -n templates/ml-intermediate -o ml-intermediate
cd ml-intermediate

Since each task executes in a different Docker container, we have to configure cloud storage for tasks to share data. Modify the environment.yml file and add the appropriate dependency:

# content...
- pip:
  # dependencies...

  # add your dependency here
  - boto3 # if you want to use S3
  - google-cloud-storage # if you want to use Google Cloud Storage

We also need to configure the pipeline to use cloud storage, open the pipeline.yaml file, and add the following next to the meta section.

meta:
    # some content...

clients:
    File: clients.get_s3

meta:
    # some content...

clients:
    File: clients.get_gcloud

Now, edit the clients.py file, you only need to change the bucket_name parameter for the corresponding function. For example, if using a bucket with the name bucket-name and S3, clients.py should look like this:

from ploomber.clients import S3Client

def get_s3():
    return S3Client(bucket_name='bucket-name',
                    parent='ml-intermediate',
                    json_credentials_path='credentials.json')

from ploomber.clients import GCloudStorageClient

def get_gcloud():
    return GCloudStorageClient(bucket_name='bucket-name',
                                parent='ml-online',
                                json_credentials_path='credentials.json')

To authenticate to the cloud storage service, add a credentials.json file in the project root (the same folder that has the environment.yml file.

{
    "aws_access_key_id": "YOUR-ACCESS-KEY-ID",
    "aws_secret_access_key": "YOU-SECRET-ACCESS-KEY"
}

{
    "type": "service_account",
    "project_id": "project-id",
    "private_key_id": "private-key-id",
    "private_key": "private-key",
    "client_email": "client-email",
    "client_id": "client-id",
    "auth_uri": "https://accounts.google.com/o/oauth2/auth",
    "token_uri": "https://oauth2.googleapis.com/token",
    "auth_provider_x509_cert_url": "https://www.googleapis.com/oauth2/v1/certs",
    "client_x509_cert_url": "https://www.googleapis.com/robot/v1/metadata/x509/service-account.iam.gserviceaccount.com"
}

Note

When running with local storage, Kubeflow takes the product path and name and stores it accordingly to its temporary outputs folder. You will need to set the path of the products in the pipeline.yaml (the file that defines the pipeline - above) to: product: '{{root}}/product/data/get.parquet' note how the product name appears in the path (mandatory). In addition we can see below that for the nb and model products we set the path according to their names.

This is how your pipeline.yaml file should look like if you’re using GCP:

meta:
  jupyter_functions_as_notebooks: True
  import_tasks_from: partial.features.yaml

clients:
  File: clients.get_gcloud

tasks:
  - source: tasks.get.fn
    name: get
    product: '{{root}}/product/data/get.parquet'
    params:
      sample: '{{sample}}'

  - source: fit.py
    name: fit
    product:
        nb: '{{root}}/nb/data/nb.html'
        model: '{{root}}/model/data/model.pickle'

Important

Make sure to adjust those files to have the same output structure: partial.features.yaml and pipeline.serve.yaml.

In addition, we also need to configure the env files to the right root location. Make sure in these files env.local.yaml, env.serve.yaml and env.yaml the root is set to 'outputs' in a similar fashion to the env.yaml below:

root: 'outputs'
sample: False

Let’s now create the virtual environment:

# configure environment
conda env create --file environment.yml

# activate environment
conda activate ml-intermediate

# generate lock file
conda env export --no-build --file environment.lock.yml

Let’s now verify that everything is configured correctly:

ploomber status

We now export the pipeline to Kubeflow:

soopervisor add train --backend kubeflow

Note

You don’t have to install soopervisor manually; it should’ve been installed when running ploomber install. If missing, install it with pip install soopervisor.

soopervisor add creates a few new files. Let’s configure soopervisor.yaml which controls some settings:

train:
  backend: kubeflow
  # we will be using a remote docker hub, we'll set this image name
  repository: idomic/general:kubeflow1
  # make sure our credentials are included when building the image
  include: [credentials.json]

Note

See how the repository is configured to the docker hub idomic/general and the image within it kubeflow1 (: separated)

Build the Docker image (takes a few minutes the first time):

soopervisor export train

Once the export process finishes, you’ll see a new train/ folder with three files: Dockerfile which is the file used to build the docker image, a kubeflow_pipeline.py which is the pythonic version of the pipeline (you can run it directly on a notebook instance within the cluster) and the file ploomber_pipeline.yaml which contains the kubeflow pipeline to run. To deploy, go to your cluster and upload the .yaml file as a new pipeline (you can also use the CLI directly (we’ll cover both options).

Option 1: UI upload of a Kubeflow pipeline#

Let’s go to the cluster and click on Pipelines (top left) and then on the top right on + Upload pipeline. (see image below)

upload_pipeline

We now can name our pipeline ml_intermidiate, describe it (or copy the name to it), click on upload file and choose file, pick the ploomber_pipeline.yaml we just created. On the bottom click on Create. (see image below)

create_pipeline

Now we can see that the pipeline is configured, we can see each step and the dependencies, we’ll need to submit our first pipeline run. To do that, click on + Create run

create_run

The run details should be filled automatically, if not give a run name and the other missing details. On the bottom, click on Start.

start_run

Now you can watch the pipeline execution by clicking on the run you’ve created. When the tasks are ready, you can view each task’s inputs and outputs, click on the task, and then on the Input/Output tab. The links contain the raw files.

task_input_output

Option 2: CLI upload of a Kubeflow pipeline#

We need to make sure we have a working notebook on the cluster, we can open it and upload/copy the content of the kubeflow_pipeline.py file. Make sure to uncomment the client rows below (6, 7, 10, 135). We’ll also need to update the kfp endpoint to your cluster and port. Once set we can run the notebook and click on the run link below to get to the pipeline and its run details.

kfp_endpoint="YOUR_KFP_ENDPOINT"
client = kfp.Client(kfp_endpoint)

# This is a sanity check to make sure your notebook and cluster can communicate
print(client.list_experiments())from ploomber.clients import S3Client
client.create_run_from_pipeline_func(ml_intermediate, arguments={})
Contents