Flyte Interactive

FlyteInteractive = Develop Flyte with the speed of flying

FlyteInteractive plugin provides users’ favorite interface to develop and debug a flyte task interactively. We support vscode, jupyter (WIP), and neovim (WIP).

Installation

To install the plugin, run the following command:

pip install flytekitplugins-flyteinteractive

Vscode

FlyteInteractive Vscode offers an easy solution for users to run Python tasks within an interactive VSCode server, compatible with any image. @vscode is a decorator which users can put within @task and user function. With @vscode, the task will install vscode dependencies (skip if they already exist) and run a vscode server instead of the user defined functions.

Starter Example

from flytekit import task
from flytekitplugins.flyteinteractive import vscode

@task
@vscode
def train():
    ...

User Guide

1. Build the image with Dockerfile.dev docker buildx build --push . -f Dockerfile.dev -t localhost:30000/flytekit:dev --build-arg PYTHON_VERSION=3.9
2. Run the decorated task on the remote. For example: pyflyte run --remote --image localhost:30000/flytekit:dev [PYTHONFILE] [WORKFLOW|TASK] [ARGS]...
3. Once the code server is prepared, you can forward a local port to the pod. For example: kubectl port-forward -n [NAMESPACE] [PODNAME] 8080:8080.
4. You can access the server by opening a web browser and navigating to localhost:8080.

VSCode example screenshot:

Build Custom Image with VSCode Plugin

If users want to skip the vscode downloading process at runtime, they have the option to create a custom image with vscode by including the following lines in their Dockerfile.

# Include this line if the image does not already have 'curl' installed.
+ RUN apt-get -y install curl
# Download and extract the binary, and ensure it's added to the system's $PATH.
+ RUN mkdir /tmp/code-server
+ RUN curl -kfL -o /tmp/code-server/code-server-4.18.0-linux-amd64.tar.gz https://github.com/coder/code-server/releases/download/v4.18.0/code-server-4.18.0-linux-amd64.tar.gz
+ RUN tar -xzf /tmp/code-server/code-server-4.18.0-linux-amd64.tar.gz -C /tmp/code-server/
+ ENV PATH="/tmp/code-server/code-server-4.18.0-linux-amd64/bin:${PATH}"

Advanced Examples

from flytekit import task, workflow
from flytekitplugins.flyteinteractive import vscode, VscodeConfig, DEFAULT_CODE_SERVER_EXTENSIONS

@task(
    container_image="localhost:30000/flytekit-vscode:0.0.2",
    environment={"FLYTE_SDK_LOGGING_LEVEL": "20"}
)
@vscode(

)
def t():
    ...



# this vscode task will be killed within 10 secs
@task(
    container_image="localhost:30000/flytekit-vscode:0.0.2",
    environment={"FLYTE_SDK_LOGGING_LEVEL": "20"}
)
@vscode(
    max_idle_seconds=10,
)
def t_short_live():
    ...



# this vscode task will download default extension + vim extension
config_with_vim = VscodeConfig(
    extension_remote_paths=DEFAULT_CODE_SERVER_EXTENSIONS+["https://open-vsx.org/api/vscodevim/vim/1.27.0/file/vscodevim.vim-1.27.0.vsix"]
)

@task(
    container_image="localhost:30000/flytekit-vscode:0.0.2",
    environment={"FLYTE_SDK_LOGGING_LEVEL": "20"}
)
@vscode(
    config=config_with_vim
)
def t_vim():
    ...



@workflow
def wf():
    t()
    t_short_live()
    t_vim()

Jupyter

FlyteInteractive Jupyter offers an easy solution for users to run Python tasks within a Jupyter Notebook server, compatible with any image. @jupyter is a decorator which users can put within @task and user function. With @jupyter, the task will run a Jupyter Notebook server instead of the user defined functions.

Starter Example

from flytekit import task
from flytekitplugins.flyteinteractive import jupyter

@task
@jupyter
def train():
    ...

User Guide

1. Build the image with Dockerfile.dev docker buildx build --push . -f Dockerfile.dev -t localhost:30000/flytekit:dev --build-arg PYTHON_VERSION=3.9
2. Run the decorated task on the remote. For example: pyflyte run --remote --image localhost:30000/flytekit:dev [PYTHONFILE] [WORKFLOW|TASK] [ARGS]...
3. Once the code server is prepared, you can forward a local port to the pod. For example: kubectl port-forward -n [NAMESPACE] [PODNAME] 8888:8888.
4. You can access the server by opening a web browser and navigating to localhost:8888.

Jupyter example screenshot: