Writing Your First Workflow

The easiest way to author a Flyte Workflow is using the provided python SDK called “FlyteKit”.

You can save some effort by cloning the flytesnacks repo, and re-initializing it as a new git repository

git clone git@github.com:lyft/flytesnacks.git myflyteproject
cd myworkflow
rm -rf .git
git init

now open the “Makefile” and change the first line to IMAGE_NAME=myflyteproject

Let’s also remove the existing python task so we can write one from scratch.

rm workflows/edges.py

Creating a Project

In Flyte, workflows are organized into namespaces called “Projects”. When you register a workflow, it must be registered under a project.

Lets create a new project called myflyteproject. Use the project creation endpoint to create the new project

curl -X POST localhost:30081/api/v1/projects -d '{"project": {"id": "myflyteproject", "name": "myflyteproject"} }'

Writing a Task

The most basic Flyte primitive is a “task”. Flyte Tasks are units of work that can be composed in a workflow. The simplest way to write a Flyte task is using the FlyteSDK.

Start by creating a new file

touch workflows/first.py

This directory has been marked in the configuration file as the location to look for workflows and tasks. Begin by importing some of the libraries that we’ll need for this example.

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import urllib.request as _request

import cv2
from flytekit.common import utils
from flytekit.sdk.tasks import python_task, outputs, inputs
from flytekit.sdk.types import Types
from flytekit.sdk.workflow import workflow_class, Output, Input

From there, we can begin to write our first task. It should look something like this.

def edge_detection_canny(wf_params, image_location, parsed_image):
    with utils.AutoDeletingTempDir('test') as tmpdir:
        plane_fname = '{}/plane.jpg'.format(tmpdir.name)
        with _request.urlopen(image_location) as d, open(plane_fname, 'wb') as opfile:
            data = d.read()

        img = cv2.imread(plane_fname, 0)
        edges = cv2.Canny(img, 50, 200)  # hysteresis thresholds

        output_file = '{}/output.jpg'.format(tmpdir.name)
        cv2.imwrite(output_file, edges)


Some of the new concepts demonstrated here are:

  • wf_params: The first argument to a python task is a Flyte SDK defined object that offers handlers like logging.

  • Inputs and outputs are first defined in the decorator, and then passed into the argument of the function. Note that the names in the function signature need to match those in the decorator arguments.

  • A Blob is a Flyte Kit type that represents binary data. It is used to offload data to a storage location like S3. Here we use it to store an image.

Writing a Workflow

Next you need to call that task from a workflow. In the same file, add these lines.

class EdgeDetectorWf(object):
    image_input = Input(Types.String, required=True, help="Image to run for")
    run_edge_detection = edge_detection_canny(image_location=image_input)
    edges = Output(run_edge_detection.outputs.parsed_image, sdk_type=Types.Blob)

This code block creates a workflow, with one task. The workflow itself has an input (the link to an image) that gets passed into the task, and an output, which is the processed image.

Interacting with Flyte

Flyte fulfills tasks using docker images. You’ll need to build a docker image from this code before it can run in Flyte. The repo has a make target to build the docker image for you

make docker_build

If you have the flyte sandbox installed on your local machine, the image will be accessible to to your Flyte system. If you’re running a remote Flyte instance, you’ll need to upload this image to a remote registry such as Dockerhub, Amazon ECR, or Google Container Registry, so that it can be used by the Flyte system.

To upload to a remote registry, use

DOCKER_REGISRY_USERNAME={username} DOCKER_REGISTRY_PASSWORD={pass} REGISTRY=docker.io make docker_build

Replace the values above with your registry username, password, and registry endpoint.

You may need to change the IMAGE_NAME in the Makefile to reflect your namespace in the docker registry. (ex {{my docker username}}/myflyteproject)

With the image built, we just need to register the tasks and workflows. The process is the same as what we had done previously.

docker run --network host -e FLYTE_PLATFORM_URL='' {{ your docker image }} pyflyte -p myflyteproject -d development -c sandbox.config register workflows

After this, you should be able to visit the Flyte UI, and run the workflow as you did with flytesnacks previously.