Skip to main content

Abstract data dispatch

Project description


The databackend package allows you to register a subclass, without needing to import the subclass itself. This is useful for implementing actions over optional dependencies.


For this example, we’ll implement a function, fill_na(), that fills in missing values in a DataFrame. It works with DataFrame objects from two popular libraries: pandas and polars. Importantly, neither library needs to be installed.


The code below defines “abstract” parent classes for each of the DataFrame classes in the two libraries.

from databackend import AbstractBackend

class AbstractPandasFrame(AbstractBackend):
    _backends = [("pandas", "DataFrame")]

class AbstractPolarsFrame(AbstractBackend):
    _backends = [("polars", "DataFrame")]

Note that the abstract classes can be used as stand-ins for the real thing in issubclass() and isinstance.

from pandas import DataFrame

issubclass(DataFrame, AbstractPandasFrame)
isinstance(DataFrame(), AbstractPandasFrame)

Simple fill_na: isinstance to switch behavior

The fill_na() function below uses custom handling for pandas and polars.

def fill_na(data, x):
    if isinstance(data, AbstractPolarsFrame):
        return data.fill_nan(x)
    elif isinstance(data, AbstractPandasFrame):
        return data.fillna(x)
        raise NotImplementedError()

Notice that neither pandas nor polars need to be imported when defining fill_na().

Here is an example of calling fill_na() on both kinds of DataFrames.

# test polars ----

import polars as pl

df = pl.DataFrame({"x": [1, 2, None]})
fill_na(df, 3)

# test pandas ----

import pandas as pd

df = pd.DataFrame({"x": [1, 2, None]})
fill_na(df, 3)
0  1.0
1  2.0
2  3.0

The key here is that a user could have only pandas, or only polars, installed. Importantly, doing the isinstance checks do not import any libraries!

Advanced fill_na: generic function dispatch

databackend shines when combined with generic function dispatch. This is a programming approach where you declare a function (e.g. fill_na()), and then register each backend specific implementation on the function.

Python has a built-in function implementing this called functools.singledispatch.

Here is an example of the previous fill_na() function written using it.

from functools import singledispatch

def fill_na2(data, x):
    raise NotImplementedError(f"No support for class: {type(data)}")

# handle polars ----

def _(data: AbstractPolarsFrame, x):
    return data.fill_nan(x)

# handle pandas ----

def _(data: AbstractPandasFrame, x):
    return data.fillna(x)

Note two important decorators:

  • @singledispatch defines a default function. This gets called if no specific implementations are found.
  • @fill_na2.register defines specific versions of the function.

Here’s an example of it in action.

# example ----

import pandas as pd
import polars as pl

df = pl.DataFrame({"x": [1, 2, None]})
fill_na2(df, 3)

df = pd.DataFrame({"x": [1, 2, None]})
fill_na2(df, 3)
0  1.0
1  2.0
2  3.0

How it works

Under the hood, AbstractBackend behaves similarly to python’s builtin abc.ABC class.

from abc import ABC

class MyABC(ABC):

from io import StringIO


# StringIO is a "virtual subclass" of MyABC
isinstance(StringIO("abc"), MyABC)

The key difference is that you can specify the virtual subclass using the tuple ("<mod_name>", "<class_name>").

When issubclass(SomeClass, AbstractBackend) runs, then…

  • The standard ABC caching mechanism is checked, and potentially returns the answer immediately.
  • Otherwise, a subclass hook cycles through registered backends.
  • The hook runs the subclass check for any backends that are imported (e.g. are in sys.modules).

Technically, AbstractBackend inherits all the useful metaclass things from abc.ABCMeta, so these can be used also.

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

databackend-0.0.1.tar.gz (10.1 kB view hashes)

Uploaded source

Built Distribution

databackend-0.0.1-py2.py3-none-any.whl (4.8 kB view hashes)

Uploaded py2 py3

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page