Scripts

The Script class is an essential part of Hera’s extension on top of Argo. As Hera is a Python library, Script templates running Python become the standard template, which is reflected by the greater feature set provided for writing them.

Script Decorator

The script decorator function is a key offering of Hera to achieve near-native Python function orchestration. It allows you to call the function under a Hera context manager such as a Workflow or Steps context, and it will be treated as the intended sub-object, which would be a template when under a Workflow, or a Step when under a Steps. The function will still behave as normal outside of any Hera contexts, meaning you can write unit tests on the given function.

For advanced users: the exact mechanism of the script decorator is to prepare a Script object within the decorator, so that when your function is invoked under a Hera context, the call is redirected to the Script.__call__ function. This takes the kwargs of a Step or Task depending on whether the context manager is a Steps or a DAG. Under a Workflow itself, your function is not expected to take arguments, so the call will add the function as a template.

When decorating a function, you should pass Script parameters to the script decorator. This includes values such as the image to use, and resources to request.

from hera.workflows import Resources, script

@script(image="python:3.11", resources=Resources(memory_request="5Gi"))
def echo(message: str):
    print(message)

When calling the function under a Steps or DAG context, you should pass Step or Task kwargs, such as the name of the Step/Task, a when clause, a with_param list to loop over a given template, or arguments for the function.

with Workflow(generate_name="dag-diamond-", entrypoint="diamond") as w:
    with DAG(name="diamond"):
        A = echo(name="A", arguments={"message": "A"})
        B = echo(name="B", arguments={"message": "B"}, when=f"{A.result == 'A'}")
        C = echo(name="C", arguments={"message": "C"}, when=f"{A.result != 'A'}")
        D = echo(name="D", arguments={"message": "D"})
        A >> [B, C] >> D

Note in the DAG above, D will still run, even though C will be skipped. This is because of the depends logic resolving to C.Succeeded || C.Skipped || C.Daemoned due to Argo’s default depends logic.

Script Constructors

InlineScriptConstructor

Script templates submitted to Argo typically run the given Python function in a Python image. By default, the Python function itself is dumped to the YAML, and the Argo cluster will run that code. For the code below, we will see it directly in the output YAML.

from hera.workflows import Workflow, script

@script(add_cwd_to_sys_path=False)
def hello(s: str):
    print("Hello, {s}!".format(s=s))


with Workflow(
    generate_name="hello-world-",
    entrypoint="hello",
    arguments={"s": "world"},
) as w:
    hello()

We added add_cwd_to_sys_path=False to remove some boilerplate from the source below. You will see Hera adds a json.loads to bridge the YAML input to a Python variable:

apiVersion: argoproj.io/v1alpha1
kind: Workflow
metadata:
  generateName: hello-world-
spec:
  arguments:
    parameters:
    - name: s
      value: world
  entrypoint: hello
  templates:
  - inputs:
      parameters:
      - name: s
    name: hello
    script:
      command:
      - python
      image: python:3.8
      source: 'import json

        try: s = json.loads(r''''''{{inputs.parameters.s}}'''''')

        except: s = r''''''{{inputs.parameters.s}}''''''


        print(''Hello, {s}!''.format(s=s))'

This method of running the function is handled by the InlineScriptConstructor, called such because it constructs the Script template to run the function “inline” in the YAML.

Importing modules

A caveat of the InlineScriptConstructor is that it is quite limited - as the InlineScriptConstructor dumps your code to the source field as-is, you must also import (within the function itself) any modules you use in the function:

@script(image="python:3.10")
def my_matcher(string: str):
    import re

    print(bool(re.match("test", string)))

Note This also applies to other functions in your code - you will not be able to call functions defined outside of the scope of the script-decorated function!

If your function uses standard library imports from Python, you will be able to run your function with any standard Python image, specified by the image argument of the decorator. Therefore, if you use non-standard imports, such as numpy, you will need to use an image that includes numpy, or build your own (e.g. as a Docker image on DockerHub).

RunnerScriptConstructor

The RunnerScriptConstructor is an alternative ScriptConstructor that uses the “Hera Runner” (think of this as being like the PyTest Runner) to run your function on Argo. This avoids dumping the function to the source of a template, keeping the YAML manageable and small, and allows you to arrange your code in natural Python fashion: imports can be anywhere in the package, the script-decorated function can call other functions in the package, and the function itself can take Pydantic objects as arguments. The use of the RunnerScriptConstructor necessitates building your own image, as the Hera Runner runs the function by referencing it as an entrypoint of your module. The image used by the script should be built from the source code package itself and its dependencies, so that the source code’s functions, dependencies, and Hera itself are available to run.

The RunnerScriptConstructor is an experimental feature and must be enabled with the script_runner feature flag, as described in the experimental features section.

global_config.experimental_features["script_runner"] = True

A function can set its constructor to "runner" to use the RunnerScriptConstructor, or use the global_config.set_class_defaults function to set it once for all script-decorated functions. We can write a script template function using Pydantic objects such as:

global_config.set_class_defaults(Script, constructor="runner")

class Input(BaseModel):
    a: int
    b: str = "foo"

class Output(BaseModel):
    output: List[Input]

@script()
def my_function(input: Input) -> Output:
    return Output(output=[input])

This creates a template in YAML that looks like:

- name: my-function
  inputs:
    parameters:
    - name: input
  script:
    command:
    - python
    args:
    - -m
    - hera.workflows.runner
    - -e
    - examples.workflows.callable_script:my_function
    image: my-image-with-python-source-code-and-dependencies
    source: '{{inputs.parameters}}'

You will notice some pecularities of this template. Firstly, it is running the hera.workflows.runner module, rather than a user-module such as examples.workflows.callable_script. Instead, the -e arg specifies the --entrypoint to be called by the runner, in this case the my_function of the examples.workflows.callable_script module. We do not give a real image here, but we assume it exists in this example. Finally, the source parameter is passed the inputs.parameters of the template. This is because the Hera Runner relies on a mechanism in Argo where the value passed to source is dumped to a file, and then the filename is passed as the final arg to the command. Therefore, the source will actually contain a list of parameters as dictionaries, which are dumped to a file which is passed to hera.workflows.runner. Of course, this is all handled for you!

Script Annotations

Annotation syntax is an experimental feature using typing.Annotated for Parameters and Artifacts to declare inputs and outputs for functions decorated as scripts. They use Annotated as the type in the function parameters and allow us to simplify writing scripts with parameters and artifacts that require additional fields such as a description or alternative name.

This feature must be enabled by setting the experimental_feature flag script_annotations on the global config.

global_config.experimental_features["script_annotations"] = True

Parameters

In Hera, we can currently specify inputs inside the @script decorator as follows:

@script(
    inputs=[
        Parameter(name="an_int", description="an_int parameter", default=1, enum=[1, 2, 3]),
        Parameter(name="a_bool", description="a_bool parameter", default=True, enum=[True, False]),
        Parameter(name="a_string", description="a_string parameter", default="a", enum=["a", "b", "c"])
    ]
)
def echo_all(an_int=1, a_bool=True, a_string="a"):
    print(an_int)
    print(a_bool)
    print(a_string)

Notice how the name and default values are duplicated for each Parameter. Using annotations, we can rewrite this as:

@script()
def echo_all(
    an_int: Annotated[int, Parameter(description="an_int parameter", default=1, enum=[1, 2, 3])],
    a_bool: Annotated[bool, Parameter(description="a_bool parameter", default=True, enum=[True, False])],
    a_string: Annotated[str, Parameter(description="a_string parameter", default="a", enum=["a", "b", "c"])]
):
    print(an_int)
    print(a_bool)
    print(a_string)

The fields allowed in the Parameter annotations are: name, default, enum, and description.

Artifacts

Note: Artifact annotations are only supported when used with the RunnerScriptConstructor.

The feature is even more powerful for Artifacts. In Hera we are currently able to specify Artifacts in inputs, but the given path is not programmatically linked to the code within the function unless defined outside the scope of the function:

@script(inputs=Artifact(name="my-artifact", path="/tmp/file"))
def read_artifact():
    with open("/tmp/file") as a_file:  # Repeating "/tmp/file" is prone to human error!
        print(a_file.read())

# or

MY_PATH = "/tmp/file"  # Now accessible outside of the function scope!
@script(inputs=Artifact(name="my-artifact", path=MY_PATH))
def read_artifact():
    with open(MY_PATH) as a_file:
        print(a_file.read())

By using annotations we can avoid repeating the path of the file, and the function can use the variable directly as a Path object, with its value already set to the given path:

@script(constructor="runner")
def read_artifact(an_artifact: Annotated[Path, Artifact(name="my-artifact", path="/tmp/file")]):
    print(an_artifact.read_text())

The fields allowed in the Artifact annotations are: name, path, and loader.

Artifact Loaders

In case you want to load an object directly from the path of the Artifact, we allow two types of loaders besides the default Path behaviour used when no loader is specified. The ArtifactLoader enum provides file and json loaders.

None loader

With None set as the loader (which is by default) in the Artifact annotation, the path attribute of Artifact is extracted and used to provide a pathlib.Path object for the given argument, which can be used directly in the function body. The following example is the same as above except for explicitly setting the loader to None:

@script(constructor="runner")
def read_artifact(
    an_artifact: Annotated[Path, Artifact(name="my-artifact", path="/tmp/file", loader=None)]
):
    print(an_artifact.read_text())

file loader

When the loader is set to file, the function parameter type should be str, and will contain the contents string representation of the file stored at path (essentially performing path.read_text() automatically):

@script(constructor="runner")
def read_artifact(
    an_artifact: Annotated[str, Artifact(name="my-artifact", path="/tmp/file", loader=ArtifactLoader.file)]
) -> str:
    return an_artifact

This loads the contents of the file at "/tmp/file" to the argument an_artifact and subsequently can be used as a string inside the function.

json loader

When the loader is set to json, the contents of the file at path are read and parsed to a dictionary via json.load (essentially performing json.load(path.open()) automatically). By specifying a Pydantic type, this dictionary can even be automatically parsed to that type:

class MyArtifact(BaseModel):
    a = "a"
    b = "b"


@script(constructor="runner")
def read_artifact(
    an_artifact: Annotated[MyArtifact, Artifact(name="my-artifact", path="/tmp/file", loader=ArtifactLoader.json)]
) -> str:
    return an_artifact.a + an_artifact.b

Here, we have a json representation of MyArtifact such as {"a": "hello ", "b": "world"} stored at "/tmp/file". We can load it with ArtifactLoader.json and then use an_artifact as an instance of MyArtifact inside the function, so the function will return "hello world".

Function parameter name aliasing

Script annotations can work on top of the RunnerScriptConstructor for name aliasing of function parameters, in particular to allow a public kebab-case parameter, while using a snake_case Python function parameter. When using a RunnerScriptConstructor, an environment variable hera__script_annotations will be added to the Script template (visible in the exported YAML file).

Outputs

Note: Output annotations are only supported when used with the RunnerScriptConstructor.

There are two ways to specify output Artifacts and Parameters.

Function return annotations

Function return annotations can be used to specify the output type information for output Artifacts and Parameters, and the function should return a value or tuple. An example can be seen here.

For a simple hello world output artifact example we currently have:

@script(outputs=Artifact(name="hello-artifact", path="/tmp/hello_world.txt"))
def hello_world():
   with open("/tmp/hello_world.txt", "w") as f:
       f.write("Hello, world!")

The new approach allows us to avoid duplication of the path, which is now optional, and results in more readable code:

@script()
def hello_world() -> Annotated[str, Artifact(name="hello-artifact")]:
    return "Hello, world!"

For Parameters we have a similar syntax:

@script()
def hello_world() -> Annotated[str, Parameter(name="hello-param")]:
    return "Hello, world!"

The returned values will be automatically saved in files within the Argo container according to this schema: * /hera/outputs/parameters/<name> * /hera/outputs/artifacts/<name>

These outputs are also exposed in the outputs section of the template in YAML.

The object returned from the function can be of any serialisable Pydantic type (or basic Python type) and must be Annotated as an Artifact or Parameter. The Parameter/Artifact’s name will be used for the path of the output unless provided: * if the annotation is an Artifact with a path, we use that path * if the annotation is a Parameter, with a value_from that contains a path, we use that path

See the following two functions for specifying custom paths:

@script()
def hello_world() -> Annotated[str, Artifact(name="hello-artifact", path="/tmp/hello_world_art.txt")]:
    return "Hello, world!"

@script()
def hello_world() -> Annotated[str, Parameter(name="hello-param", value_from={"path": "/tmp/hello_world_param.txt"})]:
    return "Hello, world!"

For multiple outputs, the return type should be a Tuple of arbitrary Pydantic types with individual Parameter/Artifact annotations, and the function must return a tuple from the function matching these types:

@script()
def func(...) -> Tuple[
    Annotated[arbitrary_pydantic_type_a, Artifact],
    Annotated[arbitrary_pydantic_type_b, Parameter],
    Annotated[arbitrary_pydantic_type_c, Parameter],
    ...]:
    return output_a, output_b, output_c

Input-Output function parameters

Hera also allows output Parameter/Artifacts as part of the function signature when specified as a Path type, allowing users to write to the path as an output, without needing an explicit return. They require an additional field output=True to distinguish them from the input parameters and must have an underlying Path type (or another type that will write to disk).

@script()
def func(..., output_param: Annotated[Path, Parameter(output=True, global_name="...", name="")]) -> Annotated[arbitrary_pydantic_type, OutputItem]:
    output_param.write_text("...")
    return output

The parent outputs directory, /hera/outputs by default, can be set by the user. This is done by adding:

global_config.set_class_defaults(RunnerScriptConstructor, outputs_directory="user/chosen/outputs")

Comments