Extending Pandas

Dr. Bryan Patrick Wood

December 21, 2021

Filed under “

Open for extension, but closed for modification is a solid1 recommendation indeed. I love software that supports extension. I've had a particular fascination with extension mechanisms in software throughout my career. I've also been lucky enough to work on several including

  • Writing plugins for a homegrown C++ data orchestration framework while at Raytheon Solipsys This is before there were a lot of quality open-source data orchestration solutions available (e.g., Airflow, Nifi, etc.).
  • Writing my own homegrown data orchestration framework from scratch in Java while at Focused Support
  • This was pure joy: marrying my love for C++, Boost, and Python! Writing a framework to support Python scriptable components in a C++ application while at Stellar Science
  • Writing a report generation framework in Python supporting pluggable section content using the pluggy More on this in the future.and reportlab Python packages at my current position with Leidos

A predilection, mayhaps. Like a moth to the flame, when I discovered one of my most frequently used Python packages supported extension mechanisms I was compelled to look into it.

Who is this for?

Although practicality beats purity2. Maybe you're not a software architecture wonk aficionado. That's fair: I'd speculate that's the exception rather than the rule for folks using pandas3. This is not a deep dive into how the extension mechanisms are implemented. Interesting in their own right but out of the intended scope.

The official documentation for extending pandas4 mentions four ways in which pandas can be extended

  1. Custom accessors
  2. Custom data types
  3. Subclassing existing pandas data structures
  4. Third-party plotting backends

The focus here will be almost exclusively on #1: custom accessors. This is the one case I had a use for and think most pandas users could benefit from knowing about.

What are some reasons to continuing reading? Maybe you've noticed a few verbose pandas statements that seem to recur often in your notebooks or scripts. Or maybe you're a teammate noticing the same mistakes or deviations from best practices during code review. Or maybe you're just a bit perverse and want to bend pandas to your will. If this is you good reader, read on: what follows is practical information that can be applied to these ends.

What is an Accessor?

If you've used pandas for any amount of time you will already be familiar with them. pandas ships with a few builtin accessors on the Series and DataFrame classes

  • Series.cat, Series.dt, Series.plot, Series.sparse, and Series.str
  • DataFrame.plot and DataFrame.sparse

I use Series.str5 all the time for manipulating string type columns in DataFrames. It provides vectorized string methods mirroring the Python standard library ones intelligently skipping over NAs. Series.plot and DataFrame.plot are another two that I use frequently whereas I have never had the occasion to use Series.sparse or DataFrame.sparse. The type of data you work with most often will dictate which will be most familiar so YMMV. Your Mileage May Vary (YMMV)

Example usage is straight forward

print(pd.Series(['bar', 'baz']).str.upper().str.replace('B', 'Z'))
0    zAR
1    zAZ
dtype: object


Pandas supports registering custom accessors for its Index, Series, and DataFrame classes via decorators in the pandas.api.extensions module. These decorators expect a class which is passed the respective object as the only non-self argument to its __init__ method.

Here is the src directory structure Pandas Accessors Boilerplate Directory Structure The minimal boilerplate necessary for creating a pandas accessor for each supported class is shown in the figure below. If you only intend to write a DataFrame accessor the series.py and index.py modules can be removed. The package naming and the accessor are both personalized with my initials. Since this effectively creates a namespace within the respectively classes it's first come, first served: sorry Bernard Piotr Wojciechowski6.

Pandas Accessors Boilerplate

Viz., the pandas documentation for a geo DataFrame accessor that requires lat and lon columns.4 Including _validate methods is generally a good idea especially for special purpose accessors that only make sense to use on objects that meet enforceable preconditions. Since this is meant to be more of a grab bag of personal conveniences, validation will get pushed down to the individual accessor methods as appropriate obviating the need for top level validation.

With that in place we can verify everything is working

Pandas Extension MVP

The custom accessors do not have any useful methods or properties yet; however, at this point it's off to the races.

What to Add?

That's completely up to you: you now have your own pandas extension module. Pretty sweet! A good place to start is looking through your projects, scripts, notebooks, etc., that use pandas for patterns and repeated statements. Easy targets are

  1. Any functions that show up in frequently in pandas HoF calls (e.g., apply, map, applymap, aggregate, transform, etc.)Higher-order Function (HoF).
  2. pandas setup code can be pushed into the module __init__.py
  3. Proxies and aliases

Still stumped? I'll share some snippets for what I added to mine after creating the boilerplate to assist in your ideation process. I've elided details in the interest of brevity. For those details, the full code can be found on GitHub and the package can be installed from PyPI.

Subject Matter Expert (SME). In my current position, the folks I'm working with are more likely to be SMEs in an area outside my own vice data scientists. As a result, I'm passed all sorts of unhygienic data and have amassed a variety of cleaning routines. Adding a common.py to collect these functions decoupled from their embedding in the accessors

PUNCTUATION_TO_SPACE_TRANSLATOR = str.maketrans(string.punctuation, ' ' * len(string.punctuation))

def clean_name(name: str):
    return '_'.join(name.lower().translate(PUNCTUATION_TO_SPACE_TRANSLATOR).split())

def unicode_normalize(text: str, form: str = DEFAULT_UNICODE_NORMALIZATION_FORM):
    return unicodedata.normalize(form, text)

Referencing it in series.py

class SeriesAccessor:
    def clean_name(self):
        series = self._series.copy()
        series.name = clean_name(series.name)
        return series

    def normalize(self, form: str = DEFAULT_UNICODE_NORMALIZATION_FORM):
        series = self._series.copy()
        return series.apply(functools.partial(unicode_normalize, form=form))

And in dataframe.py

class DataframeAccessor:
    def clean_names(self):
        df = self._df.copy()
        df.columns = [clean_name(column) for column in df.columns]
        return df

We get our first set of useful custom accessor capabilities. clean_name and clean_names While researching I found pyjanitor7 which has a more robust name cleaning capability8. A ton of other useful stuff there: strongly encourage taking a look. should be pretty self-explanatory: lowercase, replace all special characters with space, tokenize, and concatenate the tokens with underscores; normalize does Unicode normalization9.

Doing an expensive operation on a big DataFrame or Series? You're going to want an estimate of when it's likely to An example of #2 above. There are certainly other (better?) ways to do this in interactive environments like ipython and jupyter startup scripts. be done. I've settled on tqdm10 for monitoring task progress in Python. It also conveniently has builtin support for pandas. However, this support needs to be initialized which can be easily dropped into the package __init__.py as follows

from tqdm.autonotebook import tqdm

sidetable, itself a pandas custom accessor, provides a few DataFrame summarizations: specifically, I use its freq method ... frequently. missingno provides visualizations for understanding the missing information in a DataFrame. Two packages I use frequently when analyzing data with pandas are the sidetable11 and missingno12 packages. I usually want them around when I am using pandas. Baking them into the custom accessor is an easy way to do that as well as provide some usage convenience.

Here's the relevant code

class _MissignoAdapter:
    def __init__(self, df: DataFrame):
        self._df = df

    def matrix(self, **kwargs):
        return missingno.matrix(self._df, **kwargs)

    def bar(self, **kwargs):
        return missingno.bar(self._df, **kwargs)

    def heatmap(self, **kwargs):
        return missingno.heatmap(self._df, **kwargs)

    def dendrogram(self, **kwargs):
        return missingno.dendrogram(self._df, **kwargs)

class DataframeAccessor:
    def stb(self):
        return self._df.stb

    def freq(self, *args, **kwargs):
        return self._df.stb.freq(*args, **kwargs)

    def msno(self):
        return _MissignoAdapter(self._df)

In the case of sidetable, this makes sure it is available after an import bpw_pde, provides an alias in the custom accessor to it, and pulls out the freq method to the top-level of the custom accessor. In the case of missingo, its interface needs adapting to work as a top-level msno property All problems in computer science can be solved by another level of indirection, except for the problem of too many layers of indirection.13 on the custom accessor which is easily handled by introducing a level of indirection through the adapter class.


The package __init__.py imports all the individual accessor classes so registration of the custom accessors happens on package import, i.e., import bpw_pde. Following that you can use the custom accessor just like the builtin ones, e.g.,

import pandas as pd
import matplotlib.pyplot as plt

import bpw_pde

df = pd.DataFrame(...)
df = df.bpw.clean_names
df.text = df.text.bpw.normalize()

For additional usage, the package has tests which is always one of the first places to look when trying to figure out how exactly to use some code.

What Else?

Plenty as you might imagine. As mentioned above there are four different ways to extend pandas. I've only touched on one in any detail here. A few final thoughts across all extension mechanisms follow.

I'd strongly recommend checking out the pandas official documentation on the pandas ecosystem14. Most of the examples mentioned in the rest of this section can be found there as well as a treasure trove of unmentioned pandas goodies.

Custom accessors

Hopefully a well beaten horse at this point. I will site another blog post on this topic I came across and found interesting on this topic for the unsated reader.

I've already mentioned sidetable as a custom pandas accessor I use, and I've also created one, bpw_pde, you can install and use in your next project if you find it useful. pandas_path3 is another I'd recommend taking a look at as a huge proponent and user of Python's bultin pathlib functionality.

It would also be remiss not to mention an extension to pandas extension mechanism: Inception Spinner pandas_flavor15 16. The primary additional functionality you get with pandas_flavor is the ability to register methods directly onto the underlying pandas type without an intermediate accessor. Bad idea? I think it's probably safer to namespace your extensions vice effectively monkeypatching pandas core data structures; however, I can also see the counterargument in exploratory data analysis environments. Context is everything.

pyjanitor, mentioned above, uses pandas_flavor's direct method registration vice the builtin pandas custom Application Programming Interface (API). accessor extension API.

Custom data types

This is absolutely another extension mechanism I can see having a need for in the future that I did not know about before diving into this topic for the package and blog post. However, I didn't want to create a contrived one just for the blog post because there are other good, useful open-source Python packages that do this. The documentation e.g., cyberpandas17 for custom data types related to IP Addresses. on this topic is pretty good and the ecosystem page has examples. I also recommend watching this dated but still relevant talk by core developer Will Ayd on the topic.

Subclassing existing pandas data structures

This is a much more time-consuming endeavor you'll need a good reason to pursue. To some degree, with custom accessors, this may be rarely needed in new code relegated mostly to the trash heap of legacy. Not a ton of good reasons come to But, again, I can't think of a good reason why you'd need to do that either. mind but if you ever needed to override a pandas data structure method that might be one. A good example here that I've used extensively is geopandas18.

Third-party plotting backends

Very cool that this is supported but also incredibly niche. I can't imagine even most advanced users and developers of pandas and related packages will ever need to use this. However, they may well benefit from it when a new plotting package developer can take advantage of it.