Skip to content

Latest commit

 

History

History
299 lines (200 loc) · 20.1 KB

README.md

File metadata and controls

299 lines (200 loc) · 20.1 KB
Raw

Tests

Recommenders test pipeline is one of the most sophisticated MLOps pipelines in the open-source community. We execute tests in the three environments we support: CPU, GPU, and Spark, mirroring the tests in each Python version we support. We not only tests the library, but also the Jupyter notebooks in the examples folder.

The reason to have this extensive test infrastructure is to ensure that the code is reproducible by the community and that we can maintain the project with a small number of core contributors.

We currently execute over a thousand tests in the project, and we are always looking for ways to improve the test coverage. To get the exact number of tests, you can run pytest tests --collect-only, and then multiply the number of tests by the number of Python versions we support.

In this document we show our test infrastructure and how to contribute tests to the repository.

Table of Contents

Test workflows

All the tests in this repository are part of the following two workflows: the PR gate and the nightly builds.

PR gates are the set of tests executed after doing a pull request and they should be quick. The objective is to validate that the code is not breaking anything before merging it. The PR gate should not surpass 20-30 minutes.

The nightly builds are tests executed asynchronously and can take hours. Some tests take so long that they cannot be executed in a PR gate, therefore they are executed asynchronously in the nightly builds.

Notice that the errors in the nightly builds are detected after the code has been merged. This is the reason why, with nightly builds, it is interesting to have a two-level branching strategy. In the standard one-level branching strategy, all pull requests go to the main branch. If a nightly build fails, then the main branch has broken code. In the two-level branching strategy, a pre-production or staging branch is where developers send pull requests to. The main branch is only updated from the staging branch after the nightly builds are successful. This way, the main branch always has working code.

Categories of tests

The tests in this repository are divided into the following categories:

  • Data validation tests: In the data validation tests, we ensure that the schema for input and output data for each function in the pipeline matches the desired prespecified schema, that the data is available and has the correct size. They should be fast and can be added to the PR gate.
  • Unit tests: In the unit tests we just make sure the python utilities and notebooks run correctly. Unit tests are fast, ideally less than 5min and are run in every pull request. They belong to the PR gate. For this type of tests, synthetic data can be used.
  • Functional tests: These tests make sure that the components of the project not just run but their function is correct. For example, we want to test that an ML model evaluation of RMSE gives a positive number. These tests can be run asynchronously in the nightly builds and can take hours. In these tests, we want to use real data.
  • Integration tests: We want to make sure that the interaction between different components is correct. For example, the interaction between data ingestion pipelines and the compute where the model is trained, or between the compute and a database. These tests can be of variable length, if they are fast, we could add them to the PR gate, otherwise, we will add them to the nightly builds. For this type of tests, synthetic and real data can be used.
  • Smoke tests: The smoke tests are gates to the slow tests in the nightly builds to detect quick errors. If we are running a test with a large dataset that takes 4h, we want to create a faster version of the large test (maybe with a small percentage of the dataset or with 1 epoch) to ensure that it runs end-to-end without obvious failures. Smoke tests can run sequentially with functional or integration tests in the nightly builds, and should be fast, ideally less than 20min. They use the same type of data as their longer counterparts.
  • Performance test: The performance tests are tests that measure the computation time or memory footprint of a piece of code and make sure that this is bounded between some limits. Another kind of performance testing can be a load test to measure an API response time, this can be specially useful when working with large deep learning models. For this type of tests, synthetic data can be used.
  • Responsible AI tests: Responsible AI tests are test that enforce fairness, transparency, explainability, human-centeredness, and privacy.
  • Security tests: Security tests are tests that make sure that the code is not vulnerable to attacks. These can detect potential security issues either in python packages or the underlying OS, in addition to scheduled scans in the production pipelines.
  • Regression tests: In some situations, we are migrating from a deprecated version to a new version of the code, or maybe we are maintaining two versions of the same library (i.e. Tensorflow v1 and v2). Regression tests make sure that the code works in both versions of the code. These types of tests sometimes are done locally, before upgrading to the new version, or they can be included in the tests pipelines if we want to execute them recurrently.

For more information, see a quick introduction testing.

Scalable test infrastructure with AzureML

AzureML is used to run the existing unit, smoke and integration tests. AzureML benefits include being able to run the tests in parallel, managing the compute environment by automatically turning it on/off, automatic logging of artifacts from test runs and more. GitHub is used as a control plane to configure and run the tests on AzureML.

In the following figure we show a workflow on how the tests are executed via AzureML:

GitHub workflows azureml-unit-tests.yml, azureml-cpu-nightly.yml, azureml-gpu-nightly.yml and azureml-spark-nightly located in .github/workflows/ are used to run the tests on AzureML. The parameters to configure AzureML are defined in the workflow yml files. The tests are divided into groups and each workflow triggers these test groups in parallel, which significantly reduces end-to-end execution time.

There are three scripts used with each workflow, all of them are located in ci/azureml_tests:

  • submit_groupwise_azureml_pytest.py: this script uses parameters in the workflow yml to set up the AzureML environment for testing using the AzureML SDK.
  • run_groupwise_pytest.py: this script uses pytest to run the tests of the libraries and notebooks. This script runs in an AzureML workspace with the environment created by the script above.
  • test_groups.py: this script defines the groups of tests. If the tests are part of the unit tests, the total compute time of each group should be less than 15min. If the tests are part of the nightly builds, the total time of each group should be less than 35min.

How to contribute tests to the repository

In this section we show how to create tests and add them to the test pipeline. The steps you need to follow are:

  1. Create your code in the library and/or notebooks.
  2. Design the unit tests for the code.
  3. If you have written a notebook, design the notebook tests and check that the metrics they return is what you expect.
  4. Add the tests to the AzureML pipeline in the corresponding test group.

Please note that if you don't add your tests to the pipeline, they will not be executed.

How to create tests for the Recommenders library

You want to make sure that all your code works before you submit it to the repository. Here are some guidelines for creating the tests:

  • It is better to create multiple small tests than one large test that checks all the code.
  • Use @pytest.fixture to create data in your tests.
  • Follow the pattern assert computation == value, for example:
assert results["precision"] == pytest.approx(0.330753)
  • Check always the limits of your computations, for example, you want to check that the RMSE between two equal vectors is 0:
assert rmse(rating_true, rating_true) == 0
assert rmse(rating_true, rating_pred) == pytest.approx(7.254309)
  • Use the operator == with values. Use the operator is in singletons like None, True or False.
  • Make explicit asserts. In other words, make sure you assert to something (assert computation == value) and not just assert computation.
  • Use the mark @pytest.mark.gpu if you want the test to be executed in a GPU environment. Use @pytest.mark.spark if you want the test to be executed in a Spark environment.
  • Use @pytest.mark.notebooks if you are testing a notebook.

How to create tests for the notebooks

For testing the notebooks of this repo, we developed the Recommenders notebook executor, that enables you to parametrize and execute notebooks for testing.

The notebook executor is located in recommenders/utils/notebook_utils.py. The main functions are:

  • execute_notebook: Executes a notebook and saves the output in a new notebook. Optionally, you can inject parameters to the notebook. For that, you need to tag the cells with the tag parameters. Every cell tagged with parameters can be injected with the variables passed in the parameters dictionary.
  • store_metadata: Stores the output of a variable. The output is stored in the metadata of the Jupyter notebook and can be read by read_notebook function.
  • read_notebook: Reads the output notebook and returns a dictionary with the variables recorded with store_metadata.

Developing PR gate tests with the notebook executor

Executing a notebook with the Recommenders notebook executor is easy, this is what we mostly do in the unit tests. Next, we show just one of the tests that we have in tests/unit/examples/test_notebooks_python.py.

import pytest
from recommenders.utils.notebook_utils import execute_notebook

@pytest.mark.notebooks
def test_sar_single_node_runs(notebooks, output_notebook, kernel_name):
    notebook_path = notebooks["sar_single_node"]
    execute_notebook(notebook_path, output_notebook, kernel_name=kernel_name)

Notice that the input of the function is a fixture defined in conftest.py. For more information, please see the definition of fixtures in PyTest.

For executing this test, first make sure you are in the correct environment as described in the SETUP.md:

Notice that the next instruction executes the tests from the root folder.

pytest tests/unit/examples/test_notebooks_python.py::test_sar_single_node_runs

Developing nightly tests with the notebook executor

A more advanced option is used in the nightly tests, where we not only execute the notebook, but inject parameters and recover the computed metrics.

The first step is to tag the parameters that we are going to inject. For it we need to modify the notebook. We will add a tag with the name parameters. To add a tag, go the notebook menu, View, Cell Toolbar and Tags. A tag field will appear on every cell. The variables in the cell tagged with parameters can be injected. The typical variables that we inject are MOVIELENS_DATA_SIZE, EPOCHS and other configuration variables for our algorithms.

The way the notebook executor works to inject parameters is very simple, it generates a copy of the notebook (in our code we call it OUTPUT_NOTEBOOK), and replaces the cell with the tag parameters with the injected variables.

The second modification that we need to do to the notebook is to record the metrics we want to test using store_metadata("output_variable", python_variable_name). We normally use the last cell of the notebook to record all the metrics. These are the metrics that we are going to control in the smoke and functional tests.

This is an example on how we do a smoke test. The complete code can be found in smoke/examples/test_notebooks_python.py:

import pytest

from recommenders.utils.notebook_utils import execute_notebook, read_notebook

TOL = 0.05
ABS_TOL = 0.05

def test_sar_single_node_smoke(notebooks, output_notebook, kernel_name):
    notebook_path = notebooks["sar_single_node"]
    execute_notebook(
        notebook_path,
        output_notebook,
        kernel_name=kernel_name,
        parameters=dict(TOP_K=10, MOVIELENS_DATA_SIZE="100k"),
    )
    results = read_notebook(output_notebook)
    
    assert results["precision"] == pytest.approx(0.330753, rel=TOL, abs=ABS_TOL)
    assert results["recall"] == pytest.approx(0.176385, rel=TOL, abs=ABS_TOL)

As it can be seen in the code, we are injecting the dataset size and the top k and we are recovering the precision and recall at k.

For executing this test, first make sure you are in the correct environment as described in the SETUP.md:

Notice that the next instructions execute the tests from the root folder.

pytest tests/smoke/examples/test_notebooks_python.py::test_sar_single_node_smoke

How to add tests to the AzureML pipeline

To add a new test to the AzureML pipeline, add the test path to an appropriate test group listed in test_groups.py.

Tests in group_cpu_xxx groups are executed on a CPU-only AzureML compute cluster node. Tests in group_gpu_xxx groups are executed on a GPU-enabled AzureML compute cluster node with GPU related dependencies added to the AzureML run environment. Tests in group_pyspark_xxx groups are executed on a CPU-only AzureML compute cluster node, with the PySpark related dependencies added to the AzureML run environment.

It's important to keep in mind while adding a new test that the runtime of the test group should not exceed the specified threshold in test_groups.py.

Example of adding a new test:

  1. In the environment that you are running your code, first see if there is a group whose total runtime is less than the threshold.
"group_spark_001": [  # Total group time: 271.13s
    "tests/data_validation/recommenders/datasets/test_movielens.py::test_load_spark_df",  # 4.33s+ 25.58s + 101.99s + 139.23s
],
  1. Add the test to the group, add the time it takes to compute, and update the total group time.
"group_spark_001": [  # Total group time: 571.13s
    "tests/data_validation/recommenders/datasets/test_movielens.py::test_load_spark_df",  # 4.33s+ 25.58s + 101.99s + 139.23s
    #
    "tests/path/to/test_new.py::test_new_function", # 300s
],
  1. If all the groups of your environment are above the threshold, add a new group.

Setup GitHub Actions with AzureML compute clusters

In this section we explain how to create the AzureML infrastructure to run the tests in GitHub Actions.

In order to execute the tests in Recommenders, we need two types of virtual machines: ones without GPU, to execute the CPU and Spark tests, and ones with GPU, to execute the GPU tests. Therefore, the first step is to request enough quota for your subscription.

Then, follow the steps below to create the AzureML infrastructure:

  1. Create a new AzureML workspace.
    • Name: azureml-test-workspace
    • Resource group: recommenders_project_resources
    • Location: Make sure you have enough quota in the location you choose
  2. Create two new clusters: cpu-cluster and gpu-cluster. Go to compute, then compute cluster, then new.
    • Select the CPU VM base. Anything above 64GB of RAM, and 8 cores should be fine.
    • Select the GPU VM base. Anything above 56GB of RAM, and 6 cores, and an NVIDIA K80 should be fine.
  3. Add the subscription ID to GitHub action secrets here. Create a new repository secret called AZUREML_TEST_SUBID and add the subscription ID as the value.
  4. Make sure you have installed Azure CLI, and that you are logged in: az login.
  5. Select your subscription: az account set -s $AZURE_SUBSCRIPTION_ID.
  6. Create a Service Principal: az ad sp create-for-rbac --name $SERVICE_PRINCIPAL_NAME --role contributor --scopes /subscriptions/$AZURE_SUBSCRIPTION_ID --json-auth. This will output a JSON blob with the credentials of the Service Principal: 
    {
    "clientId": "XXXXXXXXXXXXXXXXXXXXX",
    "clientSecret": "XXXXXXXXXXXXXXXXXXXXX",
    "subscriptionId": "XXXXXXXXXXXXXXXXXXXXX",
    "tenantId": "XXXXXXXXXXXXXXXXXXXXX",
    "activeDirectoryEndpointUrl": "https://login.microsoftonline.com",
    "resourceManagerEndpointUrl": "https://management.azure.com/",
    "activeDirectoryGraphResourceId": "https://graph.windows.net/",
    "sqlManagementEndpointUrl": "https://management.core.windows.net:8443/",
    "galleryEndpointUrl": "https://gallery.azure.com/",
    "managementEndpointUrl": "https://management.core.windows.net/"
}
  7. Add the output as github's action secret AZUREML_TEST_CREDENTIALS under repository's Settings > Security > Secrets and variables > Actions.

How to execute tests in your local environment

To manually execute the tests in the CPU, GPU or Spark environments, first make sure you are in the correct environment as described in the SETUP.md.

CPU tests

Note that the next instructions execute the tests from the root folder.

For executing the CPU tests for the utilities:

pytest tests -m "not notebooks and not spark and not gpu" --durations 0 --disable-warnings

For executing the CPU tests for the notebooks:

pytest tests -m "notebooks and not spark and not gpu" --durations 0 --disable-warnings

If you want to execute a specific test, you can use the following command:

pytest tests/data_validation/recommenders/datasets/test_mind.py::test_mind_url --durations 0 --disable-warnings

If you want to execute any of the tests types (data_validation, unit, smoke, functional, etc.) you can use the following command:

pytest tests/data_validation -m "not notebooks and not spark and not gpu" --durations 0 --disable-warnings

GPU tests

For executing the GPU tests for the utilities:

pytest tests -m "not notebooks and not spark and gpu" --durations 0 --disable-warnings

For executing the GPU tests for the notebooks:

pytest tests -m "notebooks and not spark and gpu" --durations 0 --disable-warnings

Spark tests

For executing the PySpark tests for the utilities:

pytest tests -m "not notebooks and spark and not gpu" --durations 0 --disable-warnings

For executing the PySpark tests for the notebooks:

pytest tests -m "notebooks and spark and not gpu" --durations 0 --disable-warnings

NOTE: Adding --durations 0 shows the computation time of all tests.

NOTE: Adding --disable-warnings will disable the warning messages.

In order to skip a test because there is an OS or upstream issue which cannot be resolved you can use pytest annotations.

Example:

@pytest.mark.skip(reason="<INSERT VALID REASON>")
@pytest.mark.skipif(sys.platform == 'win32', reason="Not implemented on Windows")
def test_to_skip():
    assert False