Generally, we follow the Google Python Style Guide. This document covers Prefect-specific styles and practices.

Imports

This is a brief collection of rules and guidelines for handling imports in this repository.

Imports in __init__ files

Leave __init__ files empty unless exposing an interface. If you must expose objects to present a simpler API, please follow these rules.

Exposing objects from submodules

If importing objects from submodules, the __init__ file should use a relative import. This is required for type checkers to understand the exposed interface.

# Correct
from .flows import flow

# Wrong
from prefect.flows import flow

Exposing submodules

Generally, submodules should not be imported in the __init__ file. You should only expose submodules when the module is designed to be imported and used as a namespaced object.

For example, we do this for our schema and model modules. This is because it’s important to know if you are working with an API schema or database model—both of which may have similar names.

import prefect.server.schemas as schemas

# The full module is accessible now
schemas.core.FlowRun

If exposing a submodule, use a relative import like when you’re exposing an object.

# Correct
from . import flows

# Wrong
import prefect.flows

Importing to run side-effects

Another use case for importing submodules is to perform global side-effects that occur when they are imported.

Often, global side-effects on import are a dangerous pattern. But there are a couple acceptable use cases for this:

  • To register dispatchable types, for example, prefect.serializers.
  • To extend a CLI app, for example, prefect.cli.

Imports in modules

Importing other modules

The from syntax is recommended for importing objects from modules. You should not import modules with the from syntax.

# Correct
import prefect.server.schemas  # use with the full name
import prefect.server.schemas as schemas  # use the shorter name

# Wrong
from prefect.server import schemas

You should not use relative imports unless it’s in an __init__.py file.

# Correct
from prefect.utilities.foo import bar
# Wrong
from .utilities.foo import bar

You should never use imports that are dependent on file location without explicit indication it is relative. This avoids confusion about the source of a module.

# Correct
from . import test

Resolving circular dependencies

Sometimes, you must defer an import and perform it within a function to avoid a circular dependency:

## This function in `settings.py` requires a method from the global `context` but the context
## uses settings
def from_context():
    from prefect.context import get_profile_context

    ...

Avoid circular dependencies. They often reveal entanglement in the design.

Place all deferred imports at the top of the function.

If you are just using the imported object for a type signature, use the TYPE_CHECKING flag:

# Correct
from typing import TYPE_CHECKING

if TYPE_CHECKING:
    from prefect.server.schemas.states import State

def foo(state: "State"):
    pass

Usage of the type within the module requires quotes; for example, "State", since it is not available at runtime.

Importing optional requirements

We do not have a best practice for this yet. See the kubernetes, docker, and distributed implementations for now.

Delaying expensive imports

Sometimes imports are slow, but it’s important to keep the prefect module import times fast. In these cases, lazily import the slow module by deferring import to the relevant function body. For modules consumed by many functions, use the optional requirements pattern instead.

Command line interface (CLI) output messages

When executing a command that creates an object, the output message should offer:

  • A short description of what the command just did.
  • A bullet point list, rehashing user inputs, if possible.
  • Next steps, like the next command to run, if applicable.
  • Other relevant, pre-formatted commands that can be copied and pasted, if applicable.
  • A new line before the first line, and after the last line.

Output Example:

$ prefect work-queue create testing

Created work queue with properties:
    name - 'abcde'
    uuid - 940f9828-c820-4148-9526-ea8107082bda
    tags - None
    deployment_ids - None

Start an agent to pick up flows from the created work queue:
    prefect agent start -q 'abcde'

Inspect the created work queue:
    prefect work-queue inspect 'abcde'

Additionally:

  • Wrap generated arguments in apostrophes (’) to ensure validity by using suffixing formats with !r.
  • Indent example commands, instead of wrapping in backticks (`).
  • Use placeholders if you cannot completely pre-format the example.
  • Capitalize placeholder labels and wrap them in less than (<) and greater than (>) signs.
  • Utilize textwrap.dedent to remove extraneous spacing for strings with triple quotes (""").

Placeholder Example:

Create a work queue with tags:
    prefect work-queue create '<WORK QUEUE NAME>' -t '<OPTIONAL TAG 1>' -t '<OPTIONAL TAG 2>'

Dedent Example:

from textwrap import dedent
...
output_msg = dedent(
    f"""
    Created work queue with properties:
        name - {name!r}
        uuid - {result}
        tags - {tags or None}
        deployment_ids - {deployment_ids or None}

    Start an agent to pick up flows from the created work queue:
        prefect agent start -q {name!r}

    Inspect the created work queue:
        prefect work-queue inspect {name!r}
    """
)

API versioning

Client and server communication

You can run the Prefect client separately from Prefect server, and communicate entirely through an API. The Prefect client includes anything that runs task or flow code, (for example, agents and the Python client); or any consumer of Prefect metadata (for example, the Prefect UI and CLI).

Prefect server stores this metadata and serves it through the REST API.

API version header

Sometimes, we have to make breaking changes to the API. To check a Prefect client’s compatibility with the API it’s making requests to, every API call the client makes includes a three-component API_VERSION header with major, minor, and patch versions.

For example, a request with the X-PREFECT-API-VERSION=3.2.1 header has a major version of 3, minor version 2, and patch version 1.

Change this version header by modifying the API_VERSION constant in prefect.server.api.server.

Breaking changes to the API

A breaking change means that your code needs to change to use a new version of Prefect. We avoid breaking changes whenever possible.

When making a breaking change to the API, we consider if the change is backwards compatible for clients. This means that the previous version of the client can still make calls against the updated version of the server code. This might happen if the changes are purely additive, such as adding a non-critical API route. In these cases, we aim to bump the patch version.

In almost all other cases, we bump the minor version, which denotes a non-backwards-compatible API change. We have reserved the major version changes to denote a backwards compatible change that is significant in some way, such as a major release milestone.

Version composition

Versions are composed of three parts: MAJOR.MINOR.PATCH. For example, the version 2.5.0 has a major version of 2, a minor version of 5, and patch version of 0.

Occasionally, we add a suffix to the version such as rc, a, or b. These indicate pre-release versions that users can opt into for testing and experimentation prior to a generally available release.

Each release will increase one of the version numbers. If we increase a number other than the patch version, the versions to the right of it reset to zero.

Prefect’s versioning scheme

Prefect increases the major version when significant and widespread changes are made to the core product.

Prefect increases the minor version when:

  • Introducing a new concept that changes how to use Prefect
  • Changing an existing concept in a way that fundamentally alters its usage
  • Removing a deprecated feature

Prefect increases the patch version when:

  • Making enhancements to existing features
  • Fixing behavior in existing features
  • Adding new capabilities to existing concepts
  • Updating dependencies

Deprecation

At times, Prefect will deprecate a feature. A feature is deprecated when it will no longer be maintained. Frequently, a deprecated feature will have a new and improved alternative. Deprecated features will be retained for at least 3 minor version increases or 6 months, whichever is longer. We may retain deprecated features longer than this time period.

Prefect will sometimes include changes to behavior to fix a bug. These changes are not categorized as breaking changes.

Client compatibility with Prefect

When running a Prefect server, you are in charge of ensuring the version is compatible with those of the clients that are using the server. Prefect aims to maintain backwards compatibility with old clients for each server release. In contrast, sometimes you cannot use new clients with an old server. The new client may expect the server to support capabilities that it does not yet include. For this reason, we recommend that all clients are the same version as the server or older.

For example, you can use a client on 2.1.0 with a server on 2.5.0. You cannot use a client on 2.5.0 with a server on 2.1.0.

Client compatibility with Cloud

Prefect Cloud targets compatibility with all versions of Prefect clients. If you encounter a compatibility issue, please file a bug report.