A blog about technical stuff

Python data pipelines similar to R's '%>%'

Oct 23, 2016

Since a few years, pipelines (via %>% of the magrittr package) are quite popular in R and the grown ecosystem of the “tidyverse” is built around pipelines. Having tried both the pandas syntax (e.g. chaining like df.groupby().mean() or plain function2(function1(input))) and the R’s pipeline syntax, I have to admit that I like the pipeline syntax a lot more.

In my opinion the strengths of R’s pipeline syntax are:

The same verbs can be used for different inputs (there are SQL backends for dplyr), thanks to R’s single-dispatch mechanism (called S3 objects).
Thanks to using function instead of class methods, it’s also more easily extendable (for a new method on pandas.DataFrame you have to add that to the pandas repository or you need to use monkey patching). Fortunatelly, both functions and singledispatch are also available in python :-)
It uses normal functions as pipline parts: input %>% function() is equivalent to function(input). Unfortunately, this isn’t easily matched in python, as pythons evaluation rules would first evaluate function() (e.g. call functions without any input). So one has to make function() return a helper object which can then be used as a pipeline part.
R’s delayed evaluation rules make it easy to evaluate arguments in the context of the pipeline, e.g. df %>% select(x) would be converted to the equivalent of pandas df[["x"]], e.g. the name of the variable will be used in the selection. In python it would either error (if x is not defined) or (if x was defined, e.g. x = "column"), would take the value of x, e.g. df[["column"]]. For this, some workarounds exist by using helper objects like select(X.x), e.g. pandas-ply and its Symbolic expression.

There exist a few implementation of dplyr like pipeline verbs for python (e.g. pandas itself, pandas-ply (uses method chaining instead of a pipe operator), dplython, and dfply), but they all focus on implementing dplyr style pipelines for pandas.DataFrames and I wanted to try out a simpler but more general approach to pipelines.

The code

The following shows my take on how to implement the first three things (I left out “Symbolic expressions”). The code is available in https://github.com/janschulz/pydatapipes. The short (removed the docstrings) version is actually only a few lines of code:

from functools import singledispatch, wraps

class PipeVerb():
    """Object which represents a part of a pipeline"""
    def __init__(self, func, *args, **kwargs):
        self.pipe_func = func
        self.args = args
        self.kwargs = kwargs

    def __rrshift__(self, input):
        return self.pipe_func(input, *self.args, **self.kwargs)


def pipeverb(func):
    """Decorator to convert a function to a pipeline verb (without singledispatch)"""
    @wraps(func)
    def decorated(*args, **kwargs):
        return PipeVerb(func, *args, **kwargs)
    
    # If it is a singledispatch method, expose the register method here as well
    if hasattr(func, 'register'):
        decorated.register = func.register

    return decorated


def make_pipesource(cls):
    """Enables a class to function as a pipe source"""
    if hasattr(cls, '__rshift__') and (not getattr(cls.__rshift__, 'pipeoperator', False)):
        def __rshift__(self, other):
            """Pipeline operator if the right side is a PipeVerb"""
            if isinstance(other, PipeVerb):
                return other.__rrshift__(self)
            else:
                return self.__orig_rshift__(other)

        cls.__orig_rshift__ = cls.__rshift__
        cls.__rshift__ = __rshift__
        setattr(cls.__rshift__, "pipeoperator", True)


def singledispatch_pipeverb(func):
    """Convenience decorator to convert a function to a singledispatch pipeline verb"""
    return pipeverb(singledispatch(func))

Simple pipeline verbs

For end users wanting to build a new pipeline verb or add pipeline functionality to a new data source, there are two functions to build new pipeline parts:

#from pydatapipes.pipes import singledispatch_pipeverb, make_pipesource
import pandas as pd

# generic version which defines the API and should raise NotImplementedError
@singledispatch_pipeverb
def append_col(input, x = 1):
    """Appends x to the data source"""
    raise NotImplementedError("append_col is not implemented for data of type %s" % type(input))

# concrete implementation for pandas.DataFrame
@append_col.register(pd.DataFrame)
def append_col_df(input, x = 1):
    # always ensure that you return new data!
    copy = input.copy()
    copy["X"] = x
    return copy

# ensure that pd.DataFrame is usable as a pipe source
make_pipesource(pd.DataFrame)

This can then be used in a pipeline:

import pandas as pd
print(pd.DataFrame({"a" : [1,2,3]}) >> append_col(x=3))

The above example implements a pipeline verb for pandas.DataFrame, but due to the useage of singledispatch, this is generic. By implementing additional append_col_<data_source_type>() functions and registering it with the original append_col function, the append_col function can be used with other data sources, e.g. SQL databases, HDF5, or even builtin data types like list or dict:

@append_col.register(list)
def append_col_df(input, x = 1):
    return input + [x]

[1, 2] >> append_col()

[1, 2, 1]

If a verb has no actual implementation for a data source, it will simply raise an NotImplementedError:

try:
    1 >> append_col()
except NotImplementedError as e:
    print(e)
    

append_col is not implemented for data of type <class 'int'>

A more complex example: grouped and ungrouped aggregation on DataFrames

singledispatch also makes it easy to work with grouped and ungrouped pd.DataFrames:

@singledispatch_pipeverb
def groupby(input, columns):
    """Group the input by columns"""
    raise NotImplementedError("groupby is not implemented for data of type %s" % type(input))

@groupby.register(pd.DataFrame)
def groupby_DataFrame(input, columns):
    """Group a DataFrame"""
    return input.groupby(columns)    
    
@singledispatch_pipeverb
def summarize_mean(input):
    """Summarize the input via mean aggregation"""
    raise NotImplementedError("summarize_mean is not implemented for data of type %s" % type(input))

@summarize_mean.register(pd.DataFrame)
def summarize_mean_DataFrame(input):
    """Summarize a DataFrame via mean aggregation"""
    return input.mean()

@summarize_mean.register(pd.core.groupby.GroupBy)
def summarize_mean_GroupBy(input):
    """Summarize a grouped DataFrame via mean aggregation"""
    return input.mean()

df = pd.DataFrame({"a" : [1, 2, 3, 4], "b": [1, 1, 2, 2]})

print(df >> summarize_mean())

a    2.5
b    1.5
dtype: float64

print(df >> groupby("b") >> summarize_mean())

Limitiations

Compared to R’s implementation in the magrittr package, input >> verb(x) can’t be rewritten as verb(input, x).

The problem here is that verb(x) under the hood constructs a helper object (PipeVerb) which is used in the rshift operation. At the time of calling verb(...), we can’t always be sure whether we want an object which can be used in the pipeline or want to already compute the result. As an example consider a verb merge(*additional_data). You could call that as data >> merge(first, second) to indicate that you want all three (data, first, and second) merged. On the other hand, merge(first, second) is also valid (“merge first and second together).

Usage as function and pipeline verb

To help work around this problem, the convenience decorator singledispatch_pipeverb is actually not the best option if you want to create reusable pipeline verbs. Instead, the singledispatch_pipeverb decorator is also available in two parts, so that one can both expose the original function (with singledispatch enabled) and the final pipeline verb version:

#from pydatapipes.pipes import pipeverb, singledispatch

# first use singledispatch on the original function, but define it with a trailing underscore
@singledispatch
def my_verb_(input, x=1, y=2):
    raise NotImplemented("my_verb is not implemented for data of type %s" % type(input))

# afterwards convert the original function to the pipeline verb:
my_verb = pipeverb(my_verb_)

# concrete implementations can be registered on both ``my_verb`` and ``my_verb_``
@my_verb_.register(list)
def my_verb_df(input, x=1, y=2):
    return input + [x, y]

A user can now use both versions:

[1] >> my_verb(x=2, y=3)

[1, 2, 3]

my_verb_([9], x=2, y=3)

[9, 2, 3]

Rules and conventions

To work as a pipline verb, functions must follow these rules:

Pipelines assume that the verbs itself are side-effect free, i.e. they do not change the inputs of the data pipeline. This means that actual implementations of a verb for a specific data source must ensure that the input is not changed in any way, e.g. if you want to pass on a changed value of a pd.DataFrame, make a copy first.
The initial function (not the actual implementations for a specific data source) should usually do nothing but simply raise NotImplementedError, as it is called for all other types of data sources.

The strength of the tidyverse is it’s coherent API design. To ensure a coherent API for pipeline verbs, it would be nice if verbs would follow these conventions:

Pipeline verbs should actually be named as verbs, e.g. use input >> summarize() instead of input >> Summary()
If you expose both the pipeline verb and a normal function (which can be called directly), the pipeline verb should get the “normal” verb name and the function version should get an underscore _ appended: x >> verb() -> verb_(x)
The actual implementation function of a verb() for a data source of class Type should be called verb_Type(...), e.g. select_DataFrame()

Missing parts

So what is missing? Quite a lot :-)

Symbolic expressions: e.g. select(X.x) instead of select("x")
Helper for dplyr style column selection (e.g. select(starts_with("y2016_")) and select(X[X.first_column:X.last_column]))
all the dplyr, tidyr, … verbs which make the tidyverse so great

Some of this is already implemented in the other dplyr like python libs (pandas-ply, dplython, and dfply), so I’m not sure how to go on. I really like my versions of pipelines but duplicating the works of them feels like a waste of time. So my next step is seeing if it’s possible to integrate this with one of these solutions, probably dfply as that looks the closest implementation.

[This post is also available as a jupyter notebook]