A few months ago, I was trying to get some bulk data into ebay's proprietary TurboLister[1] program. Guess what, it can import csv but not JSON.
SQLite[2] can import csv but not JSON.
Google's terabyte ngram dataset[3] is csv (tsv) instead of JSON. I'm glad it's not JSON because it would have required extra disk space.
... plus tons of other real-world csv examples out in the wild.
Unfortunately, the csv format is very easy for programs to write but it's very difficult for programs to properly read because of the tricky parsing.
[1] http://pages.ebay.com/sellerinformation/sellingresources/tur...
[2] http://www.sqlite.org/cvstrac/wiki?p=ImportingFiles
[3] http://storage.googleapis.com/books/ngrams/books/datasetsv2....
So no, if you control input and output, JSON is still easier to use than CSV, and just as performant. JSON stores straight arrays just fine. It's not the format's fault so many people choose to store hashes with it.
...and, the moment you declare you handle json, people will send non-string data ("that is a number, of course it isn't quoted"), attempt to include nested data, leave out the opening and closing [] (because people will grep a file with one array per line to filter a json file; that is no way robust, but people will do it, anyways)
The main advantage of json vs csv is that there is only one json, but quite a few different variant of csv. That is a huge advantage, but csv is engrained; there are many, many tools that handle csv but not json.
And the "if you control input and output" case is not the interesting or the problematic one.
Sometimes, simplicity is better for everyone.
I've had to() parse json embedded in a field in a csv file. Unquoted of course.
Until I explained to the other developer just how stupid that was.
CSV is bullshit, it's not good for anything except scenarios where you control both the export process, and the parser (so you know what delimiter is used and so on).
We process csv files containing payment information, export csv as product feeds for perhaps 10 different partners. It works, it's faster than XML or JSON and is easy for non-technical people to inspect using tools like Excel. Most cvs data exports are not something that's made up on the fly every day, it batch jobs that create the same type of CSV file each and every time.
CSV is alright. XML is bullshit.
Obviously, you can make CSV files work for alternating record types, in the same way that the old mainframe files used to work with multiple record types in a file, with a type descriminator field in a know place, usually field 1, of each record. But it starts to get cluttered.