Skip to content

Latest commit

 

History

History
250 lines (167 loc) · 10.5 KB

README.md

File metadata and controls

250 lines (167 loc) · 10.5 KB
Raw

Kubeflow Pipelines SDK for Tekton

The Kubeflow Pipelines SDK allows data scientists to define end-to-end machine learning and data pipelines. The output of the Kubeflow Pipelines SDK compiler is YAML for Argo.

The kfp-tekton SDK is extending the Compiler and the Client of the Kubeflow Pipelines SDK to generate Tekton YAML and to subsequently upload and run the pipeline with the Kubeflow Pipelines engine backed by Tekton.

Table of Contents

Extensions to the Kubeflow Pipelines SDK

In addition to the functionality provided by the Kubeflow Pipelines SDK the kfp-tekton SDK provides a TektonCompiler and a TektonClient:

TektonCompiler:

  • kfp_tekton.compiler.TektonCompiler.compile compiles Python DSL code into a YAML file containing a Tekton PipelineRun which can be deployed directly to a Tekton enabled Kubernetes cluster or uploaded to the Kubeflow Pipelines dashboard with the Tekton backend.

TektonClient:

  • kfp_tekton.TektonClient.create_run_from_pipeline_func compiles DSL pipeline function and runs the pipeline on a Kubernetes cluster with KFP and Tekton

Project Prerequisites

Follow the instructions for installing project prerequisites and take note of some important caveats.

Installation

You can install the latest release of the kfp-tekton compiler from PyPi. We recommend to create a Python virtual environment first:

python3 -m venv .venv
source .venv/bin/activate

pip install kfp-tekton

Alternatively you can install the latest version of the kfp-tekton compiler from source by cloning the repository https://github.com/kubeflow/kfp-tekton:

  1. Clone the kfp-tekton repo:

    git clone https://github.com/kubeflow/kfp-tekton.git
cd kfp-tekton

  2. Setup Python environment with Conda or a Python virtual environment:

    python3 -m venv .venv
source .venv/bin/activate

  3. Build the compiler:

    pip install -e sdk/python

  4. Run the compiler tests (optional):

    make test

Compiling a Kubeflow Pipelines DSL Script

The kfp-tekton Python package comes with the dsl-compile-tekton command line executable, which should be available in your terminal shell environment after installing the kfp-tekton Python package.

If you cloned the kfp-tekton project, you can find example pipelines in the samples folder or under sdk/python/tests/compiler/testdata folder.

dsl-compile-tekton \
    --py sdk/python/tests/compiler/testdata/parallel_join.py \
    --output pipeline.yaml

Note: If the KFP DSL script contains a __main__ method calling the kfp_tekton.compiler.TektonCompiler.compile() function:

if __name__ == "__main__":
    from kfp_tekton.compiler import TektonCompiler
    TektonCompiler().compile(pipeline_func, "pipeline.yaml")

... then the pipeline can be compiled by running the DSL script with python3 executable from a command line shell, producing a Tekton YAML file pipeline.yaml in the same directory:

python3 pipeline.py

Running the Compiled Pipeline on a Tekton Cluster

After compiling the sdk/python/tests/compiler/testdata/parallel_join.py DSL script in the step above, we need to deploy the generated Tekton YAML to our Kubernetes cluster with kubectl. The Tekton server will automatically start a pipeline run for which we can follow the logs using the tkn CLI.

Here we have to deploy the pipeline in the kubeflow namespace because all the pipelines with metadata and artifacts tracking rely on the minio object storage credentials in the kubeflow namespace.

kubectl apply -f pipeline.yaml -n kubeflow

tkn pipelinerun logs --last -n kubeflow

Once the Tekton Pipeline is running, the logs should start streaming:

Waiting for logs to be available...

[gcs-download : main] With which he yoketh your rebellious necks Razeth your cities and subverts your towns And in a moment makes them desolate

[gcs-download : copy-artifacts] Added `storage` successfully.
[gcs-download : copy-artifacts] tar: removing leading '/' from member names
[gcs-download : copy-artifacts] tekton/results/data
[gcs-download : copy-artifacts] `data.tgz` -> `storage/mlpipeline/artifacts/parallel-pipeline/gcs-download/data.tgz`
[gcs-download : copy-artifacts] Total: 0 B, Transferred: 195 B, Speed: 1 B/s

[gcs-download-2 : main] I find thou art no less than fame hath bruited And more than may be gatherd by thy shape Let my presumption not provoke thy wrath

[gcs-download-2 : copy-artifacts] Added `storage` successfully.
[gcs-download-2 : copy-artifacts] tar: removing leading '/' from member names
[gcs-download-2 : copy-artifacts] tekton/results/data
[gcs-download-2 : copy-artifacts] `data.tgz` -> `storage/mlpipeline/artifacts/parallel-pipeline/gcs-download-2/data.tgz`
[gcs-download-2 : copy-artifacts] Total: 0 B, Transferred: 205 B, Speed: 1 B/s

[echo : main] Text 1: With which he yoketh your rebellious necks Razeth your cities and subverts your towns And in a moment makes them desolate
[echo : main]
[echo : main] Text 2: I find thou art no less than fame hath bruited And more than may be gatherd by thy shape Let my presumption not provoke thy wrath
[echo : main]

Building Tekton from Master

In order to utilize the latest features and functions of the kfp-tekton compiler, we suggest to install Tekton from a nightly built or build it from the master branch. Features that require a special build, different from the 'Tested Version', will be listed below.

Optional Features

Compile Kubeflow Pipelines without Artifacts

By default, artifacts are enabled because the KFP DSL are designed to run on Kubeflow Pipeline's engine with artifacts to be stored on Minio storage. When artifacts are enabled, all the output parameters are also treated as artifacts and persisted to the default object storage. Enabling artifacts also allows files to be downloaded or stored as artifact inputs/outputs. Since artifacts are dependent on the Kubeflow Pipeline's deployment, the generated Tekton pipeline must be deployed to the same namespace as Kubeflow Pipelines.

To run Tekton pipelines without installing Kubeflow pipeline, or if you need to compile the Kubeflow Pipelines DSL without artifacts, add the --disable-artifacts argument to your dsl-compile-tekton commands. Then, run the pipeline in the same namespace that is used by Kubeflow Pipelines (typically kubeflow) by specifying the -n flag:

dsl-compile-tekton \
    --py sdk/python/tests/compiler/testdata/parallel_join.py \
    --output pipeline.yaml \
    --disable-artifacts
    
kubectl apply -f pipeline.yaml -n kubeflow

tkn pipelinerun logs --last -n kubeflow

You should see log messages without any artifact reference:

Waiting for logs to be available...

[gcs-download : main] With which he yoketh your rebellious necks Razeth your cities and subverts your towns And in a moment makes them desolate

[gcs-download-2 : main] I find thou art no less than fame hath bruited And more than may be gatherd by thy shape Let my presumption not provoke thy wrath

[echo : main] Text 1: With which he yoketh your rebellious necks Razeth your cities and subverts your towns And in a moment makes them desolate
[echo : main]
[echo : main] Text 2: I find thou art no less than fame hath bruited And more than may be gatherd by thy shape Let my presumption not provoke thy wrath
[echo : main]

List of Available Features

To understand how each feature is implemented and its current status, please visit the FEATURES doc.

Tested Pipelines

We are testing the compiler on more than 80 pipelines found in the Kubeflow Pipelines repository, specifically the pipelines in KFP compiler testdata folder, the KFP core samples and the samples contributed by third parties.

A report card of Kubeflow Pipelines samples that are currently supported by the kfp-tekton compiler can be found here. If you work on a PR that enables another of the missing features please ensure that your code changes are improving the number of successfully compiled KFP pipeline samples.

Troubleshooting

  • When you encounter permission issues related to ServiceAccount, refer to Servince Account and RBAC doc

  • If you run into the error bad interpreter: No such file or director when trying to use python's venv, remove the current virtual environment in the .venv directory and create a new one using virtualenv .venv

  • For big data passing, user need to create PV manually, or enable dynamic volume provisioning, refer to the link of: https://kubernetes.io/docs/concepts/storage/dynamic-provisioning

    User need to create pvc manually with the pvc name same as pipelinerun name until issue #181 addressed