Django REST Pandas

Django REST Framework + pandas = A Model-driven Visualization API

Django REST Pandas (DRP) provides a simple way to generate and serve pandas DataFrames via the Django REST Framework. The resulting API can serve up CSV (and a number of other formats) for consumption by a client-side visualization tool like d3.js.

The design philosophy of DRP enforces a strict separation between data and presentation. This keeps the implementation simple, but also has the nice side effect of making it trivial to provide the source data for your visualizations. This capability can often be leveraged by sending users to the same URL that your visualization code uses internally to load the data.

DRP does not include any JavaScript code, leaving the implementation of interactive visualizations as an exercise for the implementer. That said, DRP is commonly used in conjunction with the wq.app library, which provides wq/chart.js and wq/pandas.js, a collection of chart functions and data loaders that work well with CSV served by DRP and wq.db's chart module.

Live Demo

The climata-viewer project uses Django REST Pandas and wq/chart.js to provide interactive visualizations and spreadsheet downloads.

Related Work

The field of Python-powered data analysis and visualization is growing, and there are a number of similar solutions that may fit your needs better.

• Django Pandas provides a custom ORM model manager with pandas support. By contrast, Django REST Pandas works at the view level, by integrating pandas via custom Django REST Framework serializers and renderers.
• DRF-CSV provides straightforward CSV renderers for use with Django REST Framework. It may be useful if you just want a CSV API and don't have a need for the pandas DataFrame functionality.
• mpld3 provides a direct bridge from matplotlib to d3.js, complete with seamless IPython integration. It is restricted to the (large) matplotlib chart vocabularly but should be sufficient for many use cases.
• Bokeh is a complete client-server visualization platform. It does not leverage d3 or Django, but is notable as a comprehensive, forward-looking approach to addressing similar use cases.

The goal of Django REST Pandas is to provide a generic REST API for serving up pandas dataframes. In this sense, it is similar to the Plot Server in Bokeh, but more generic in that it does not assume any particular visualization format or technology. Further, DRP is optimized for integration with public-facing Django-powered websites (unlike mpld3 which is primarily intended for use within IPython).

In summary, DRP is designed for use cases where:

• You want to support live spreadsheet downloads as well as interactive visualizations, and/or
• You want full control over the client visualization stack in order to integrate it with the rest of your website and/or build process. This usually means writing JavaScript code by hand. mpld3 may be a better choice for data exploration if you are more comfortable with (I)Python and need something that can generate interactive visualizations out of the box.

Supported Formats

The following output formats are provided by default. These are provided as renderer classes in order to leverage the content type negotiation built into Django REST Framework. This means clients can specify a format via:

• an HTTP "Accept" header (Accept: text/csv),
• a format parameter (/path?format=csv), or
• a format extension (/path.csv)

The HTTP header and format parameter are enabled by default on every pandas view. Using the extension requires a custom URL configuration (see below).

Format	Content Type	pandas DataFrame Function	Notes
CSV	text/csv	to_csv()
TXT	text/plain	to_csv()	Useful for testing, as most browsers will download a CSV file instead of displaying it
JSON	application/json	to_json()	date_format and orient can be provided in URL (e.g. /path.json?orient=columns)
XLSX	application/vnd.openxml...sheet	to_excel()
XLS	application/vnd.ms-excel	to_excel()
PNG	image/png	plot()	Currently not very customizable, but a simple way to view the data as an image.
SVG	image/svg	plot()	Eventually these could become a fallback for clients that can't handle d3.js

The underlying implementation is a set of serializers that take the normal serializer result and put it into a dataframe. Then, the included renderers generate the output using the built in pandas functionality.

Usage

Getting Started

pip3 install rest-pandas

NOTE: Django REST Pandas relies on pandas, which itself relies on NumPy and other scientific Python libraries. If you are having trouble installing DRP due to dependency issues, you may want to pre-install pandas using another tool. For example, on Ubuntu 14.04 LTS you can pre-install pandas using this command:

sudo apt-get install python3-pandas
sudo pip3 install rest-pandas

The pandas documentation recommends using conda to install pandas for similar reasons. We've found the apt-get approach to be the fastest route to getting DRP running with the default Apache WSGI implementation on Ubuntu.

Usage Example

No Model

The example below allows you to create a simple API for an existing Pandas DataFrame, e.g. generated from an existing file.

# views.py
from rest_pandas import PandasSimpleView
import pandas as pd

class TimeSeriesView(PandasSimpleView):
    def get_data(self, request, *args, **kwargs):
        return pd.read_csv('data.csv')

Model-Backed

The example below assumes you already have a Django project set up with a single TimeSeries model.

# views.py
from rest_pandas import PandasView
from .models import TimeSeries
from .serializers import TimeSeriesSerializer

# Short version (leverages default DRP settings):
class TimeSeriesView(PandasView):
    queryset = TimeSeries.objects.all()
    serializer_class = TimeSeriesSerializer
    # That's it!  The view will be able to export the model dataset to any of
    # the included formats listed above.  No further customization is needed to
    # leverage the defaults.

# Long Version and step-by-step explanation
class TimeSeriesView(PandasView):
    # Assign a default model queryset to the view
    queryset = TimeSeries.objects.all()

    # Step 1. In response to get(), the underlying Django REST Framework view
    # will load the queryset and then pass it to the following function.
    def filter_queryset(self, qs): 
        # At this point, you can filter queryset based on self.request or other
        # settings (useful for limiting memory usage).  This function can be
        # omitted if you are using a filter backend or do not need filtering.
        return qs
        
    # Step 2. A Django REST Framework serializer class should serialize each
    # row in the queryset into a simple dict format.  A simple ModelSerializer
    # should be sufficient for most cases.
    serializer_class = TimeSeriesSerializer  # extends ModelSerializer

    # Step 3.  The included PandasSerializer will load all of the row dicts
    # into array and convert the array into a pandas DataFrame.  The DataFrame
    # is essentially an intermediate format between Step 2 (dict) and Step 4
    # (output format).  The default DataFrame simply maps each model field to a
    # column heading, and will be sufficient in many cases.  If you do not need
    # to transform the dataframe, you can skip to step 4.
    
    # If you would like to transform the dataframe (e.g. to pivot or add
    # columns), you can do so in one of two ways:

    # A. Create a subclass of PandasSerializer, define a function called
    # transform_dataframe(self, dataframe) on the subclass, and assign it to
    # pandas_serializer_class on the view.  You can also use one of the three
    # provided pivoting serializers (see Advanced Usage below).
    #
    # class MyCustomPandasSerializer(PandasSerializer):
    #     def transform_dataframe(self, dataframe):
    #         dataframe.some_pivot_function(in_place=True)
    #         return dataframe
    #
    pandas_serializer_class = MyCustomPandasSerializer

    # B. Alternatively, you can create a custom transform_dataframe function
    # directly on the view.  Again, if no custom transformations are needed,
    # this function does not need to be defined.
    def transform_dataframe(self, dataframe):
        dataframe.some_pivot_function(in_place=True)
        return dataframe
    
    # NOTE: As the name implies, the primary purpose of transform_dataframe()
    # is to apply a transformation to an existing dataframe.  In PandasView,
    # this dataframe is created by serializing data queried from a Django
    # model.  If you would like to supply your own custom DataFrame from the
    # start (without using a Django model), you can do so with PandasSimpleView
    # as shown in the first example.

    # Step 4. Finally, the provided renderer classes will convert the DataFrame
    # to any of the supported output formats (see above).  By default, all of
    # the formats above are enabled.  To restrict output to only the formats
    # you are interested in, you can define renderer_classes on the view:
    renderer_classes = [PandasCSVRenderer, PandasExcelRenderer]
    # You can also set the default renderers for all of your pandas views by
    # defining the PANDAS_RENDERERS in your settings.py.

Registering URLs

# urls.py
from django.conf.urls import patterns, include, url

from .views import TimeSeriesView
urlpatterns = patterns('',
    url(r'^data', TimeSeriesView.as_view()),
)

# This is only required to support extension-style formats (e.g. /data.csv)
from rest_framework.urlpatterns import format_suffix_patterns
urlpatterns = format_suffix_patterns(urlpatterns)

The default PandasView will serve up all of the available data from the provided model in a simple tabular form. You can also use a PandasViewSet if you are using Django REST Framework's ViewSets and Routers.

Building Interactive Charts

In addition to use as a data export tool, DRP is well-suited for creating data API backends for interactive charts. In particular, DRP can be used with d3.js, wq/pandas.js, and wq/chart.js, to create interactive time series, scatter, and box plot charts - as well as any of the infinite other charting possibilities d3.js provides.

To facilitate data API building, the CSV renderer is the default in Django REST Pandas. While the pandas JSON serializer is improving, the primary reason for making CSV the default is the compactness it provides over JSON when serializing time series data. The default CSV output from DRP will have single row of column headers, making it suitable as-is for use with e.g. d3.csv(). However, DRP is often used with the custom serializers below to produce a dataframe with nested multi-row column headers. This is harder to parse with d3.csv() but can be easily processed by wq/pandas.js, an extension to d3.js.

// mychart.js
define(['d3', 'wq/pandas', 'wq/chart'], function(d3, pandas, chart) {

// Unpivoted data (single-row header)
d3.csv("/data.csv", render);

// Pivoted data (multi-row header)
pandas.get('/data.csv', render);

function render(error, data) {
    d3.select('svg')
       .selectAll('rect')
       .data(data)
       // ...
}

});

You can override the default renderers by setting PANDAS_RENDERERS in your settings.py, or by overriding renderer_classes in your PandasView subclass. PANDAS_RENDERERS is intentionally set separately from Django REST Framework's own DEFAULT_RENDERER_CLASSES setting, as it is likely that you will be mixing DRP views with regular DRF views.

As of version 0.4, DRP includes three custom serializers with transform_dataframe() functions that address common use cases. These serializer classes can be leveraged by assigning them to pandas_serializer_class on your view.

For documentation purposes, the examples below assume the following dataset:

Location	Measurement	Date	Value
site1	temperature	2016-01-01	3
site1	humidity	2016-01-01	30
site2	temperature	2016-01-01	4
site2	temperature	2016-01-02	5

PandasUnstackedSerializer

PandasUnstackedSerializer unstacks the dataframe so a few key attributes are listed in a multi-row column header. This makes it easier to include metadata about e.g. a time series without repeating the same values on every data row.

To specify which attributes to use in column headers, define the attribute pandas_unstacked_header on your ModelSerializer subclass. You will generally also want to define pandas_index, which is a list of metadata fields unique to each row (e.g. the timestamp).

# serializers.py
from rest_framework import serializers
from .models import TimeSeries

class TimeSeriesSerializer(serializers.ModelSerializer):
    class Meta:
        model = MultiTimeSeries
        pandas_index = ['date']
        pandas_unstacked_header = ['location', 'measurement']

# views.py
from rest_pandas import PandasView, PandasUnstackedSerializer
from .models import TimeSeries
from .serializers import TimeSeriesSerializer

class TimeSeriesView(PandasView):
    queryset = TimeSeries.objects.all()
    serializer_class = TimeSeriesSerializer
    pandas_serializer_class = PandasUnstackedSerializer

With the above example data, this configuration would output a CSV file with the following layout:

	Value	Value	Value
Location	site1	site1	site2
Measurement	temperature	humidity	temperature
Date
2016-01-01	3	30	4
2016-01-02			5

This could then be processed by wq/pandas.js into the following structure:

[
    {
        "location": "site1",
        "measurement": "temperature",
        "data": [
            {"date": "2016-01-01", "value": 3}
        ]
    },
    {
        "location": "site1",
        "measurement": "humidity",
        "data": [
            {"date": "2016-01-01", "value": 30}
        ]
    },
    {
        "location": "site2",
        "measurement": "temperature",
        "data": [
            {"date": "2016-01-01", "value": 4},
            {"date": "2016-01-02", "value": 5}
        ]
    }
]

The output of PandasUnstackedSerializer can be used with the timeSeries() chart provided by wq/chart.js:

define(['d3', 'wq/pandas', 'wq/chart'], function(d3, pandas, chart) {

var svg = d3.select('svg');
var plot = chart.timeSeries();
pandas.get('/data/timeseries.csv', function(data) {
    svg.datum(data).call(plot);
});

});

PandasScatterSerializer

PandasScatterSerializer unstacks the dataframe and also combines selected attributes to make it easier to plot two measurements against each other in an x-y scatterplot.

To specify which attributes to use for the coordinate names, define the attribute pandas_scatter_coord on your ModelSerializer subclass. You can also specify additional metadata attributes to include in the header with pandas_scatter_header. You will generally also want to define pandas_index, which is a list of metadata fields unique to each row (e.g. the timestamp).

# serializers.py
from rest_framework import serializers
from .models import TimeSeries

class TimeSeriesSerializer(serializers.ModelSerializer):
    class Meta:
        model = MultiTimeSeries
        pandas_index = ['date']
        pandas_scatter_coord = ['measurement']
        pandas_scatter_header = ['location']

# views.py
from rest_pandas import PandasView, PandasScatterSerializer
from .models import TimeSeries
from .serializers import TimeSeriesSerializer

class TimeSeriesView(PandasView):
    queryset = TimeSeries.objects.all()
    serializer_class = TimeSeriesSerializer
    pandas_serializer_class = PandasScatterSerializer

With the above example data, this configuration would output a CSV file with the following layout:

	temperature-value	humidity-value	temperature-value
Location	site1	site1	site2
Date
2014-01-01	3	30	4
2014-01-02			5

This could then be processed by wq/pandas.js into the following structure:

[
    {
        "location": "site1",
        "data": [
            {
                "date": "2016-01-01",
                "temperature-value": 3,
                "humidity-value": 30
            }
        ]
    },
    {
        "location": "site2",
        "data": [
            {
                "date": "2016-01-01",
                "temperature-value": 4
            },
            {
                "date": "2016-01-02",
                "temperature-value": 5
            }
        ]
    }
]

The output of PandasScatterSerializer can be used with the scatter() chart provided by wq/chart.js:

define(['d3', 'wq/pandas', 'wq/chart'], function(d3, pandas, chart) {

var svg = d3.select('svg');
var plot = chart.scatter()
    .xvalue(function(d) {
        return d['temperature-value'];
    })
    .yvalue(function(d) {
        return d['humidity-value'];
    });

pandas.get('/data/scatter.csv', function(data) {
    svg.datum(data).call(plot);
});

});

PandasBoxplotSerializer

PandasBoxplotSerializer computes boxplot statistics (via matplotlib's boxplot_stats) and pushes the results out via an unstacked dataframe. The statistics can be aggregated for a specified group column as well as by date.

To specify which attribute to use for the group column, define the attribute pandas_boxplot_group on your ModelSerializer subclass. To specify an attribute to use for date-based grouping, define pandas_boxplot_date. You will generally also want to define pandas_boxplot_header, which will unstack any metadata columns and exclude them from statistics.

# serializers.py
from rest_framework import serializers
from .models import TimeSeries

class TimeSeriesSerializer(serializers.ModelSerializer):
    class Meta:
        model = MultiTimeSeries
        pandas_boxplot_group = 'site'
        pandas_boxplot_date = 'date'
        pandas_boxplot_header = ['measurement']

# views.py
from rest_pandas import PandasView, PandasBoxplotSerializer
from .models import TimeSeries
from .serializers import TimeSeriesSerializer

class TimeSeriesView(PandasView):
    queryset = TimeSeries.objects.all()
    serializer_class = TimeSeriesSerializer
    pandas_serializer_class = PandasBoxplotSerializer

With the above example data, this configuration will output a CSV file with the same general structure as PandasUnstackedSerializer, but with the value spread across multiple boxplot statistics columns (value-mean, value-q1,value-whishi, etc.). An optional group` parameter can be added to the query string to switch between various groupings:

name	purpose
?group=series	Group by series (pandas_boxplot_group)
?group=series-year	Group by series, then by year
?group=series-month	Group by series, then by month
?group=year	Summarize all data by year
?group=month	Summarize all data by month

The output of PandasBoxplotSerializer can be used with the boxplot() chart provided by wq/chart.js:

define(['d3', 'wq/pandas', 'wq/chart'], function(d3, pandas, chart) {

var svg = d3.select('svg');
var plot = chart.boxplot();
pandas.get('/data/boxplot.csv?group=year', function(data) {
    svg.datum(data).call(plot);
});

});