What file formats are on the data portals?

I found some more open data about open data to study! While at Socrata's office on Friday, I learned of the /data.json endpoint. It contains an entry for each dataset, uploaded by the data publisher; it doesn't contain all of the other views that are based on these source datasets. And it has this format.

How many datasets?

Socrata portals have 50,000 different views, but only 8922 are original datasets.

The /data.json files include federated datasets, so some of these datasets are duplicated. I did not remove duplicates, so I'm working with 15699 datasets, with a median of 96 datasets per portal.

Datasets per portal, based on the /data.json file

No derived datasets

This is much lower than my earlier figure because the present figure does not include derived views (map, charts, &c.).

Some time, I'll compare the within-portal counts of original datasets and derived datasets. But not right now.

Federated data

It still includes federated data (duplicates), however, and this file doesn't make it easy to determine which direction the federation is in. The following plot gives us an idea of how many of these datasets are duplicates.

The scale of data federation within this subset of Socrata datasets

Most of the datasets are not duplicates, but some are duplicated many times.

Some day, I'll look more at federation, probably by reading the federation information from home pages of the portals or by following the links in the /data.json file.

Cutoff at 1000?

I find it highly suspicious the following nine portals have exactly 1000 datasets and no portals have more than 1000 datasets.

paste(names(sort(table(datasets$portal), decreasing = T)[1:9]), collapse = "\n")

Half of these portals federate explore.data.gov, which has quite a few datasets, and the JSON files seem to just include some of the explore.data.gov data. I think these files have only the first 1000 datasets, and I haven't figured out how to look at the next pages, so I'll focus the present analysis on portals with fewer than 1000 datasets.

My curiosity about file formats

I've recently become curious about what formats the datasets come from. When tabular data get loaded into a Socrata data portal, they get converted to a tabular representation within the portal software. From that, they get converted to a range of different tabular formats.

The Socrata data portal doesn't explicitly store the source format because of how the import process works. Most of the data probably come from Excel, and the data that aren't from Excel typically come from inside of a government network where policies would make it inconvenient to expose the database to the world. Because of this, Socrata doesn't query database servers. Instead, data publishers write middlemen that act as both database clients and Socrata clients. They query the database and then make web requests to the Socrata portal.

/data.json contains file format information

The /data.json endpoint contains a file format field per the Project Open Data schema. This refers to the format of the data as served from the Socrata portal, not the format it was stored in before it got to the Socrata portal. But this still tells us something about the source file formats.

People sometimes upload things that Socrata doesn't interpret as tables. PDFs are a major example. Other times, people upload or link to files that could be tables but don't specify that they are tables; those are called "external links". Read more on dataset types here.

Data formats are represented in two fields, format and distribution. distribution seems to contain all of the different available formats. If the data are imported as tabular data, it contains CSV, JSON, XML, &c., all served from the Socrata site. And if the data are external links, it will contain a few external links, still specifying the file types. The format field contains one of the formats that are specified in the distribution field. I think it's just the first of the formats. For the present analysis, I'm using the format field.

Recall that the present dataset of datasets counts federated datasets multiple times. The following plot shows the file types of the deduplicated dataset dataset, across all portals.

Formats of datasets across all portals

And here are some of the main types by portal, counting federated datasets in all of their portals.

Data formats on all the portals

csv mostly refers to data that Socrata represents as a table; this is the sort of data that Socrata can convert to a range of different tabular data formats. It is also my preferred file format.

Everything that is not CSV appears to be an external link.


Most datasets are CSV (8143 of 15699). I was curious as to how this varies by portal and over time, and the following image addresses that.

Dataset formats by portal over time

The image above contains one plot per data portal. The x-axis of each plot is the date, the y-axis is the proportion[^proportion] of datasets that are tabular (CSV), and the width of the line is the number of datasets on the portal. For example, if there were 100 datasets on a portal in June 2011 and 80 were CSV, the line would be near the top of the graph and quite skinny at June 2011.

We can thus see how many datasets each portal has and what the different formats are.

Some interesting portals

Some portals have only CSV data (like data.medicare.gov), but most contain other data. I am curious both as to what other data formats they have and what prompted the shifts in dataset format.

Missouri mostly has PDFs.

Data formats on data.mo.gov

Also interesting about Missouri is that it federates Kansas City, which didn't appear in my list of portals.

I know I said I'd focus on portals with fewer than 1000 datasets, but Lehman College is interesting because it has lots of zipped files.

Data formats on bronx.lehman.cuny.edu

San Francisco has a lot of CSVs, a lot of externally linked zip files, and a lot of externally linked files of unknown format.

Data formats on data.sfgov.org

It looks like the format of external links is determined by the file name. For example, Edmonton's Road and Traffic Updates are marked as application/rss+xml because the external link, http://www.trumba.com/calendars/construction-and-special-events-road-closures.rss, ends in .rss.

In contrast, the Maryland Land Use/ Land Cover: 1973, 2002, 2010 dataset is marked as having the format application/octet-stream because the external link, http://planning.maryland.gov/OurWork/landUseDownload.shtml, ends in .shtml.

Dates of significant changes

A few portals have only CSV data since the beginning, but most have had other formats. Looking at the plots, we can see dates where there was a sudden change in the proportion of datasets that were CSV.

Sudden changes at the beginning

When the first dataset gets uploaded, the proportion of datasets that are CSV is either zero or one. Thus, the line for all of these datasets starts either at zero or one. Most datasets sharply change after that; data.austintexas.gov and data.mo.gov are examples.

Others stay at this level for quite a while because no datasets were uploaded for a while. data.raleighnc.gov is an example of this. Here are its first ten datasets.

title created format
Building Permit Data 2012-03-14 text/csv
Building Permit Data 2012-03-14 text/csv
City of Raleigh Quickfacts 2013-02-19 text/csv
Raleigh Parking 2013-02-28 application/zip
Raleigh Electric Utilities 2011 2013-02-28 application/zip
Raleigh Communications 2011 2013-02-28 application/zip
Raleigh Buildings 2011 2013-02-28 application/zip
Raleigh Parks and Trails 2013-02-28 application/zip
Raleigh Trail Areas 2013-02-28 application/zip
Family Income In The Past 12 Months 2013-02-28 text/csv

The first two datasets were uploaded in the middle of March 2012 and were CSV format, making the datasets 100% CSV. The next was uploaded in the middle of February 2013 and was also CSV, so the proportion was still 100% CSV. At the end of the month, six zip files were uploaded, reducing the proportion to 33% CSV.

Sudden changes after the beginning

data.sfgov.org had a sudden change in the proportion of datasets that were CSV, but it was after a long while, so a lot of related datasets might have been uploaded all at once. Let's look at when datasets were uploaded.

Formats of newly open San Francisco datasets over time

This has the quite similar information to the earlier plots, but it's a bit more precise. San Francisco added lots of datasets in January 2012, November 2012, and December 2012, and proportionately few of these datasets were CSV. What were they?

Here are ten of the January datasets.

title created format
Arterial Streets of San Francisco (Zipped Shapefile Format) 2012-01-01 application/zip
Orthophoto 1ft resolution (1993) - (Zipped MrSID Format) 2012-01-01 application/zip
City Lots (Zipped Shapefile Format) 2012-01-01 application/zip
Census 2000 Block Group (No Water) (Zipped Shapefile Format) 2012-01-01 application/zip
Orthophoto 1ft (1993) - Treasure Island (Zipped MrSID Format) 2012-01-01 application/zip
SFPD Sectors 2012-01-01 application/zip
SFPD Crime Reporting Plots (Zipped Shapefile Format) 2012-01-01 application/zip
Neighborhood Marketplace Initiative Corridors (Zipped Shapefile Format) 2012-01-01 application/zip
San Francisco Basemap Street Centerlines (Zipped Shapefile Format) 2012-01-01 application/zip
The Presidio of San Francisco (Zipped Shapefile Format) 2012-01-01 application/zip

It looks like January is mostly externally linked, zipped shapefiles. Most of the datasets say "shapefile" in their title, description or distribution fields.

Formats of newly open San Francisco datasets over time

And a lot of the rest of the January files look like zipped shapefiles, even though the titles and descriptions don't say so.

So San Francisco suddenly uploaded a bunch of shapefiles in January and November/December.

Plateaus of dataset counts

Some of the plots make it look like lots of datasets were suddenly uploaded one time and no datasets were uploaded again. This is mainly for the portals with more than 1000 datasets, so I think this is because we're seeing only the first 1000 datasets.

CSV, pdf, zip and octet-stream

Based on the examples above, it seems like a lot of datasets are PDF, zip or unknown external links. I made the following series of plots to check it. It is just like the similar image above except for the y-axes; instead of representing the proportion of datasets that are CSV, each y-axis represents the proportion of datasets that are CSV, PDF, zip or unknown external links.

Dataset formats by portal over time

Most of the curves are pretty straight and stay near 1, meaning that the proportion doesn't change much and that the proportion is quite high. Thus, it looks like most datasets are either CSV, pdf, zip or external links of unknown format.



I was mainly wondering about the source formats of the data. It turns out that most datasets are tables. Aside from tables, there are lots of externally linked files, mostly PDFs and zip archives.


I've been thinking recently about how to infer the organizational structure of a municipality based on the open data that they release, and the finding with the San Francisco shapefiles alludes to this. It might be that one department within the San Francisco government manages and uses most of the geospatial data and that the open data team happened to work with them in January, November and December of 2012.

Future study

This got me thinking about other ways of studying file formats.

The attribution field

Socrata's SODA 1 API, which I've used before, contains an attribution field, which references the URL from which the dataset was taken.[^attribution] This would be one way of figuring out the source format, or at least some related information about the source.

Externally linked CSVs

Not all of the CSV-formatted datasets necessarily come from Socrata; some might be links to external CSV files. There is enough information in /data.json to determine which of these categories a CSV dataset falls into.

It looks to me like the format type of external links is determined based on the file extension of the URL; octet-stream datasets seem to correspond to URLs without file extensions or with file extensions like aspx that don't clearly correspond to a particular file types. One could determine the formats of these datasets by downloading the files.


[^proportion]: It's actually a tad bit more complicated than that. These dates are the creation dates of the datasets that are available today; I do not know about datasets that were historically on the portal and have since been deleted.

[^attribution]: I presume that this is entered manually.