Charred

Efficient character-based file parsing for csv and json formats.

• Zero dependencies.

• As fast as univocity or jackson.

• Same API as clojure.data.csv and clojure.data.json implemented far more efficiently.

• API Documentation

• Simple JSON benchmarks

• Large CSV Support

Usage

user> (require '[charred.api :as charred])
nil
user> (charred/read-json "{\"a\": 1, \"b\": 2}")
{"a" 1, "b" 2}
user> (charred/read-json "{\"a\": 1, \"b\": 2}" :key-fn keyword)
{:a 1, :b 2}
user> (println (charred/write-json-str *1))
{
  "a": 1,
  "b": 2
}

A Note About Efficiency

If you are reading or writing a lot of small JSON objects the best option is to create a specialized parse fn to exactly the options that you need and pass in strings or char[] data. A similar pathway exists for high performance writing of json objects. The returned functions are safe to use in multithreaded contexts.

The system is overall tuned for large files. Small files or input streams should be setup with :async? false and smaller :bufsize arguments such as 8192 as there is no gain for async loading when the file/stream is smaller than 1MB. For smaller streams slurping into strings in an offline threadpool will lead to the highest performance. For a particular file size if you know you are going to parse many of these then you should gridsearch :bufsize and :async? as that is a tuning pathway that I haven't put a ton of time into. In general the system is tuned towards larger files as that is when performance really does matter.

All the parsing systems have mutable options. These can be somewhat faster and it is interesting to look at the tradeoffs involved. Parsing a csv using the raw supplier interface is a bit faster than using the Clojure sequence pathway into persistent vectors and it probably doesn't really change your consume pathway so it may be worth trying it.

Development

Before running a REPL you must compile the java files into target/classes. This directory will then be on your classpath.

scripts/compile

Tests can be run with scripts/run-tests which will compile the java and then run the tests.

Lies, Damn Lies, and Benchmarks!

See the fast-json project. These times are for parsing a 100k json document using keywords for map keys - :key-fn keyword.

Intel JDK-8

method	performance µs
data.json	4275
jsonista	754
charred	638
charred-hamf	486

Intel JDK-19

method	performance µs
data.json	5608
jsonista	856
charred	673
charred-hamf	531

Mac m-1 JDK-19

method	performance µs
data.json	3164
jsonista	285
charred	249
charred-hamf	227

License

MIT license.