[go: up one dir, main page]
Skip to content

Latest commit

 

History

History
275 lines (191 loc) · 10.7 KB

CONTRIBUTING.md

File metadata and controls

275 lines (191 loc) · 10.7 KB
Raw

Kubeflow Contributor Guide

Welcome to the Kubeflow project! We'd love to accept your patches and contributions to this project. Please read the contributor's guide in our docs.

The contributor's guide

  • shows you where to find the Contributor License Agreement (CLA) that you need to sign,
  • helps you get started with your first contribution to Kubeflow,
  • and describes the pull request and review workflow in detail, including the OWNERS files and automated workflow tool.

This document focus on technical aspects while contributing to the Model Registry project

Contributing to Model Registry using Apple-silicon/ARM-based computers

Some limitations apply when developing on this project, specifically using Apple-silicon and Mac OSX. The content from this guide might also be applicable in part for general ARM-based developers/users, beyond Mac OSX.

Makefile

The make command shipped with Mac OSX (at the time of writing) is a bit old:

% make --version
GNU Make 3.81
This program built for i386-apple-darwin11.3.0

and known to cause problems when using this project's Makefile:

% make build
openapi-generator-cli validate -i api/openapi/model-registry.yaml
make: openapi-generator-cli: No such file or directory
make: *** [openapi/validate] Error 1

i.e. failing to locate the bin/ executables managed in the Makefile.

The solution is to use and updated version of make.

You can install it with Homebrew:

% brew install make
...
==> Pouring make--4.4.1.arm64_ventura.bottle.tar.gz
==> Caveats
GNU "make" has been installed as "gmake".
If you need to use it as "make", you can add a "gnubin" directory
to your PATH from your bashrc like:

    PATH="/opt/homebrew/opt/make/libexec/gnubin:$PATH"
...

and now you can substitute gmake every time the make command is mentioned in guides (or perform the path management per the caveat).

Local kubernetes deployment of Model Registry

To test the Model Registry locally without mocking the k8s calls, the Model Registry backend can be deployed using Kind.

Prerequisites

The following tools need to be installed in your local environment:

Choose the networking setup that fits your needs, either port-forwarding or Ingress.

Port-forwarding guide

Create a Kind cluster with the following command:

kind create cluster

and then follow the steps from the Installation guide on the Kubeflow website, to set up the port-forwarding and deploy the Model Registry on the cluster.

Ingress guide

Follow the Ingress guide to set up the Ingress controller and deploy the Model Registry on the cluster.

Docker engine

Several options of docker engines are available for Mac. Having Docker installed is also helpful for Testcontainers.

Podman

Podman also supports Rosetta (Apple specific) emulation which is handy since the Google MLMD project dependency is x86 specific.

We recommend setting up the Podman machine with root privileges, and setting the environment variable

export TESTCONTAINERS_RYUK_PRIVILEGED=true

when running TestContainer-based Model Registry Python tests (for more information, see here).

Fedora

Fedora requires further setup to make testcontainers work with Podman, the following steps are required:

  • start the podman socket service
systemctl --user start podman.socket
  • set the environment variable
export DOCKER_HOST=unix://$XDG_RUNTIME_DIR/podman/podman.sock

If you need more information, please refer to the Testcontainers using Podman.

Colima

Colima offers Rosetta (Apple specific) emulation which is handy since the Google MLMD project dependency is x86 specific. You can install Colima (and Docker) with Homebrew.

You can create a Colima "docker context" focusing on x86 emulation with:

colima start --vz-rosetta --vm-type vz --arch x86_64 --cpu 4 --memory 8

To use with Testcontainers for Go you can use these commands:

export DOCKER_HOST="unix://${HOME}/.colima/default/docker.sock"
export TESTCONTAINERS_DOCKER_SOCKET_OVERRIDE="/var/run/docker.sock"

as instructed in this guide.

This colima setups allows to:

  • launch Integration tests in Go (used in Core go layer) with Testcontainers for Go
  • launch DevContainer to be able to install MLMD python wheel dependency (which is x86 specific)

DevContainer

Using a DevContainer is helpful to develop with the Model Registry Python client, since it needs to wrap MLMD python dependency (which is x86 specific).

This allows for instance with VSCode DevContainer extension to re-open VSCode window "inside" an x86 emulated docker container. The experience is very similar to when on GitHub.com you press dot . and you get a VSCode "inside something", except it is local to your machine. It's not super fast because x86 is emulated via Rosetta, but works "good enough" to complete most tasks without requiring remote connection to a real x86 server.

To use DevContainer as a framework directly, a command-line tool and an SDK is available as well on the upstream project: https://containers.dev.

Don't forget you will need a Docker context for x86 emulation, for instance with colima (see previous step) this can be achieved with:

colima start --vz-rosetta --vm-type vz --arch x86_64 --cpu 4 --memory 8

Define this .devcontainer/devcontainer.json file :

{
	"name": "Python 3",
	// "image": "mcr.microsoft.com/devcontainers/python:1-3.10-bullseye"
	"build": {
		"dockerfile": "Dockerfile"
	},
	"runArgs": [
		"--network=host"
	],
	"customizations": {
		"vscode": {
			"extensions": [
				// does not work well in DevContainer: "robocorp.robotframework-lsp",
				"d-biehl.robotcode"
			]
		}
	},
//...
}

The network=host allow from inside the devcontainer to reach any "service" exposed on your computer (host). This is helpful if other containers are started on your computer (eg: a PostgreSQL or DB in another container).

The customizations.vscode.extensions pre-loads additional extensions needed in VSCode to be executing from inside the DevContainer.

Define this .devcontainer/Dockerfile file:

FROM mcr.microsoft.com/devcontainers/python:1-3.10-bullseye

# Here I use the USER from the FROM image
ARG USERNAME=vscode
ARG GROUPNAME=vscode

# Here I use the UID/GID from _my_ computer, as I'm _not_ using Docker Desktop
ARG USER_UID=501
ARG USER_GID=20

RUN groupmod --gid $USER_GID -o $GROUPNAME \
    && usermod --uid $USER_UID --gid $USER_GID $USERNAME \
    && chown -R $USER_UID:$USER_GID /home/$USERNAME

# General setup which is "cached" for convenience
RUN pip install -U pip setuptools
RUN pip install -U poetry
RUN pip install -U "ml-metadata==1.14.0"
RUN pip install -U robotframework
RUN pip install -U robotframework-requests
RUN pip install -U PyYAML

The group/user is needed as on Mac anything but Docker Desktop will need to set correct FS permissions. (more details here: https://github.com/devcontainers/spec/issues/325).

The RUN pip install "caches" the local installation (inside the DevContainer) of some wheels which are almost always needed, for convenience.

Please notice the line RUN pip install -U "ml-metadata==1.14.0": that pip install would otherwise fail on Apple-silicon/ARM-based machines.

E.g. when issued on bare-metal-Apple-silicon fails with:

% pip install -U "ml-metadata==1.14.0"
ERROR: Could not find a version that satisfies the requirement ml-metadata==1.14.0 (from versions: 0.12.0.dev0, 0.13.0.dev0, 0.13.1.dev0)
ERROR: No matching distribution found for ml-metadata==1.14.0

As all the wheels are x86 specific: https://pypi.org/project/ml-metadata/1.14.0/#files

So it's not even possible to receive code assists. However, after clicking to re-open the project inside an (emulated) DevContainer:

Then with the given setup MLMD is already installed inside the DevContainer:

At this point Poetry is already installed as well and can be used to build and run test of the Model Registry Python client.

FAQ

Error docker.errors.NotFound: 404 Client Error for http+docker://localhost/v1.41/containers ...

This happens on Mac OSX when running Testcontainers-based Python tests, and Ryuk does not have the correct access priviledges.

-------------------------------------------- Captured stderr setup ---------------------------------------------
Pulling image testcontainers/ryuk:0.7.0
Container started: 4af385c2d670
---------------------------------------------- Captured log setup ----------------------------------------------
WARNING  root:config.py:59 DOCKER_AUTH_CONFIG is experimental, see testcontainers/testcontainers-python#566
INFO     testcontainers.core.container:container.py:88 Pulling image testcontainers/ryuk:0.7.0
INFO     testcontainers.core.container:container.py:101 Container started: 4af385c2d670
=========================================== short test summary info ============================================
ERROR tests/store/test_wrapper.py::test_get_undefined_artifact_type_id - docker.errors.NotFound: 404 Client Error for http+docker://localhost/v1.41/containers/4af385c2d67067875f6f0...
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! stopping after 1 failures !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
=============================================== 1 error in 0.50s ===============================================

Solution:

  • Use export TESTCONTAINERS_RYUK_PRIVILEGED=true (we recommend this option)
  • Use export TESTCONTAINERS_RYUK_DISABLED=true to disable Ryuk entirely; the tests are configured to close all container resources, but we do not recommend this option

For more information: