Data transformers¶

Before a Vega-Lite or Vega specification can be passed to a renderer, it typically has to be transformed in a number of ways:

Pandas Dataframe has to be sanitized and serialized to JSON.
The rows of a Dataframe might need to be sampled or limited to a maximum number.
The Dataframe might be written to a .csv of .json file for performance reasons.

These data transformations are managed by the data transformation API of Altair.

Note

The data transformation API of Altair should not be confused with the transform API of Vega and Vega-Lite.

A data transformer is a Python function that takes a Vega-Lite data dict or Pandas DataFrame and returns a transformed version of either of these types:

from typing import Union
Data = Union[dict, pd.DataFrame]

def data_transformer(data: Data) -> Data:
    # Transform and return the data
    return transformed_data

Dataset Consolidation¶

Datasets passed as Pandas dataframes can be represented in the chart in two ways:

As literal dataset values in the data attribute at any level of the specification
As a named dataset in the datasets attribute of the top-level specification.

The former is a bit more simple, but common patterns of usage in Altair can often lead to full datasets being listed multiple times in their entirety within a single specification.

For this reason, Altair 2.2 and newer will by default move all directly-specified datasets into the top-level datasets entry, and reference them by a unique name determined from the hash of the data representation. The benefit of using a hash-based name is that even if the user specifies a dataset in multiple places when building the chart, the specification will only include one copy.

This behavior can be modified by setting the consolidate_datasets attribute of the data transformer.

For example, consider this simple layered chart:

import altair as alt
import pandas as pd

df = pd.DataFrame({'x': range(5),
                   'y': [1, 3, 4, 3, 5]})

line = alt.Chart(df).mark_line().encode(x='x', y='y')
points = alt.Chart(df).mark_point().encode(x='x', y='y')
chart = line + points

If we look at the resulting specification, we see that although the dataset was specified twice, only one copy of it is output in the spec:

from pprint import pprint
pprint(chart.to_dict())

{'$schema': 'https://vega.github.io/schema/vega-lite/v4.8.1.json',
 'config': {'view': {'continuousHeight': 300, 'continuousWidth': 400}},
 'data': {'name': 'data-137453ccbb3f929636eb2f9dd7052f16'},
 'datasets': {'data-137453ccbb3f929636eb2f9dd7052f16': [{'x': 0, 'y': 1},
                                                        {'x': 1, 'y': 3},
                                                        {'x': 2, 'y': 4},
                                                        {'x': 3, 'y': 3},
                                                        {'x': 4, 'y': 5}]},
 'layer': [{'encoding': {'x': {'field': 'x', 'type': 'quantitative'},
                         'y': {'field': 'y', 'type': 'quantitative'}},
            'mark': 'line'},
           {'encoding': {'x': {'field': 'x', 'type': 'quantitative'},
                         'y': {'field': 'y', 'type': 'quantitative'}},
            'mark': 'point'}]}

This consolidation of datasets is an extra bit of processing that is turned on by default in all renderers.

If you would like to disable this dataset consolidation for any reason, you can do so by setting alt.data_transformers.consolidate_datasets = False, or by using the enable() context manager to do it only temporarily:

with alt.data_transformers.enable(consolidate_datasets=False):
    pprint(chart.to_dict())

{'$schema': 'https://vega.github.io/schema/vega-lite/v4.8.1.json',
 'config': {'view': {'continuousHeight': 300, 'continuousWidth': 400}},
 'data': {'values': [{'x': 0, 'y': 1},
                     {'x': 1, 'y': 3},
                     {'x': 2, 'y': 4},
                     {'x': 3, 'y': 3},
                     {'x': 4, 'y': 5}]},
 'layer': [{'encoding': {'x': {'field': 'x', 'type': 'quantitative'},
                         'y': {'field': 'y', 'type': 'quantitative'}},
            'mark': 'line'},
           {'encoding': {'x': {'field': 'x', 'type': 'quantitative'},
                         'y': {'field': 'y', 'type': 'quantitative'}},
            'mark': 'point'}]}

Notice that now the dataset is not specified within the top-level datasets attribute, but rather as values within the data attribute of each individual layer. This duplication of data is the reason that dataset consolidation is set to True by default.

Built-in data transformers¶

Altair includes a default set of data transformers with the following signatures.

Raise a MaxRowsError if a Dataframe has more than max_rows rows:

limit_rows(data, max_rows=5000)

Randomly sample a DataFrame (without replacement) before visualizing:

sample(data, n=None, frac=None)

Convert a Dataframe to a separate .json file before visualization:

to_json(data, prefix='altair-data'):

Convert a Dataframe to a separate .csv file before visualization:

to_csv(data, prefix='altair-data'):

Convert a Dataframe to inline JSON values before visualization:

to_values(data):

Piping¶

Multiple data transformers can be piped together using pipe:

from altair import limit_rows, to_values
from toolz.curried import pipe
pipe(data, limit_rows(10000), to_values)

Managing data transformers¶

Altair maintains a registry of data transformers, which includes a default data transformer that is automatically applied to all Dataframes before rendering.

To see the registered transformers:

>>> import altair as alt
>>> alt.data_transformers.names()
['default', 'json', 'csv']

The default data transformer is the following:

def default_data_transformer(data):
    return pipe(data, limit_rows, to_values)

The json and csv data transformers will save a Dataframe to a temporary .json or .csv file before rendering. There are a number of performance advantages to these two data transformers:

The full dataset will not be saved in the notebook document.
The performance of the Vega-Lite/Vega JavaScript appears to be better for standalone JSON/CSV files than for inline values.

There are disadvantages of the JSON/CSV data transformers:

The Dataframe will be exported to a temporary .json or .csv file that sits next to the notebook.
That notebook will not be able to re-render the visualization without that temporary file (or re-running the cell).

In our experience, the performance improvement is significant enough that we recommend using the json data transformer for any large datasets:

alt.data_transformers.enable('json')

We hope that others will write additional data transformers - imagine a transformer which saves the dataset to a JSON file on S3, which could be registered and enabled as:

alt.data_transformers.register('s3', lambda data: pipe(data, to_s3('mybucket')))
alt.data_transformers.enable('s3')

Storing JSON data in a separate directory¶

When creating many charts with alt.data_transformers.enable('json') the working directory can get a bit cluttered. To avoid this we can build a simple custom data transformer that stores all JSON files in separate directory.:

import os
import altair as alt
from toolz.curried import pipe


def json_dir(data, data_dir='altairdata'):
    os.makedirs(data_dir, exist_ok=True)
    return pipe(data, alt.to_json(filename=data_dir + '/{prefix}-{hash}.{extension}') )


alt.data_transformers.register('json_dir', json_dir)
alt.data_transformers.enable('json_dir', data_dir='mydata')

After enabling this data transformer, the JSON files will be stored in what data_dir was set to when enabling the transformer or ‘altairdata’ by default. All we had to do was to prefix the filename argument of the alt.to_json function with our desired directory and make sure that the directory actually exists.