I have test-driven the pandera package over the past few days, and I'm extremely happy that this package exists!

Let me tell you why.

Data testing

"Data testing" is the colloquial term that I have used in place of the more formal term, "schema validation", or "data validation". Ever since I was burned once by CSV files with invalid column names, I put together a data testing tutorial to kickstart the idea, and have been keeping an eye on the space ever since.

But let’s step back a bit.

What do we mean by schema/data validation? It’s merely a fancy way of checking, “are my data as I expect them to be?”

Here’s an example. Imagine you have a data table, in which the semantic meaning of a row is a "sample" (i.e. one row = one sample), and columns are observations about that sample. In this data, there are definitely some things you may want to guarantee to be true of our data.

For example, I may have numeric measurements of each sample, and I need to guarantee that they cannot be negative or zero. Hence, for those columns, I need “greater than zero” check. For another column, it may be that we expect that the values in there take one of two values. An example dataframe could be as follows:

df = pd.DataFrame(
        "non_negative": [1.1, 0.34, 2.5],
        "two_values": ["val1", "val2", "val2"],

If I were implementing these checks in a test suite in vanilla Python, I would write them as follows:

from project_source.data import load_data
import numpy as np

def test_load_data():
    data = load_data(**kwargs)
    assert np.all(data["non_negative"] > 0)
    assert set(data["two_values"].unique()) == set(["val1",

Having the checks in the test suite is nice, as it means if the checks are run as part of a CI pipeline, they’ll continually be checked. However, they cannot be checked whenever I call on load_data(), which is the most important time I’d need those checks to appear — whenever I’m going to use that data loading function, i.e. at runtime, and have to depend on them being correct. To have those checks available at runtime, I’d have to duplicate all of the checks into the load_data() function. And we know that duplication of code increases risk of stale code.

Enter pandera. Instead of writing the checks in two separate places, I can start by writing the following declarative schema in a Python module, that serves as a defined source of truth:

# project_source/schemas.py
from pandera import Column,
import pandera as pa 

schema = DataFrameSchema(
        "non_zero_column": Column(pa.Float, Check.greater_than(0)),
        "categorical_column": Column(pa.String, Check.isin(["value_1", "value_2"])),

Now, the flexibility with which this schema gets used is simply superb! Because I've declared it as a Python object in code, I can import it anywhere I want validation to happen.

For example, I can validate the inputs to a function at runtime:

# project_source/data.py
from pandera import check_input,
from .schemas import dataframe_schema,

def clean(dataframe):
    # stuff happens
    return cleaned_df

And because the validation happens as soon as the function is executed, I can also validate the schema as part of an automatically-executed test suite that continuously runs and checks data alongside code.

from project_src.data import load_some_data

def test_load_some_data():
    data = load_some_data()

The data test suite becomes simpler to maintain, as it gets reduced to a simple execution test. Validating data at every step that it could go wrong is pretty important!

Additionally, if there is a new data file that needs to be processed in the same way, when we pass in the dataframe version into the clean_data function, it is automatically validated by pandera’s check_input. If a dataframe that fails check_input, then pandera will give us an error message straight away!

These are just the basics. Pandera comes with even more:

  • Validating the Index/MultiIndex.
  • Automatically inferring a draft schema that you can use as a base.

I'm excited to see how this goes. Congratulations to the package author Niels Bantilan, it's a package that has all of the right API ideas in place! Its design is very lightweight, which helps encourage fast iteration and turnaround on schema validation. Being one of those “eagerly executable” and “interactively debuggable” tools helps its case as well. The package focuses on doing only one thing well, and doesn't try to be a "God package" that does all sorts of things in one package. For a person of my tastes, this increases the attraction factor! Finally, it targets pandas DataFrames, an idiomatic data structure in the Python data science world, which is always a plus. Having incorporated pandera into a project just recently, I'm excited to see where its future is headed!

Data Validation as Statistical Evaluation

I can see that data validation actually has its roots in statistical evaluation of data.

Humour me for a moment.

Say we have a column that can only take two values, such as “yes” and “no”.

From a statistical viewpoint, that data are generated from Bernoulli distribution, with “yes”/1 and “no”/0 being the outcomes.

If you get a “maybe” in that column, that value is outside of the support of the distribution. It’s like trying to evaluate the likelihood of observing a 2 from a Bernoulli. It won’t compute.

So every time we run a pandera check, we are effectively expressing a statistical check of some kind. The byline of the package, “Statistical Data Validation for Pandas”, is even more apt once we consider this viewpoint!


I hope this post encourages you to give pandera a test-drive! Its implementation of "runtime data validation" is very handy, allowing me to declare the assumptions I have about my data up-front, and catch any violations of those assumptions as early as possible.