pipelines

Kubeflow Pipelines v2

There are two modes Kubeflow Pipelines (KFP) v2 can run in:

• v2 compatible -- supports core v2 features in KFP v1
• v2 (v2 engine) -- full feature v2 in KFP v2's new architecture

Code for both modes live inside this folder.

Kubeflow Pipelines v2 compatible

Status: Beta, to be deprecated in KFP SDK v1.9. Documentation: Introducing Kubeflow Pipelines SDK v2 Known Caveats & breaking changes Design: bit.ly/kfp-v2-compatible Github Project: KFP v2 compatible mode project

Plan:

• 2021 late May - first release
• 2021 early August (delayed from mid July) - feature complete
• End of support in KFP SDK v1.9

Kubeflow Pipelines v2 engine

Status: not released (WIP) Design: bit.ly/kfp-v2 (You need to join kubeflow-discuss google group to get access.) Github Project: KFP v2 project Tracker Issue: KFP v2 tracker

Plan:

• 2021 October - Alpha release
• TBD - Beta/Stable release

Developing

Developing KFP v2

Prerequisites:

• Install a KFP standalone instance on Google Cloud:

  This does not currently work on other envs, because some tests use GCS & GCS client. Welcome contributions to make it portable.

• Install go, python, kfp pypi package, docker.

• Install ko CLI tool:

  make install-ko
  

• Configure dev environment by creating a config file called .env in this folder, it should have the following content:

  DEV_IMAGE_PREFIX=<an container image registry prefix you own>
  

  For example:

  DEV_IMAGE_PREFIX?=gcr.io/ml-pipeline-test/dev/kfp-
  

  Then after images are built, they will be pushed to locations like gcr.io/ml-pipeline-test/dev/kfp-driver.

  The .env file is ignored by git, it's your local development configuration.

  Verify you can push images to the registry:

  # push all built dev images to DEV_IMAGE_PREFIX
  make image-dev
  

  set up go environment value

  # .env is a Makefile local config (ignored by git)
  echo "GOOS_VALUE=$(go env GOOS)" >> .env
  echo "GOARCH_VALUE="$(go env GOARCH) >> .env
  

• Install sample test python dependencies (require Python 3.7 or 3.8 due to ml-metadata limitation):

  cd test
  pip install -r requirements.txt
  

• Connecting to Kubeflow Pipelines using the SDK client.

  Recommend adding the env vars to your .bashrc or .zshrc etc to persist your config.

  Verify your configuration and connectivity:

  kfp experiment list
  

  Requirements on the KFP backend installation:

  • Current limitation, this only works for KFP Standalone, not tested on full Kubeflow yet.
  • KFP backend version should be at least 1.7.0-rc.2.

  Requirements on the KFP SDK package:

  • KFP v2 components defined using @component decorator installs KFP SDK package at runtime. To use a compatible KFP SDK, define the following environment variable before running e2e test:

  export KFP_PACKAGE_PATH="git+https://github.com/kubeflow/pipelines.git@master#subdirectory=sdk/python"
  

Instructions:

• Run everything e2e: build images, backend compiler, compile pipelines and run them:

  make dev
  

• Run go unit tests:

  make test
  # or watch file changes and rerun automatically
  make test-watch
  

• For individual targets, read the Makefile directly.

• Run one sample test:

  cd "${REPO_ROOT}/v2"
  make install-compiler # This needs to be run each time you update compiler code.
  cd "${REPO_ROOT}"
  python -m samples.path.to.sample_test
  

  Read v2 sample test documentation for more details.

Update licenses

Note, this is currently outdated instructions for v2 compatible mode. We haven't set up licensing workflow for v2 engine.

Download the license tool binary from https://github.com/Bobgy/go-licenses/releases and put it into $PATH.

Update licenses info by:

make license-launcher

or run the following to enable verbose output:

GO_LICENSES_FLAGS=-v4 make license-launcher

After the update, check generated third_party/licenses/launcher.csv file to make sure licenses of new dependencies are correctly identified.

If something is unexpected, examine the unexpected dependencies by yourself and add overrides to go-licenses.yaml.

For detailed documentation about the tool: https://github.com/Bobgy/go-licenses/tree/main/v2.

Releasing KFP v2 compatible image

Note, this is currently outdated instructions for v2 compatible mode. We haven't set up releasing workflow for v2 engine.

1. Build & release gcr.io/ml-pipeline/kfp-launcher:$TAG_NAME (this step needs write permission to gcr.io/ml-pipeline):

   # login locally
   gcloud auth login
   gcloud auth configure-docker
   
   cd kubeflow/pipelines/v2
   TAG_NAME=1.6.0
   LAUNCHER_IMAGE="gcr.io/ml-pipeline/kfp-launcher:${TAG_NAME}" make push-launcher
   

2. Edit v2_compat.py -- pin _DEFAULT_LAUNCHER_IMAGE to the tag we will release.

3. Continue with KFP python SDK release instructions.