CLLD – Cross-Linguistic Linked Data

Data curation strategies of CLLD databases

posted Thursday, July 09, 2015 by Robert Forkel

Two and a half years into the CLLD project we see two main strategies of coping with the versioning problem. In this post I will describe these strategies as exemplified by the WALS and Tsammalex database respectively and investigate their strengths and weaknesses.

Both databases are published by corresponding clld applications, so data access is pretty much the same. The main difference in their data curation strategies is in the role, the PostgreSQL database of this application plays.

The WALS model

In the case of WALS, the data is curated in this database, i.e.

  • the PostgreSQL software is used to ensure integrity of the data and must be used to interact with the data.
  • WALS data in plain text formats is derived from this database via code (to export or dump).

Given these constraints, updating and versioning are handled as follows:

  • Modification of the data is effected through migrations, i.e. code which acts on the data using SQL. Each migration script constitutes a version of the database.
  • Each version of database can be reconstructed by applying the appropriate number of migrations to a known state of the database.
  • To reconstruct versions of the derived text formats, the relevant code used to create these formats needs to be reconstructed (and versioned) as well.

The Tsammalex model

In the spirit of “data is code” Tsammalex data is curated as set of csv files managed with the version control system git. This git repository is hosted on GitHub.

The database used by the clld app serving is derived from these csv files, or more precisely from releases of the git repository storing these files. Thus, the clld app and its database are purely a publishing mechanism for the data, i.e. they provide a user interface and an API.

Strengths and weaknesses

  • Changes of the database schema: Since - in the WALS model - migration scripts, i.e. modifications of the data, are part of the application code, application code and corresponding database version are synchronized automatically; whereas the Tsammalex model requires out-of-band mechanisms (like naming conventions for data and application code releases) to enable synchronization.

  • Errata or intra-edition modification of the database: With the WALS model, errata can be fixed transparently in the published database; whereas the Tsammalex model only allows modification of the data in the git repository, but requires a new edition (i.e. release) to push these modifications to the clld application.

  • Collaboration:

    • The barrier to entry for collaboration on the data is significantly higher for WALS, since it requires proficiency in programming, both in the python language and in SQL. It could be argued, though, that collaborating via editing csv files kept in git is not particularly easy either. But leaving the issue of editing text files aside, the collaboration workflow with git and GitHub is well documented and becoming standard in the open source world.

    • Since migration scripts are ordered linearly, the WALS model does only allow reconstruction of stages in this linear version history. With the Tsammalex model, reconstruction of the database for any state of the git data repository is possible - including constructing the database and running the clld application on top of private forks of the data repository. Thus, Tsammalex trivially provides functionality often requested for WALS: How can I compare my own data with WALS?

  • History: Since the WALS application database keeps a history of all changes it is possible to present a list of changes on the WALS website. However, tracing these changes back to corresponding migrations, e.g. to determine authorship, requires investigation of the commit history of the application software. With Tsammalex keeping a history of changes is entirely left to git (which was built for exactly this purpose, after all); and considering that collaboration is easier and more transparent in the Tsammalex case, the history of changes may also reveal more information.

  • Public vs. private: While with the WALS model preparation of new editions happens in private by default (because there is no public curation platform), with Tsammalex public is the default. But in the case of Tsammalex, it would be easy to fork the data repository into a private repos, modify/add to the data, when done, submit a pull request to clld/tsammalex-data.

  • Applicability: Clearly the Tsammalex model is limited to data which can be serialized and edited as text files. So it can not be applied to non-textual data or too much data. These limitations do not exist for the WALS model (although there are limits to the amount of data you should store in a PostgreSQL database).


Considering the advantages and disadvantages of these two data curation models, it seems clear that both are of value. Thus, the value proposition of the CLLD platform - a uniform API for cross-linguistic data - makes sense even without support for data curation within the clld software.