Data Science Pipelines Operator

The Data Science Pipelines Operator (DSPO) is an OpenShift Operator that is used to deploy single namespace scoped Data Science Pipeline stacks onto individual OCP namespaces.

Overview

Data Science Pipelines (DSP) allows data scientists to track progress as they iterate over development of ML models. With DSP, a data scientist can create workflows for data preparation, model training, model validation, and more. They can create and track experiements to arrive at the best version of of training data, model hyperparameters, model code, etc., and repeatably rerun these experiments.

Data Science Pipelines is based on the upstream Kubeflow Pipelines (KFP) project.

Data Scientists can use tools like the kfp SDK or Elyra to author their workflows, and interact with them in the ODH dashboard.

Quickstart

To get started you will first need to satisfy the following pre-requisites:

Pre-requisites

An OpenShift cluster that is 4.11 or higher.
You will need to be logged into this cluster as cluster admin via oc client.
Based on which DSP version to install you will need to do the following:
1. For DSPv1: The OpenShift Cluster must have OpenShift Pipelines 1.8 or higher installed. We recommend channel pipelines-1.8 on OCP 4.10 and pipelines-1.9 or pipelines-1.10 for OCP 4.11, 4.12 and 4.13. Instructions here.
2. For DSPv2: The DSPO will install a namespace-scoped Argo Workflow Controller, so ensure the OpenShift Cluster does not have a cluster-scoped Argo Workflows installation already present.
Based on installation type you will need one of the following:
1. For Standalone method: You will need to have Kustomize version 4.5+ installed
2. For ODH method: The Open Data Hub operator needs to be installed. You can install it via OperatorHub.

Deploy the Operator via ODH

Use this method to deploy DSPO using the OpenDataHub operator

On a cluster with ODH installed, create a namespace where you would like to install DSPO: Deploy the following DataScienceCluster:

ODH_NS=opendatahub
oc new-project ${ODH_NS}

Then deploy the following DataScienceCluster into the namespace created above:

cat <<EOF | oc apply -f -
kind: DataScienceCluster
apiVersion: datasciencecluster.opendatahub.io/v1
metadata:
  name: data-science-cluster
  namespace: ${ODH_NS}
spec:
  components:
    dashboard:
      managementState: Managed
    datasciencepipelines:
      managementState: Managed
EOF

ℹ️ Note:

You can also deploy other ODH components using DataScienceCluster`. See https://github.com/opendatahub-io/opendatahub-operator#example-datasciencecluster for more information.

Confirm the pods are successfully deployed and reach running state:

oc get pods -n ${ODH_NS}

Once all pods are ready, we can proceed to deploying the first Data Science Pipelines (DSP) instance. Instructions here.

Using a development image

You can use custom manifests from a branch or tag to use a different Data Science Pipelines Operator image. To do so, modify the IMAGES_DSPO config in config/base/params.env and push the changes to a branch or tag.

Create a (or edit the existent) DataSciencePipelines adding devFlags.manifests with the URL of your branch or tag. For example, given the following repository and branch:

Repository: https://github.com/a_user/data-science-pipelines-operator
Branch: my_branch

The DataSciencePipelines YAML should look like:

kind: DataScienceCluster
apiVersion: datasciencecluster.opendatahub.io/v1
metadata:
   name: data-science-cluster
   namespace: ${ODH_NS}
spec:
  components:
    dashboard:
      managementState: Managed
    datasciencepipelines:
      managementState: Managed
      devFlags:
        manifests:
          - uri: https://github.com/a_user/data-science-pipelines-operator/tarball/my_branch
            contextDir: config
            sourcePath: base

Deploy the standalone Operator

Use this method to deploy DSPO without OpenDataHub operator

First clone this repository:

WORKING_DIR=$(mktemp -d)
git clone https://github.com/opendatahub-io/data-science-pipelines-operator.git ${WORKING_DIR}

If you already have the repository cloned, set WORKING_DIR= to its absolute location accordingly.

DSPO can be installed in any namespace, we'll deploy it in the following namespace, you may update this environment variable accordingly.

ODH_NS=opendatahub

# Create the namespace if it doesn't already exist
oc new-project ${ODH_NS}

Now we will navigate to the DSPO manifests then build and deploy them to this namespace.

Run the following to deploy the ful DSPO v2 stack.

cd ${WORKING_DIR}
make deploy OPERATOR_NS=${ODH_NS}

Confirm the pods are successfully deployed and reach running state:

oc get pods -n ${ODH_NS}

Once all pods are ready, we can proceed to deploying the first Data Science Pipelines (DSP) instance. Instructions here.

Deploy DSP instance

We'll deploy the first instance in the following namespace.

# You may update this value to the namespace you would like to use.
DSP_Namespace=test-ds-project-1
oc new-project ${DSP_Namespace}

DSPO introduces a new custom resource to your cluster called DataSciencePipelinesApplication. This resource is where you configure your DSP components, your DB and Storage configurations, and so forth. For now, we'll use the sample one provided, feel free to inspect this sample resource to see other configurable options.

cd ${WORKING_DIR}/config/samples/v2/dspa-simple
kustomize build . | oc -n ${DSP_Namespace} apply -f -

Note: the sample CR used here deploys a minio instance so DSP may work out of the box this is unsupported in production environments and we recommend to provide your own object storage connection details via spec.objectStorage.externalStorage see ${WORKING_DIR}/config/samples/v2/external-object-storage/dspa.yaml for an example.

Confirm all pods reach ready state by running:

oc get pods -n ${DSP_Namespace}

For instructions on how to use this DSP instance refer to these instructions: here.

Deploy another DSP instance

You can use the DSPO to deploy multiple DataSciencePipelinesApplication instances in different OpenShift namespaces, for example earlier we deployed a DataSciencePipelinesApplication resource named sample. We can use this again and deploy it to a different namespace:

DSP_Namespace_2=test-ds-project-2
oc new-project ${DSP_Namespace_2}
cd ${WORKING_DIR}/config/samples/v2/dspa-simple
kustomize build . | oc -n ${DSP_Namespace_2} apply -f -

Deploy a DSP with custom credentials

Using DSPO you can specify custom credentials for Database and Object storage. If specifying external connections, this is required. You can also provide secrets for the built in MariaDB and Minio deployments. To see a sample configuration you can simply investigate and deploy the following path:

DSP_Namespace_3=test-ds-project-3
oc new-project ${DSP_Namespace_3}
cd ${WORKING_DIR}/config/samples/v2/custom-configs
kustomize build . | oc -n ${DSP_Namespace_3} apply -f -

Notice the introduction of 2 secrets testdbsecret, teststoragesecret and 2 configmaps custom-ui-configmap and custom-artifact-script. The secrets allow you to provide your own credentials for the DB and MariaDB connections.

These can be configured by the end user as needed.

Deploy a DSP with external Object Storage

To specify a custom Object Storage (example an AWS s3 bucket) you will need to provide DSPO with your S3 credentials in the form of a k8s Secret, see an example of such a secret here config/samples/v2/external-object-storage/storage-creds.yaml.

DSPO can deploy a DSPA instance and use this S3 bucket for storing its metadata and pipeline artifacts. A sample configuration for a DSPA that does this is found in config/samples/v2/external-object-storage, you can update this as needed, and deploy this DSPA by running the following:

DSP_Namespace_3=test-ds-project-4
oc new-project ${DSP_Namespace_4}
cd ${WORKING_DIR}/config/samples/v2/external-object-storage
kustomize build . | oc -n ${DSP_Namespace_3} apply -f -

DataSciencePipelinesApplication Component Overview

When a DataSciencePipelinesApplication is deployed, the following components are deployed in the target namespace:

APIServer
Persistence Agent
Scheduled Workflow controller

If specified in the DataSciencePipelinesApplication resource, the following components may also be additionally deployed:

MariaDB
Minio
MLPipelines UI
MLMD (ML Metadata)

To understand how these components interact with each other please refer to the upstream Kubeflow Pipelines Architectural Overview documentation.

Deploying Optional Components

MariaDB

To deploy a standalone MariaDB metadata database (rather than providing your own database connection details), simply add a mariaDB item under the spec.database in your DSPA definition with an deploy key set to true. All other fields are defaultable/optional, see All Fields DSPA Example for full details. Note that this component is mutually exclusive with externally-provided databases (defined by spec.database.externalDB).

apiVersion: datasciencepipelinesapplications.opendatahub.io/v1
kind: DataSciencePipelinesApplication
metadata:
  name: sample
spec:
   ...
  database:
    mariaDB:   # mutually exclusive with externalDB
      deploy: true

Minio

To deploy a Minio Object Storage component (rather than providing your own object storage connection details), simply add a minio item under the spec.objectStorage in your DSPA definition with an image key set to a valid minio component container image. All other fields are defaultable/optional, see All Fields DSPA Example for full details. Note that this component is mutually exclusive with externally-provided object stores (defined by spec.objectStorage.externalStorage).

apiVersion: datasciencepipelinesapplications.opendatahub.io/v1
kind: DataSciencePipelinesApplication
metadata:
  name: sample
spec:
   ...
  objectStorage:
    minio:  # mutually exclusive with externalStorage
      deploy: true
      # Image field is required
      image: 'quay.io/opendatahub/minio:RELEASE.2019-08-14T20-37-41Z-license-compliance'

ML Pipelines UI

To deploy the standalone DS Pipelines UI component, simply add a spec.mlpipelineUI item to your DSPA with an image key set to a valid ui component container image. All other fields are defaultable/optional, see All Fields DSPA Example for full details.

apiVersion: datasciencepipelinesapplications.opendatahub.io/v1
kind: DataSciencePipelinesApplication
metadata:
  name: sample
spec:
   ...
  mlpipelineUI:
    deploy: true
    # Image field is required
    image: 'quay.io/opendatahub/odh-ml-pipelines-frontend-container:beta-ui'

ML Metadata

To deploy the ML Metadata artifact linage/metadata component, simply add a spec.mlmd item to your DSPA with deploy set to true. All other fields are defaultable/optional, see All Fields DSPA Example for full details.

apiVersion: datasciencepipelinesapplications.opendatahub.io/v1
kind: DataSciencePipelinesApplication
metadata:
  name: sample
spec:
   ...
   mlmd:
      deploy: true

Using a DataSciencePipelinesApplication

When a DataSciencePipelinesApplication is deployed, use the MLPipelines UI endpoint to interact with DSP, either via a GUI or via API calls.

You can retrieve this route by running the following in your terminal:

DSP_CR_NAME=sample
DSP_Namespace=test-ds-project-1
echo https://$(oc get routes -n ${DSP_Namespace} ds-pipeline-ui-${DSP_CR_NAME} --template={{.spec.host}})

Using the Graphical UI

Note the UI presented below is the upstream Kubeflow Pipelines UI, this is not supported in DSP and will be replaced with the ODH Dashboard UI. Until then, this UI can be deployed via DSPO for experimentation/development purposes. Note however that this UI is not a supported feature of DSPO/ODH.

Navigate to the route retrieved in the last step. You will be presented with the MLPipelines UI. In this walkthrough we will upload a pipeline and start a run based off it.

To start, click the "Upload Pipeline" button.

Choose a file, you can use the flipcoin example. Download this example and select it for the first step. Then click "Create" for the second step.

Once done, you can now use this Pipeline to create a Run, do this by pressing "+ Create Run".

On this page specify the Pipeline we just uploaded if it's not already auto selected. Similarly with the version, with the example there will only be one version listed. Give this Run a name, or keep the default as is. Then click "Start".

Once you click start you will be navigated to the Runs page where you can see your previous runs that are have been executed, or currently executing. You can click the most recently started Run to view it's execution graph.

You should see something similar to the following once the Run completes.

Click the first "flip-coin" step. This step produces an output message either "heads" or "tails", confirm that you can see these logs after clicking this step and navigating to "Logs."

Using the API

Note: By default we use kfp-tekton 1.5.x for this section so you will need kfp-tekton v1.5.x sdk installed in your environment

In the previous step we submitted a generated Pipeline yaml via the GUI. We can also submit the Pipeline code directly either locally or via a notebook.

You can find the Pipeline code example here. We can submit this to the DSP API Server by including this code in the following Python script:

cd ${WORKING_DIR}/docs/example_pipelines
touch execute_pipeline.py

We can utilize the flip coin example in this location and submit it directly to the API server by doing the following:

# Add this to execute_pipeline.py we created earlier
import os
import kfp_tekton
from condition import flipcoin_pipeline
token = os.getenv("OCP_AUTH_TOKEN")
route = os.getenv("DSP_ROUTE")
client = kfp_tekton.TektonClient(host=route, existing_token=token)

client.create_run_from_pipeline_func(pipeline_func=flipcoin_pipeline, arguments={})

Note: If you are in an unsecured cluster, you may encounter CERTIFICATE_VERIFY_FAILED error, to work around this you can pass in the self-signed certs to the kfp-tekton client. For instance, if running inside a notebook or pod on the same cluster, you can do the following:

...
cert = "/run/secrets/kubernetes.io/serviceaccount/ca.crt"
client = kfp_tekton.TektonClient(host=route, existing_token=token, ssl_ca_cert=cert)
client.create_run_from_pipeline_func(pipeline_func=flipcoin_pipeline, arguments={})

Retrieve your token and DSP route:

# This is the namespace you deployed the DataSciencePipelinesApplication Custom Resource
DSP_Namespace=test-ds-project-1
# This is the metadata.name of that DataSciencePipelinesApplication Custom Resource
DSP_CR_NAME=sample
export DSP_ROUTE="https://$(oc get routes -n ${DSP_Namespace} ds-pipeline-ui-${DSP_CR_NAME} --template={{.spec.host}})"
export OCP_AUTH_TOKEN=$(oc whoami --show-token)

And finally execute this script and submit the flip coin example:

python execute_pipeline.py

You can navigate to the UI again and find your newly created run there, or you could amend the script above and list the runs via client.list_runs().

Cleanup

To remove a DataSciencePipelinesApplication from your cluster, run:

# Replace environment variables accordingly
oc delete ${YOUR_DSPIPELINE_NAME} -n ${YOUR_DSPIPELINES_NAMESPACE}

The DSPO will clean up all manifests associated with each DataSciencePipelinesApplication instance.

Or you can remove all DataSciencePipelinesApplication instances in your whole cluster by running the following:

oc delete DataSciencePipelinesApplication --all -A

Depending on how you installed DSPO, follow the instructions below accordingly to remove the operator:

Cleanup ODH Installation

To uninstall DSPO via ODH run the following:

DSC_NAME=$(oc get DataScienceCluster -o jsonpath='{.items[0].metadata.name}')
ODH_NS=$(oc get DataScienceCluster -o jsonpath='{.items[0].metadata.namespace}')
oc delete datasciencecluster ${DSC_NAME} -n "${ODH_NS}"

Cleanup Standalone Installation

To clean up standalone DSPO deployment:

# WORKING_DIR must be the root of this repository's clone
cd ${WORKING_DIR}
make undeploy OPERATOR_NS=${ODH_NS}
oc delete project ${ODH_NS}

Run tests

See .github/scripts/tests/README.md(https://github.com/opendatahub-io/data-science-pipelines-operator/blob/main/.github/scripts/tests/README.md)

Metrics

The Data Science Pipelines Operator exposes standard operator-sdk metrics for controller monitoring purposes.
In addition to these metrics, DSPO also exposes several custom metrics for monitoring the status of the DataSciencePipelinesApplications that it owns.

They are as follows:

data_science_pipelines_application_apiserver_ready - Gauge that indicates if the DSPA's APIServer is in a Ready state (1 => Ready, 0 => Not Ready)
data_science_pipelines_application_persistenceagent_ready - Gauge that indicates if the DSPA's PersistenceAgent is in a Ready state (1 => Ready, 0 => Not Ready)
data_science_pipelines_application_scheduledworkflow_ready - Gauge that indicates if the DSPA's ScheduledWorkflow manager is in a Ready state (1 => Ready, 0 => Not Ready)
data_science_pipelines_application_ready - Gauge that indicates if the DSPA is in a fully Ready state (1 => Ready, 0 => Not Ready)

Configuring Log Levels for the Operator

By default, the operator's log messages are set to info severity. If you wish to adjust the log verbosity, you can do so by modifying the ZAP_LOG_LEVEL parameter in params.env file to your preferred severity level.

For a comprehensive list of available values, please consult the Zap documentation.

Deployment and Testing Guidelines for Developers

To build the DSPO locally :

Login oc using following command:

oc login

Install CRDs using the following command:

kubectl apply -k config/crd

Execute the following command:

go build main.go

To run the DSPO locally :

Execute the following command:

go run main.go --config=$config

Below is the sample config file (the tags for the images can be edited as required):

Images:
  ApiServer: quay.io/opendatahub/ds-pipelines-api-server:latest
  Artifact: quay.io/opendatahub/ds-pipelines-artifact-manager:latest
  OAuthProxy: registry.redhat.io/openshift4/ose-oauth-proxy:v4.12.0
  PersistenceAgent: quay.io/opendatahub/ds-pipelines-persistenceagent:latest
  ScheduledWorkflow: quay.io/opendatahub/ds-pipelines-scheduledworkflow:latest
  Cache: registry.access.redhat.com/ubi8/ubi-minimal
  MoveResultsImage: registry.access.redhat.com/ubi8/ubi-micro
  MariaDB: registry.redhat.io/rhel8/mariadb-103:1-188
  MlmdEnvoy: quay.io/opendatahub/ds-pipelines-metadata-envoy:latest
  MlmdGRPC: quay.io/opendatahub/ds-pipelines-metadata-grpc:latest
  MlmdWriter: quay.io/opendatahub/ds-pipelines-metadata-writer:latest

To build your own images :

All the component images are available here and for thirdparty images here. Build these images from root as shown in the below example:

podman build . -f backend/Dockerfile -t quay.io/your_repo/dsp-apiserver:sometag

To run the tests:

Execute make test or make unittest or make functest based on the level of testing as mentioned below:

make unittest is a command that is often used to run only unit tests for individual units or components. These tests verify that each unit of code (e.g., functions or methods) behaves as expected. It is suggested to run this command often during the development process.

make functest is a command used to run functional tests which assess the overall functionality of the software by testing its features and user interactions. Functional tests help ensure that the software works as a whole and that its features are functioning correctly from a user's perspective. It is suggested to run make functest for every commit.

The specific tests that are executed when you run make test can include unit tests, functional tests and more. It is a test to check if the software behaves correctly and meets the desired quality standards. It helps identify and fix issues early in the development process. It is suggested to run make test before creating a PR.

To deploy DSPO as a developer :

Follow the instructions from here to deploy the standalone operator.

Follow the instructions from here to deploy the operator via ODH.

How to deploy with a custom image:

Run the following command using the custom image:

make deploy IMG=my-registry/my-operator:v1

How to regenerate manifests:

After updating the Kubebuilder annotations in your code, run the following command to regenerate code and manifests:

make generate manifests

How to regenerate crd on api changes:

After making your API changes, run the following command to regenerate code and CRDs based on your updated API definitions:

make generate

Refer to kubebuilder docs here for more info.

How to run pre-commit tests:

Install pre-commit following the instructions here. Before creating a PR, developers should run the following command which will auto fix any simple errors:

pre-commit run --all-files

How to do disable health checks when dev testing:

To disable the health checks set the values to true in the DSPA yaml file you apply. Refer to his sample file here.

In certain scenarios, it may be necessary to disable health checks within our environment. When the DSPO is executed either locally or on a different cluster, the health checks can't reach the database and Object Store endpoints. Consequently, they remain unsuccessful, preventing the deployment of essential pipeline infrastructure components by the DSPA. To address this challenge, we have introduced the disableHealthCheck mechanism as a viable solution.

How to enable kfp ui and minio:

Refer to this sample yaml file for enabling the upstream kubeflow pipelines ui and minio.

Refer to this repo to see examples of different pipelines for dev testing.

Releases

Refer to RELEASE.md for details on the release process

Name		Name	Last commit message	Last commit date
Latest commit History 1,098 Commits
.github		.github
api		api
config		config
controllers		controllers
datasciencecluster		datasciencecluster
docs		docs
hack		hack
scripts/release		scripts/release
tests		tests
.dockerignore		.dockerignore
.gitignore		.gitignore
.gitleaks.toml		.gitleaks.toml
.golangci.yaml		.golangci.yaml
.pre-commit-config.yaml		.pre-commit-config.yaml
.yamllint.yaml		.yamllint.yaml
Dockerfile		Dockerfile
LICENSE		LICENSE
Makefile		Makefile
OWNERS		OWNERS
PROJECT		PROJECT
README.md		README.md
go.mod		go.mod
go.sum		go.sum
main.go		main.go

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

Data Science Pipelines Operator

Table of Contents

Overview

Quickstart

Pre-requisites

Deploy the Operator via ODH

Using a development image

Deploy the standalone Operator

Deploy DSP instance

Deploy another DSP instance

Deploy a DSP with custom credentials

Deploy a DSP with external Object Storage

DataSciencePipelinesApplication Component Overview

Deploying Optional Components

MariaDB

Minio

ML Pipelines UI

ML Metadata

Using a DataSciencePipelinesApplication

Using the Graphical UI

Using the API

Cleanup

Cleanup ODH Installation

Cleanup Standalone Installation

Run tests

Metrics

Configuring Log Levels for the Operator

Deployment and Testing Guidelines for Developers

Releases

About

Releases

Packages

Languages

License

gregsheremeta/data-science-pipelines-operator

Folders and files

Latest commit

History

Repository files navigation

Data Science Pipelines Operator

Table of Contents

Overview

Quickstart

Pre-requisites

Deploy the Operator via ODH

Using a development image

Deploy the standalone Operator

Deploy DSP instance

Deploy another DSP instance

Deploy a DSP with custom credentials

Deploy a DSP with external Object Storage

DataSciencePipelinesApplication Component Overview

Deploying Optional Components

MariaDB

Minio

ML Pipelines UI

ML Metadata

Using a DataSciencePipelinesApplication

Using the Graphical UI

Using the API

Cleanup

Cleanup ODH Installation

Cleanup Standalone Installation

Run tests

Metrics

Configuring Log Levels for the Operator

Deployment and Testing Guidelines for Developers

Releases

About

Resources

License

Stars

Watchers

Forks

Releases

Packages 0

Languages

Packages