New package: jsonlite. A smart(er) JSON encoder/decoder.
December 6, 2013
This week we released a new package on CRAN: jsonlite. This package is a fork of
RJSONIO by Duncan Temple Lang and builds on the same parser, but uses a different mapping between R objects and JSON data. The package vignette goes in great detail and has many examples on how JSON data are converted to R objects and vice versa. To try it:
So what’s new?
jsonlite package implements functions
fromJSON similar to those in packages as
rjson, but options and output are quite different. The primary goal in the design of
jsonlite is to recognize and comply with conventional ways of encoding data in JSON (outside the R community), in particular (relational) datasets. This increases interoperability when dealing with external data from within R, or when reading/writing R objects from an external client (e.g. through OpenCPU). For example, consider structures as returned by the Github API:
- Simple dataset: https://api.github.com/users/hadley/orgs
- Nested dataset: https://api.github.com/users/hadley/repos
These JSON structures obviously represent data tables, or in R terminology: data frames. The first dataset is a single table; the second dataset has a relational structure with two tables: the
owner property in the main table was generated from a foreign key that points to a record in a second table (owners). However, in their JSON representation these tables are structured by row, wereas R likes data frames by column. This is one example where
jsonlite goes beyond other packages, and actually returns a data frame:
The second example is a bit more complicated because of the relational structure.
jsonlite tries to stay as close as possible to the original structure by returing a nested data frame:
The package vignette gives many more examples of how various structures map to R objects.
On correctness and performance
The initial emphasis in jsonlite has been on correctness: rather than rushing towards performance, we want to explicity specify intended behavior covering all important structures. The complexity of this problem is easily understimated, which can result in unexpected behavior, ambiguous edge cases and differences between implementations. A set of conventions for a consistent and practical mapping are proposed in the package vignette. If you are using JSON with R, free to join the discussion.
Premature optimization is the root of all evil.Donald Knuth
We hope that a clear specifiction will make it much easier to optimize performance or write alternate implementations. The package vignette and package unit tests are intended to take away ambiguity on what exactly
fromJSON are supposed to do. From here we will start optimizing R code, port pieces to C++, or perhaps even write an entirely new implementation, without breaking software that depends on it.
If you would like to contribute to
jsonlite, you can submit patches or pull requests on github, as long as they don’t alter the behavior of the functions. At a minimum, they should pass the package unit tests… or you should modify the unit tests that are overly strict :-)