Skip to main content
Flyte organizes execution around a small set of composable primitives. Understanding these four concepts gives you a complete mental model of how Flyte works.

Tasks

The fundamental unit of compute in Flyte. A task is a strongly typed, versioned, containerized Python function that runs on a Kubernetes Pod.

Workflows

A directed acyclic graph (DAG) that composes tasks and other workflows into a meaningful execution pipeline.

Launch Plans

Templates for workflow executions. Bind default or fixed inputs, define schedules, and configure notifications.

Projects and Domains

Multi-tenancy primitives. Projects namespace your workflows, and domains (development, staging, production) control execution environments.

​
How these concepts fit together

Here is how the four concepts relate to each other at execution time:
  1. You write tasks — Python functions decorated with @task. Each task has typed inputs and outputs.
  2. You compose tasks into workflows — Python functions decorated with @workflow. The workflow defines the DAG structure and the data flow between tasks.
  3. You create launch plans to trigger workflow executions. Every registered workflow automatically gets a default launch plan. You can create additional launch plans with different default inputs, fixed inputs, or schedules.
  4. You register all of this under a project and domain. The project is a logical namespace (e.g., my-ml-project). The domain controls the execution environment (e.g., development vs. production).
Project: my-ml-project
├── Domain: development
│   ├── Workflow: training_wf (v1, v2, v3)
│   └── Launch Plan: nightly_training_lp (cron schedule)
└── Domain: production
    ├── Workflow: training_wf (v3)
    └── Launch Plan: nightly_training_lp (cron schedule, fixed inputs)

​
A quick example

Here is a complete example showing all four concepts working together: 
from typing import List
from flytekit import task, workflow, LaunchPlan, CronSchedule


# 1. Define tasks
@task
def mean(values: List[float]) -> float:
    return sum(values) / len(values)


@task
def standard_deviation(values: List[float], mu: float) -> float:
    from math import sqrt
    variance = sum([(x - mu) ** 2 for x in values])
    return sqrt(variance)


@task
def standard_scale(values: List[float], mu: float, sigma: float) -> List[float]:
    return [(x - mu) / sigma for x in values]


# 2. Compose tasks into a workflow
@workflow
def standard_scale_workflow(values: List[float]) -> List[float]:
    mu = mean(values=values)
    sigma = standard_deviation(values=values, mu=mu)
    return standard_scale(values=values, mu=mu, sigma=sigma)


# 3. Create a launch plan with default inputs and a schedule
nightly_lp = LaunchPlan.get_or_create(
    standard_scale_workflow,
    name="nightly_scale_lp",
    default_inputs={"values": [1.0, 2.0, 3.0, 4.0, 5.0]},
    schedule=CronSchedule(schedule="0 0 * * *"),  # daily at midnight
)
To run this workflow locally: 
pyflyte run example.py standard_scale_workflow --values '[1.0,2.0,3.0,4.0,5.0]'
To register and run on a Flyte cluster: 
pyflyte run --remote example.py standard_scale_workflow --values '[1.0,2.0,3.0,4.0,5.0]'

​
Versioning

Each task and workflow in Flyte is versioned and immutable. Once registered, a specific version of a task or workflow (identified by its project, domain, name, and version) cannot be changed. This is a deliberate design decision to guarantee reproducibility: you can always re-run an old workflow version with identical behavior. By default, pyflyte register uses the current Git SHA as the version string.

​
Explore each concept

Tasks in depth

Task types, resource allocation, caching, retries, timeouts, and plugins.

Workflows in depth

DAG construction, promises, subworkflows, and the >> dependency operator.

Launch Plans in depth

Default inputs, fixed inputs, cron schedules, fixed-rate schedules, and activation.

Projects and Domains in depth

Multi-tenancy, execution isolation, and promoting workflows across environments.

Build docs developers (and LLMs) love

Get started for freeTalk to us