2

I'm trying to local test a kubeflow component from kfp.v2.ds1 (which works on a pipeline) using pytest, but struggling with the input/output arguments together with fixtures.

Here is a code example to illustrate the issue:

First, I created a fixture to mock a dataset. This fixture is also a kubeflow component.

# ./fixtures/

    @pytest.fixture
    @component()
    def sample_df(dataset: Output[Dataset]):
         df = pd.DataFrame(
             {
                 'name': ['Ana', 'Maria', 'Josh'],
                 'age': [15, 19, 22],
             }
         )
         dataset.path += '.csv'
         df.to_csv(dataset.path, index=False)
         return

Lets suppose the component double the ages.

# ./src/
    @component()
    def double_ages(df_input: Input[Dataset], df_output: Output[Dataset]):
         df = pd.read_csv(df_input.path)
         
         double_df = df.copy()
         double_df['age'] = double_df['age']*2

         df_output.path += '.csv'
         double_df.to_csv(df_output.path, index=False)

Then, the test:

#./tests/

@pytest.mark.usefixtures("sample_df")
def test_double_ages(sample_df):

    expected_df = pd.DataFrame(
        {
            'name': ['Ana', 'Maria', 'Josh'],
            'age': [30, 38, 44],
        }
    )

    df_component = double_ages(sample_df)    # This is where I call the component, sample_df is an Input[Dataset]
    df_output = df_component.outputs['df_output']
    df = pd.read_csv(df_output.path)
    
    assert df['age'].tolist() == expected_df['age'].tolist()

But that's when the problem occurs. The Output[Dataset] that should be passed as an output, is not, so the component cannot properly work with it, then I would get the following error on assert df['age'].tolist() == expected_df['age'].tolist():

AttributeError: 'TaskOutputArgument' object has no attribute 'path'

Aparently, the object is of the type TaskOutputArgument, instead of Dataset.

Does anyone knows how to fix this? Or how to properly use pytest with kfp components? I've searched a lot on internet but couldn't find a clue about it.

Gabriel Caldas
  • 371
  • 2
  • 13

3 Answers3

7

After spending my afternoon on this, I finally figured out a way to pytest a python-based KFP component. As I found no other lead on this subject, I hope this can help:

Access the function to test

The trick is not to directly test the KFP component created by the @component decorator. However you can access the inner decorated Python function through the component attribute python_func.

Mock artifacts

Regarding the Input and Output artifacts, as you get around KFP to access and call the tested function, you have to create them manually and pass them to the function:

input_artifact = Dataset(uri='input_df_previously_saved.csv')
output_artifact = Dataset(uri='target_output_path.csv')

I had to come up with a workaround for how the Artifact.path property works (which also applies for all KFP Artifact subclasses: Dataset, Model, ...). If you look in KFP source code, you'll find that it uses the _get_path() method that returns None if the uri attribute does not start with one of the defined cloud prefixes: "gs://", "s3://" or "minio://". As we're manually building artifacts with local paths, the tested component that wants to read the path property of an artifact would read a None value.

So I made a simple method that builds a subclass of an Artifact (or a Dataset or any other Artifact child class). The built subclass is simply altered to return the uri value instead of None in this specific case of a non-cloud uri.

Your example

Putting this all together for your test and your fixture, we can get the following code to work:

  • src/double_ages_component.py: your component to test

Nothing changes here. I just added the pandas import:

from kfp.v2.dsl import component, Input, Dataset, Output

@component
def double_ages(df_input: Input[Dataset], df_output: Output[Dataset]):
    import pandas as pd

    df = pd.read_csv(df_input.path)

    double_df = df.copy()
    double_df['age'] = double_df['age'] * 2

    df_output.path += '.csv'
    double_df.to_csv(df_output.path, index=False)
  • tests/utils.py: the Artifact subclass builder
import typing

def make_test_artifact(artifact_type: typing.Type):
    class TestArtifact(artifact_type):
        def _get_path(self):
            return super()._get_path() or self.uri

    return TestArtifact

I am still not sure it is the most proper workaround. You could also manually create a subclass for each Artifact that you use (Dataset in your example). Or you could directly mock the kfp.v2.dsl.Artifact class using pytest-mock.

  • tests/conftest.py: your fixture

I separated the sample dataframe creator component from the fixture. Hence we have a standard KFP component definition + a fixture that builds its output artifact and calls its python function:

from kfp.v2.dsl import component, Dataset, Output
import pytest

from tests.utils import make_test_artifact

@component
def sample_df_component(dataset: Output[Dataset]):
    import pandas as pd

    df = pd.DataFrame({
        'name': ['Ana', 'Maria', 'Josh'],
        'age': [15, 19, 22],
    })
    dataset.path += '.csv'
    df.to_csv(dataset.path, index=False)

@pytest.fixture
def sample_df():
    # define output artifact
    output_path = 'local_sample_df.csv'  # any writable local path. I'd recommend to use pytest `tmp_path` fixture.
    sample_df_artifact = make_test_artifact(Dataset)(uri=output_path)

    # call component python_func by passing the artifact yourself
    sample_df_component.python_func(dataset=sample_df_artifact)
    # the artifact object is now altered with the new path that you define in sample_df_component (".csv" extension added)

    return sample_df_artifact

The fixture returns an artifact object referencing a selected local path where the sample dataframe has been saved to.

  • tests/test_component.py: your actual component test

Once again, the idea is to build the I/O artifact(s) and to call the component's python_func:

from kfp.v2.dsl import Dataset
import pandas as pd

from src.double_ages_component import double_ages
from tests.utils import make_test_artifact

def test_double_ages(sample_df):
    expected_df = pd.DataFrame({
        'name': ['Ana', 'Maria', 'Josh'],
        'age': [30, 38, 44],
    })

    # input artifact is passed in parameter via sample_df fixture
    # create output artifact
    output_path = 'local_test_output_df.csv'
    output_df_artifact = make_test_artifact(Dataset)(uri=output_path)

    # call component python_func
    double_ages.python_func(df_input=sample_df, df_output=output_df_artifact)

    # read output data
    df = pd.read_csv(output_df_artifact.path)

    # write your tests
    assert df['age'].tolist() == expected_df['age'].tolist()

Result

> pytest
================ test session starts ================
platform linux -- Python 3.8.13, pytest-7.1.3, pluggy-1.0.0
rootdir: /home/USER/code/kfp_tests
collected 1 item                                                                                                                                                                                                                        

tests/test_component.py .                      [100%]

================ 1 passed in 0.28s ================
nicolasg
  • 71
  • 4
  • This helped me to raise test coverage because calling just the component function does not run the code that creates the component. When I use `.python_func` I can go through all the logic of creating a component. But I need to mock real calls to `aiplatform` for example. – havryliuk Nov 09 '22 at 11:50
  • Bravo. nicely done – John R Nov 09 '22 at 21:49
  • I found this so useful I wrote it up as a blog post - hope that's OK? https://medium.com/kingfisher-technology/vertex-ai-pipelines-notebooks-in-disguise-d5956ff0434 – Michael Jones Mar 16 '23 at 19:31
2

I spent some time investigating this and my conclusion is that individual components are not meant to be unit tested by kfp's design. That means that you must rely on unit testing each component's logic, wrapping each piece of that logic in a component, and then testing the end-to-end functionality of the kfp pipeline.

I agree that it would be quite cool if there were a way to easily mock Inputs and Outputs but I dug quite deep and it does not seem like this is an intended use (or an easy hack) at this point in time.

John R
  • 1,505
  • 10
  • 18
  • Yeah, I also spent much time studying that topic and got to the same conclusion as you, that's actually sad that there's no way to use unit tests like on normal python functions. Anyways, thanks for your contribution! – Gabriel Caldas Oct 23 '22 at 03:12
  • Yes, considering integration with the cloud, you need to mock a lot of things. – havryliuk Nov 09 '22 at 11:48
1

This has worked for me. I've used create_autospec to mock the output parameters.

@dsl.component(
    base_image="pipeline:latest",
    target_image="simple:latest",
)
def simple(
    word: str,
    number: int,
    output_path: Output[Dataset],
    output_metric: Output[Metrics],
) -> None:

    output_path.metadata["meta"] = "my meta data"
    output_metric.log_metric("numbers", number)
    output_metric.log_metric("other numbers", 5678)

    simple_stage(output_path.path, word, number)
    
def test_simple(uses_temp_directory: str) -> None:
    # arrange
    dataset_file = f"{uses_temp_directory}/dataset"
    dataset = create_autospec(Dataset, metadata=dict(), path=dataset_file)
    metrics = create_autospec(Metrics)
    
    # act
    simple.python_func(
        word="my word",
        number=1234,
        output_path=dataset,
        output_metric=metrics,
    )
    
    # assert
    result = pd.read_csv(dataset_file)
    assert 1234 == len(result.index)
    metrics.log_metric.assert_has_calls(
        [call("numbers", 1234), call("other numbers", 5678)]
    )
Alexander Townsend
  • 249
  • 2
  • 10