Development workflow#

Prerequisites#

Required software#

You need the following installed to contribute to Orchest:

Python version 3.x
Go: Used by the orchest-controller and needed to run our scripts/build_container.sh to build Orchest’s images.
Docker: To build Orchest’s images.
minikube: To deploy Orchest on a local cluster.
kubectl: To manage k8s clusters.
helm: Needed to run our scripts/build_container.sh to create the manifests to deploy the orchest-controller.
pre-commit: Running pre-commit hooks, e.g. linters.
npm and pnpm: To develop the front-end code of Orchest.
Google Chrome: Requirement to run integration tests locally.

Optional, but highly recommended:

k9s: Terminal UI to manage k8s clusters.
jq: Useful when working with JSON in your terminal.
gron: Make JSON greppable.

Dependencies#

After installing the required software, you need to configure the tools and install additional dependencies.

Note

Make sure you are inside the root of the orchest repository.

# Set-up pre-commit:
pre-commit install
# Install frontend dependencies for local development:
npm run setup --install && pnpm i
# Install the Orchest CLI to manage the Orchest Cluster in k8s:
python3 -m pip install -e orchest-cli
# Install dependencies to build the docs:
python3 -m pip install -r docs/requirements.txt

Orchest’s integration tests require a MySQL client to be installed:

Linux

sudo apt install -y default-libmysqlclient-dev

macOS

brew install mysql

Cluster for development#

Currently, the development tools assume that you have Orchest installed on a local minikube cluster. To get the best development experience, it is recommended to mount the Orchest repository in minikube which allows for incremental development.

Note

Make sure you are inside the root of the orchest repository.

# Delete any existing cluster
minikube delete

# Start minikube with the repository mounted in the required place
# for hot reloading to work.
minikube start \
  --cpus max \
  --memory max \
  --addons ingress metrics-server \
  --mount-string="$(pwd):/orchest-dev-repo" --mount

Installing Orchest for development#

Now when all dependencies are installed and your cluster is set up, you can install Orchest!

But before doing so, it is important to realize that the Docker daemon of your host is different from the Docker daemon of minikube. This means that you need to build Orchest’s images on the minikube node in order for minikube to be able to use them, otherwise it will pull the images from DockerHub. Note that DockerHub only contains images of Orchest releases and not active code changes from GitHub branches. Therefore it is important to configure your environment to use minikube’s Docker daemon before building images.

Note

The command below needs to be run in every terminal window you open!

# Use minikube's Docker daemon:
eval $(minikube -p minikube docker-env)

Next, you can build Orchest’s images. Again, there is an important realization to make here and that is that the images you build are given a specific tag. This tag is used by the Orchest Controller (orchest-controller) to manage the Orchest Cluster, thus if you build images with tag X but deploy the orchest-controller with tag Y, then the orchest-controller will start pulling the images with tag Y from DockerHub (instead of using the locally built images with tag X). This will become important when rebuilding images after making code changes.

# Verify whether you are using minikube's Docker daemon
echo "$MINIKUBE_ACTIVE_DOCKERD"

Let’s build the minimal set of required images for Orchest to run:

# Set the *tag* to the latest Orchest version available
export TAG="$(orchest version --latest)"

# Build the minimal set of images for Orchest to run
scripts/build_container.sh -M -t $TAG -o $TAG

And finally, install Orchest:

# The --dev flag is used so that it doesn't pull in the release assets
# from GitHub, but instead uses the manifests from the local filesystem
# to deploy the Orchest Controller. NOTE: these manifests are automatically
# generated when running the above `build_container.sh` script ;)
orchest install --dev

Take a look in k9s and see how Orchest is getting installed.

Once the installation is completed you can reach Orchest using one of the following approaches, depending on your operating system:

Linux

Simply access the Orchest UI by browsing to the IP returned by:

minikube ip

macOS

Run the tunnel daemon and browse to localhost.

sudo minikube tunnel

Does everything look good? Awesome! You’re all set up and ready to start coding now! 🎉

Have a look at our best practices and our GitHub to find interesting issues to work on.

Redeploying Orchest after code changes#

Warning

Running minikube delete is not recommend because it will lose the Docker cache on the minikube node, making rebuilding images very slow. Luckily, it is unlikely you will need to (ever) run minikube delete.

In this section we will go over the three ways to “redeploy” Orchest to reflect your code changes. Note that each approach is best used in specific circumstances.

Using development mode to automatically reflect code changes. (link) Best used when working on a PR and you would like to see your code changes immediately, especially useful when developing the front-end (e.g. orchest-webserver).
Rebuilding and redeploying only the service’s image that you made changes to. (link) Best used when you know the code changes affect only one service and you don’t want to fully re-install Orchest. For example, you want to test a PR that only changed the front-end and want to run in production mode instead of development mode.
Completely uninstalling and installing Orchest again. (link) When making larger changes that touch different parts of Orchest, it is a good idea to fully re-install Orchest. Do note that this should be rather fast because the Docker cache is used when rebuilding images.

Development mode (incremental development)#

For the next steps, we assume you already installed Orchest.

To get “hot reloading”, you need to make sure your minikube cluster was created using the above mount command and have Orchest serve files from your local filesystem (that contains code changes) instead of the files baked into the Docker images. To achieve the latter, simply run:

Note: Don’t forget to disable cache (DevTools -> Disable cache) or force reload (Command/Ctrl + Shift + R) to see frontend changes propagate.

# In case any new dependencies were changed or added they need
# to be installed
pnpm i

# Run the client dev server for hot reloading of client (i.e. FE)
# files
pnpm run dev

# Get the Orchest Cluster to serve files from your local filesystem.
orchest patch --dev

Note: Your cluster will stay in --dev mode until you unpatch it (using orchest patch --no-dev).

The services that support incremental development are:

orchest-webserver
orchest-api
auth-server

For changes to all other services, you need to redeploy the respective image as described in the next section.

Note

Even if you do incremental development, it is good practice to rebuild the containers and run in production mode before opening a PR (see before committing).

Teardown#

If at any point you want to disable incremental development, proceed as follows:

# Kill the client dev server
kill $(pidof pnpm)

# Revert the patch
orchest patch --no-dev

To stop the cluster, it’s enough to call minikube stop, which will stop all the pods.

Switching branches#

If you have a running development installation with hot reloading, every time you make a change to the code it will be automatically reloaded. However, when switching git branches that are very different, or if changes to certain core components were made, this procedure might produce inconsistent results. A safer way to proceed is to uninstall Orchest before making the switch, see re-installing Orchest.

Rebuilding images#

To easily test code changes of an arbitrary service, you will need to:

rebuild the respective Docker image and
make it available to the k8s deployment.

The procedure changes slightly depending on the deployment type, i.e. single-node or multi-node. Luckily, in the majority of cases you will be using a local single-node cluster (like the one you created in the previous steps).

For the sake of simplicity (without loss of generality), let’s assume you made changes to the orchest-api.

Single node

Generally, single node deployments make it far easier to test changes. First of all, configure your environment to use minikube’s Docker daemon if you haven’t already:

# If not active, set it
eval $(minikube -p minikube docker-env)

Now you’re ready to rebuild the images, to which you made changes, using the build_container.sh script.

Note: It is very important (otherwise your code changes will not be reflected) to use the tag equal to the currently running Orchest version.

export TAG="$(orchest version)"

# Rebuild the images that need it
scripts/build_container.sh -i orchest-api -t $TAG -o $TAG

Alternatively, you can run scripts/build_container.sh -M -t $TAG -o $TAG to rebuild the absolute minimal required set of images instead of cherry picking. This is not a bad idea given that the Docker cache will be used and thus rebuilds of unchanged images is quick.

Lastly, you need to make sure that your new orchest-api image is used by minikube. This can be done by deleting the respective orchest-api pod (which will automatically get replaced with a new pod serving your updated image thanks to Kubernetes deployments):

# Kill the pods of the orchest-api, so that the new image gets used
# when new pod gets automatically deployed.
kubectl delete pods -n orchest -l "app.kubernetes.io/name=orchest-api"

Check out k9s if you want to use a visual interface instead (highly recommended!).

Multi node

The procedure for single-node is not possible in multi node deployments though. Since this is slightly more involved, we provide the following scripts:

# Redeploy a service after building the image using the repo code.
# This is the script that you will likely use the most. This script
# assumes Orchest is installed and running, since it interacts with
# an Orchest service.
bash scripts/redeploy_orchest_service_on_minikube.sh orchest-api

# Remove an image from minikube. Can be useful to force a pull from
# a registry.
bash scripts/remove_image_from_minikube.sh orchest/orchest-api

# Build an image with a given tag, on all nodes.
bash scripts/build_image_in_minikube.sh orchest-api v2022.03.7

# Run arbitrary commands on all nodes.
bash scripts/run_in_minikube.sh echo "hello"

Warning

The redeploy and build_image scripts require the Orchest repository to be mounted in minikube. However, note that multi node mounting might not be supported by all minikube drivers. We have tested with docker, the default driver.

Re-installing Orchest#

When making larger changes or when wanting to check out a different branch for example, it is a good idea to re-install Orchest. Rest assured, this should be fairly quick!

# Uninstall Orchest before proceeding
orchest uninstall

# Switch git branches if applicable
git switch feature-branch

# Rebuild containers, if needed
eval $(minikube -p minikube docker-env)
export TAG="$(orchest version --latest)"
scripts/build_container.sh -M -t $TAG -o $TAG

# Install Orchest again
orchest install --dev

Making changes#

Before committing#

Make sure your development environment is set up correctly (see prerequisites) so that pre-commit can automatically take care of running the appropriate formatters and linters when running git commit.

In our CI we also run a bunch of checks, such as unit tests and integration tests to make sure the codebase remains stable. To read more about testing, check out the testing section.

Opening a PR#

Note

When opening a PR please change the base in which you want to merge from master to dev. The GitHub docs describe how this can be done.

We use gitflow as our branching model with master and dev being the described master and develop branches respectively. Therefore, we require PRs to be merged into dev instead of master.

When opening the PR a checklist will automatically appear to guide you to successfully completing your PR 🏁

Changing Python dependencies#

Python dependencies for the microservices are specified using pip’s requirements.txt files. Those files are automatically generated by pip-tools from requirements.in files by calling pip-compile, which locks all the transitive dependencies. After a locked requirements.txt file is in place, subsequent calls to pip-compile will not upgrade any of the dependencies unless the constraints in requirements.in are modified.

To manually upgrade a dependency to a newer version, there are several options:

pip-compile -P <dep>  # Upgrades <dep> to latest version
pip-compile -U  # Try to upgrade everything

As a general rule, avoid writing exact pins in requirements.in unless there are known incompatibilities. In addition, avoid manually editing requirements.txt files, since they will be automatically generated.

Warning

A bug in pip-tools affects local dependencies. Older versions are not affected, but they are not compatible with modern pip. At the time of writing, the best way forward is to install this fork (see this PR for details):

pip install -U "pip-tools @ git+https://github.com/richafrank/pip-tools.git@combine-without-copy"

Database schema migrations#

Whenever one of the services’s database models (in their respective models.py) have been changed, a database migration has to be performed so that all existing users are unaffected by the schema change on update (since they can then be automatically migrated to the latest version).

# Depending on the service that requires schema changes.
scripts/migration_manager.sh orchest-api migrate
scripts/migration_manager.sh orchest-webserver migrate

# For more options run:
scripts/migration_manager.sh --help

Building the docs#

Our docs are built using Read the Docs with Sphinx and written in reStructuredText.

To build the docs, run:

cd docs
make html

Tip

👉 If you didn’t follow the prerequisites, then make sure you’ve installed the needed requirements to builds the docs:

python3 -m pip install -r docs/requirements.txt

Example VS Code monorepo set-up#

Note

👉 This section is for VS Code and pyright users.

If you use VS Code (or the pyright language server to be more precise) then this section is for you. The different services contain their own pyrightconfig.json file that configures smart features such as auto complete, go to definition, find all references, and more. For this to work, you need to install the dependencies of the services in the correct virtual environment by running:

scripts/run_tests.sh

Next you can create a workspace file that sets up VS Code to use the right Python interpreters (do note that this won’t include all the files defined in the Orchest repo), e.g.:

{
  "folders": [
    {
      "path": "services/orchest-api"
    },
    {
      "path": "services/orchest-webserver"
    },
    {
      "path": "services/base-images/runnable-shared"
    },
    {
      "path": "services/session-sidecar"
    },
    {
      "name": "orchest-sdk",
      "path": "orchest-sdk/python"
    },
    {
      "name": "internal lib Python",
      "path": "lib/python/orchest-internals/"
    }
  ],
  "settings": {}
}

Automated tests#

Unit tests#

Unit tests are being ported to k8s, stay tuned :)!

Integration tests#

Integration tests are being ported to k8s, stay tuned :)!

Manual testing#

Test Environment or custom Jupyter base image changes#

When building environment or custom Jupyter images the image builder mounts the socket of the container runtime running on the node. This means that, to test changes to a base image, all that is needed is to build or load the new base image in the container runtime. Example:

# Make changes to services/base-images/base-kernel-py/Dockerfile, then:
eval $(minikube -p minikube docker-env)
bash scripts/build_container.sh -o v2022.08.11 -t v2022.08.11  -i base-kernel-py
# That's it, you can now build an environment image in Orchest using the
# new python base image.

Currently, this has only been tested with docker as the container runtime.

Test running Orchest on `containerd`#

To test running Orchest on containerd, we recommend installing MicroK8s. Alternatively, you can also set up Orchest on GKE (see installation) or install MicroK8s in a VM (e.g. using VirtualBox).

Next, enable the following addons:

microk8s enable hostpath-storage \
    && microk8s enable dns \
    && microk8s enable ingress

Now that MicroK8s is correctly configured we need to rebuild Orchest’s images and save them to a .tar file so that containerd can unpack the file and use the images.

export TAG=v2022.06.4
scripts/build_container.sh -M -t $TAG -o $TAG
docker save \
    $(docker images | awk '{if ($1 ~ /^orchest\//) new_var=sprintf("%s:%s", $1, $2); print new_var}' | grep $TAG | sort | uniq) \
    -o orchest-images.tar

Next, inside the MicroK8s node (which can be your host), you can import the images using:

microk8s ctr --namespace k8s.io --address /var/snap/microk8s/common/run/containerd.sock image import orchest-images.tar

# OR, requires ctr to be installed: https://github.com/containerd/containerd/releases
sudo ctr -n k8s.io -a /var/snap/microk8s/common/run/containerd.sock i import orchest-images.tar

Now you can install Orchest:

orchest install --dev --socket-path=/var/snap/microk8s/common/run/containerd.sock

Run Orchest Controller locally#

For easier debugging it is possible to run the orchest-controller locally with a debugger. We will explain how to do so using VSCode. Make sure your cluster is set up and you’ve installed Go, then follow the steps below:

Run the orchest-controller with a debugger in VSCode, example launch.json:

{
  "configurations": [
    {
      "name": "Launch ctrl",
      "type": "go",
      "request": "launch",
      "mode": "debug",
      "program": "${workspaceFolder}/cmd/controller/main.go",
      "args": [
        "--inCluster=false",
        "--defaultVersion=<INSERT VERSION, e.g. v2022.05.0>",
        "--assetsDir=${workspaceFolder}/deploy",
        "--endpoint=:5000"
      ],
      "env": {
        "KUBECONFIG": "~/.kube/config"
      }
    }
  ]
}

Next install Orchest and afterwards issue other commands to test the controller with:

# Asuming you are in the root of the orchest git repository
orchest install --dev

# Delete orchest-controller deployment so that the one started with
# the debugger does everything
kubectl delete -n orchest deploy orchest-controller

The Orchest Controller should now be running inside a debugger session.

Without using VSCode#

Build the orchest-controller binary via the Makefile in services/orchest-controller and run the orchest-controller by passing the following command line arguments:

# Asuming you have built the controller via "make controller" command
./bin/controller --inCluster=false --defaultVersion=v2022.05.3 \
--endpoint=:5000 --assetsDir=./deploy

Test updating Orchest#

Through the CLI#

orchest uninstall
scripts/build_container.sh -M -t "v2022.04.4" -o "v2022.04.4"
orchest install --dev
scripts/build_container.sh -M -t "v2022.04.5" -o "v2022.04.5"
orchest update --dev --version=v2022.04.5
scripts/build_container.sh -M -t "v2022.04.6" -o "v2022.04.6"
orchest update --dev --version=v2022.04.6

Through the UI#

For this to work you need to be running in dev mode and have the orchest-dev-repo mounted (as per setting up minikube for development).

# Start from a clean slate so that we know what version we are on
# before invoking the update.
orchest uninstall

# Build whatever version you like! In case you want to test out
# the product after the update, build the X-1 latest release
# tag.
scripts/build_container.sh -m -t "v2022.04.4" -o "v2022.04.4"

# Installing and making sure running in dev.
orchest install --dev
orchest patch --dev
pnpm run dev

# Build the version to update to
scripts/build_container.sh -M -t "v2022.04.5" -o "v2022.04.5"

# Invoke the update through the UI go to:
# http://localorchest.io/update
...

# In case you want to test it again
scripts/build_container.sh -M -t "v2022.04.6" -o "v2022.04.6"
# Invoke the update through the UI go to:
# http://localorchest.io/update
...

# And repeat if you like.

Development workflow#

Prerequisites#

Required software#

Dependencies#

Cluster for development#

Installing Orchest for development#

Redeploying Orchest after code changes#

Development mode (incremental development)#

Teardown#

Switching branches#

Rebuilding images#

Re-installing Orchest#

Making changes#

Before committing#

Opening a PR#

Changing Python dependencies#

Database schema migrations#

Building the docs#

Example VS Code monorepo set-up#

Automated tests#

Unit tests#

Integration tests#

Manual testing#

Test Environment or custom Jupyter base image changes#

Test running Orchest on containerd#

Run Orchest Controller locally#

Without using VSCode#

Test updating Orchest#

Through the CLI#

Through the UI#

Test running Orchest on `containerd`#