Moved. See All new content goes to the new site. This is a legacy, and will likely be dropped five years after the last post in Jan 2023.

Tuesday, November 22, 2022

Testing with PySpark

This isn't about details of pySpark. This is about the philosophy of testing when working with a large, complex framework, like pySpark, pandas, numpy, or whatever. 


Use data subsets. 

Write unit tests for the functions that process the data.

Don't test pyspark itself. Test the code you write.

Some History

I've worked with folks -- data scientists specifically -- without a deep background in software engineering.

When we said their model-building applications needed a test case, they supplied the test case they used to validate the model.

Essentially, their test script ran the entire training set. Built the model. Did extensive statistical testing on the resulting decisions made by the model. The test case asserted that the stats were "good." In fact, they recapitulated the entire model review process that had gone on in the data science community to get the model from "someone's idea" to a "central piece of the business." 

The test case ran for hours and required a huge server loaded up with GPUs. It cost a fortune to run. And. It tended to timeout the deployment pipeline.

This isn't what we mean by "test." Our mistake.

We had to explain that a unit test demonstrates the code works. That was all. It shouldn't involve the full training set of data and the full training process with all the hyperparameter tuning and hours of compute time. We don't need to revalidate your model. We want to know the code won't crash. We'd like 100% code coverage. But the objective is little more than show it won't crash when we deploy it.

It was difficult to talk them down from full training sets. They couldn't see the value in testing code in isolation. A phrase like "just enough data to prove the thing could plausibly work with real data" seemed to resonate. 

A few folks complained that a numpy array with a few rows didn't really show very much. We had to explain (more than once) that we didn't really want to know all the algorithmic and performance nuances. We mostly wanted to know it wouldn't crash when we applied it to production data. We agreed with them the test case didn't show much. We weren't qualified to revalidate the model; we were only qualified to run their training process for them. If they had done enough work to be sure we *could* run it.

(It was a bank. Software deployments have rules. An AI model-building app is still an app. It still goes through the same CI/CD pipeline as demand deposit account software changes. It's a batch job, really, just a bit more internally sophisticated than the thing that clears checks.)

Some Structure

I lean toward the following tiers of testing:

  1. Unit tests of every class and function. 100% code coverage here. I suggest using pytest and pytest-cov packages to tracking testing and make sure every line of code has some test case. For a few particularly tricky things, every logic path is better than simply testing lines of code. In some cases, every line of code will tend to touch every logic path, but seems less burdensome.
  2. Use hypothesis for the more sensitive numeric functions. In “data wrangling” applications there may not be too many of these. In the machine learning and model application software, there may be more sophisticated math that benefits from hypothesis testing.
  3. Write larger integration tests that mimic pyspark processing, using multiple functions or classes to be sure they work together correctly, but without the added complication of actually using pySpark. This means creating mocks for some of the libraries using unittest.mock objects. This is a fair bit of work, but it pays handsome dividends when debugging. For well-understood pyspark APIs, it should be easy to provide mocked results for the app components under test to use. For the less well-understood parts, the time spent building a mock will often provide useful insight into how (and why) it works the way it does. In rare cases, building the mock suggests a better design that's easier to test.
  4. Finally. Write a few overall acceptance tests that use your modules and also start and run a small pyspark instance from the command line. For this, I really like using behave, and writing the acceptance testing cases using the Gherkin language. This enforces a very formal “Given-When-Then” structure on the test scenarios, and allows you to write in English. You can share the Gherkin with users and other stakeholders to be sure they agree on what the application should do.


Each tier of testing builds up a larger, and more complete picture of the overall application. 

More important, we don't emphasize running pySpark and testing it. It already works. It has it's own tests. We need to test the stuff we wrote, not the framework.

We need to test our code in isolation.

We need to test integrated code with mocked pySpark.

Once we're sure our code is likely to work, the next step is confirmation that the important parts do work with pySpark. For life-critical applications, the integration tests will need to touch 100% of the logic paths. For data analytics, extensive integration testing is a lot of cost for relatively little benefit.

Even for data analytics, testing is a lot of work. The alternative is hope and prayer. I suggest starting with small unit tests, and expanding from there.

No comments:

Post a Comment

Note: Only a member of this blog may post a comment.