Intake: Parsing Data from Filenames and Paths



Do you have data in collections of files, where information is encoded both in the contents and the file/directory names? Perhaps something like 


? This is a very common problem for which people build custom code all the time. Intake provides a systematic way to declare that information in a concise spec.

What is Intake?

Intake is a lightweight set of tools for loading and sharing data. You might have seen earlier blog posts introducing Intakeand describing caching. Intake separates the concepts of the data engineer – the person curating, managing, and disseminating data – from the data user – the person analyzing and visualizing the data. The data engineer sets up catalog files describing data sources and the data user loads data without needing to know how it is stored. Intake makes creating functionality easy. Here we show new functionality for dealing with structured file-names.

How to use it – data user

Intake abstracts away messy data storage practices so data users don’t need to know about it. You get a full dataset from all the files, together with the information from the filenames, in just two lines of code:

cat = intake.load_catalog('catalog.yml')
data = cat.data_source().read()

As an example we’ll use a catalog with real satellite imagery data from the landsat project to calculate Normalized Difference Vegetation Index (NDVI):

Click the button below to launch an interactive session of the example or download the notebook here.


How to use it – data engineer

When loading multiple files, the locations of the files can often be provided as a list or as a glob (a path containing 


wildcards). For instance, let’s suppose that we have a number of CSV files containing precipitation forecasts for a number of models run under different emissions scenarios. Each forecast is stored in one file with the model and emissions scenario encoded in the filename. The following glob pattern would match all of the files:

urlpath: 'data/SRLCC_*_Precip_*.csv'

In order to capture the data encoded into the names of the files, we can replace the


wildcards with field-names, as follows, making what we’ll refer to as a path pattern:

urlpath: 'data/SRLCC_{emissions}_Precip_{model}.csv'

When the data source is opened, the values for each declared field are populated from the path or filename and returned on the data.

When passing an explicit list of paths, the argument path_as_pattern can be used to pass the pattern we want applied to the filenames:

  - 'data/SRLCC_a1b_Precip_ECHAM5-MPI.csv'
  - 'data/SRLCC_b1_Precip_PCM-NCAR.csv'
path_as_pattern: 'SRLCC_{emissions}_Precip_{model}.csv'

In this case the pattern is used to populate new columns (




) on the data with the values for each set of data being populated from the paths. Note that the pattern can just be a piece of the path as long as it is unambiguous where the piece starts and stops. For instance,


would not yield the intended outcome since emissions would match everything before




) and model would match everything after (



Click the button below to launch an interactive example of setting up a catalog with path_as_pattern or download the notebook here.


How it works

The formatting of the pattern is python format string syntax , but inverse: the set of arguments required such that

pattern.format(**arguments) == path


The logic is implemented in the function 


, which we can use to demonstrate how this works. We’ll call this function directly to demonstrate that any (reasonable) format string syntax is supported and the parsed values will match the implied type of the format string:

>>> reverse_format('data_{year}_{month}_{day}.csv', 'data_2014_01_03.csv')
{'year': '2014', 'month': '01', 'day': '03'}
>>> reverse_format('data_{year:d}_{month:d}_{day:d}.csv', 'data_2014_01_03.csv')
{'year': 2014, 'month': 1, 'day': 3}
>>> reverse_format('data_{date:%Y_%m_%d}.csv', 'data_2016_10_01.csv')
{'date': datetime.datetime(2016, 10, 1, 0, 0)}
>>> reverse_format('{state:2}{zip:5}', 'PA19104')
{'state': 'PA', 'zip': '19104'}

What’s next

  • So far only the 

     plugin and 


     plugin with 


     driver support this behavior. More plugins can be made to respect 


     notation, using the helper classes provided in Intake. The specific implementation may depend on the specifics of the third-party library.

  • This feature seems most helpful in the context of multiple file loading, however, the functionality may also be useful for parsing other similarly structured text in general.
  • Add 

     support for pattern as a path.

  • Swap out 

     for an external library such as 



You May Also Like

Company Blog
VS Code in Anaconda Distribution 5.1
A few months ago, Anaconda, Inc., creator of the world’s most popular Python data science platform, announced a partnership with Microsoft that included providing Anaconda D...
Read More
Data Science Blog
Continuum Analytics To Speak At TDWI Las Vegas Leadership Summit
Chief Data Scientist and co-founder Travis Oliphant to discuss Open Data Science and How it Drives Significant Value LAS VEGAS, Nev.—February 9, 2017—Continuum Analytics, ...
Read More
Data Science Blog
Intake: Caching Data on First Read Makes Future Analysis Faster
By Mike McCarty Intake provides easy access data sources from remote/cloud storage. However, for large files, the cost of downloading files every time data is read can be ext...
Read More