Streaming/chunked parsing support (gzread(), gzgets() across boundaries)

[mutluit]

Does nlohmann::json support true incremental parsing where JSON tokens span arbitrary chunk boundaries, like when reading via zlib's gzread(64KB) and gzgets() → parse_chunk(buffer) → repeat?

Use case: Parse gzip compressed big files like ~35MB .json.gz (by multiple chunked reads via gzread()), and ~25MB .csv.gz (via multiple gzgets() calls) memory-efficiently:

//...
#include <zlib.h>
#include <nlohmann/json.hpp>

using json = nlohmann::json;

//....
json j;
gzFile f = gzopen("large.json.gz", "r");
char buf[64*1024];
while (gzread(f, buf, sizeof(buf)) > 0) {
    j.feed_chunk(buf);  // Resume parser state across mid-number/mid-string splits
}
gzclose(f);
//...

Current >> operator and sax_parse() require complete input upfront. Discussion #3411 confirms streaming multiple complete objects, but what about single large JSON documents with tokens split across chunks (like RapidJSON's Reader::Parse())?

I just tested this feature of the RapidJSON library and it works good, but for me it would be nice if also nlohmann::json had this very useful feature.
Rationale is of course: keep the files compressed (like in my case daily about 2000 files of size each 35MB...).

(I of course don't mean reformatting the data into JSONL/NDJSON format, rather keeping the original JSON format + compressed.)

Is this already possible via (undocumented) API, or on the TODO list? Thx.

[gregmarr]

I think you would need to create an istream that wraps it, and pass that istream to the parse() function. I thought maybe an input_stream_adapter would work, but that's just an overload set of functions, so you'd need to create a function in the detail namespace with the name input_stream_adapter.

[mutluit]

I think you would need to create an istream that wraps it, and pass that istream to the parse() function. I thought maybe an input_stream_adapter would work, but that's just an overload set of functions, so you'd need to create a function in the detail namespace with the name input_stream_adapter.

When reading blockwise (like in the example above), after parsing the initial block of data, the parser of course must ask for subsequent input data. How does parse() request the next block after it processed a block? Is there a callback function to implement? And how is EOF/end_of_stream to signal to the parser?

[gregmarr]

If you do it like this, it stops at the end of the complete object.

    istream istr(...);
    json j;
    istr >> j

[gregmarr]

The parse() function requests a character at a time from the istream.

[mutluit]

The parse() function requests a character at a time from the istream.

Does the library allow also blockwise processing of the data, ie. w/o first loading all into memory?

[mutluit]

I thought I solved it by using an old library named "gzstream" from the year 2001 (!): by creating an own class derived from std::istream, and passing that to nlohmann::json :
https://www.cs.unc.edu/Research/compgeom/gzstream/

BUT json gives parsing error:
terminate called after throwing an instance of 'nlohmann::json_abi_v3_11_2::detail::parse_error'
what(): [json.exception.parse_error.101] parse error at line 1, column 5: syntax error while parsing value - invalid literal; last read: '<U+0009><U+0009> G'
Aborted

:-(
Is that maybe an error in nlohmann::json ? Like reading the (compressed) stream itself, instead of reading via the derived class (which does provide the uncompressed stream to json).

    igzstream ifs("file.gz");
    json j;
#if 0
    ifs >> j;
#else
    j = json::parse(ifs);
#endif
    std::cout << j.dump(2) << std::endl;

[gregmarr]

I would make sure that you're getting the expected text by reading from the stream and writing to cout.

[mutluit]

I would make sure that you're getting the expected text by reading from the stream and writing to cout.

You are right! The file.gz is a just a text file, not a json file! :-) I'll test again....
YES, it now works! :-)

$ ./my.exe test.json.gz 
{
  "a": [
    1,
    2,
    3,
    4
  ],
  "f": false,
  "hello": "world",
  "i": 123,
  "n": null,
  "pi": 3.1416,
  "t": true
}