Parse large JSON file in Nodejs

Question

To process a file line-by-line, you simply need to decouple the reading of the file and the code that acts upon that input. You can accomplish this by buffering your input until you hit a newline. Assuming we have one JSON object per line (basically, format B):

var stream = fs.createReadStream(filePath, {flags: 'r', encoding: 'utf-8'});
var buf = '';

stream.on('data', function(d) {
    buf += d.toString(); // when data is read, stash it in a string buffer
    pump(); // then process the buffer
});

function pump() {
    var pos;

    while ((pos = buf.indexOf('\n')) >= 0) { // keep going while there's a newline somewhere in the buffer
        if (pos == 0) { // if there's more than one newline in a row, the buffer will now start with a newline
            buf = buf.slice(1); // discard it
            continue; // so that the next iteration will start with data
        }
        processLine(buf.slice(0,pos)); // hand off the line
        buf = buf.slice(pos+1); // and slice the processed data off the buffer
    }
}

function processLine(line) { // here's where we do something with a line

    if (line[line.length-1] == '\r') line=line.substr(0,line.length-1); // discard CR (0x0D)

    if (line.length > 0) { // ignore empty lines
        var obj = JSON.parse(line); // parse the JSON
        console.log(obj); // do something with the data here!
    }
}

Each time the file stream receives data from the file system, it's stashed in a buffer, and then pump is called.

If there's no newline in the buffer, pump simply returns without doing anything. More data (and potentially a newline) will be added to the buffer the next time the stream gets data, and then we'll have a complete object.

If there is a newline, pump slices off the buffer from the beginning to the newline and hands it off to process. It then checks again if there's another newline in the buffer (the while loop). In this way, we can process all of the lines that were read in the current chunk.

Finally, process is called once per input line. If present, it strips off the carriage return character (to avoid issues with line endings – LF vs CRLF), and then calls JSON.parse one the line. At this point, you can do whatever you need to with your object.

Note that JSON.parse is strict about what it accepts as input; you must quote your identifiers and string values with double quotes. In other words, {name:'thing1'} will throw an error; you must use {"name":"thing1"}.

Because no more than a chunk of data will ever be in memory at a time, this will be extremely memory efficient. It will also be extremely fast. A quick test showed I processed 10,000 rows in under 15ms.

Just as I was thinking that it would be fun to write a streaming JSON parser, I also thought that maybe I should do a quick search to see if there's one already available.

Turns out there is.

• JSONStream "streaming JSON.parse and stringify"

Since I just found it, I've obviously not used it, so I can't comment on its quality, but I'll be interested to hear if it works.

It does work consider the following Javascript and _.isString:

stream.pipe(JSONStream.parse('*'))
  .on('data', (d) => {
    console.log(typeof d);
    console.log("isString: " + _.isString(d))
  });

This will log objects as they come in if the stream is an array of objects. Therefore the only thing being buffered is one object at a time.

As of October 2014, you can just do something like the following (using JSONStream) - https://www.npmjs.org/package/JSONStream

var fs = require('fs'),
    JSONStream = require('JSONStream'),

var getStream() = function () {
    var jsonData = 'myData.json',
        stream = fs.createReadStream(jsonData, { encoding: 'utf8' }),
        parser = JSONStream.parse('*');
    return stream.pipe(parser);
}

getStream().pipe(MyTransformToDoWhateverProcessingAsNeeded).on('error', function (err) {
    // handle any errors
});

To demonstrate with a working example:

npm install JSONStream event-stream

data.json:

{
  "greeting": "hello world"
}

hello.js:

var fs = require('fs'),
    JSONStream = require('JSONStream'),
    es = require('event-stream');

var getStream = function () {
    var jsonData = 'data.json',
        stream = fs.createReadStream(jsonData, { encoding: 'utf8' }),
        parser = JSONStream.parse('*');
    return stream.pipe(parser);
};

getStream()
    .pipe(es.mapSync(function (data) {
        console.log(data);
    }));

$ node hello.js
// hello world

I had similar requirement, i need to read a large json file in node js and process data in chunks and call a api and save in mongodb. inputFile.json is like:

{
 "customers":[
       { /*customer data*/},
       { /*customer data*/},
       { /*customer data*/}....
      ]
}

Now i used JsonStream and EventStream to achieve this synchronously.

var JSONStream = require("JSONStream");
var es = require("event-stream");

fileStream = fs.createReadStream(filePath, { encoding: "utf8" });
fileStream.pipe(JSONStream.parse("customers.*")).pipe(
  es.through(function(data) {
    console.log("printing one customer object read from file ::");
    console.log(data);
    this.pause();
    processOneCustomer(data, this);
    return data;
  }),
  function end() {
    console.log("stream reading ended");
    this.emit("end");
  }
);

function processOneCustomer(data, es) {
  DataModel.save(function(err, dataModel) {
    es.resume();
  });
}

I realize that you want to avoid reading the whole JSON file into memory if possible, however if you have the memory available it may not be a bad idea performance-wise. Using node.js's require() on a json file loads the data into memory really fast.

I ran two tests to see what the performance looked like on printing out an attribute from each feature from a 81MB geojson file.

In the 1st test, I read the entire geojson file into memory using var data = require('./geo.json'). That took 3330 milliseconds and then printing out an attribute from each feature took 804 milliseconds for a grand total of 4134 milliseconds. However, it appeared that node.js was using 411MB of memory.

In the second test, I used @arcseldon's answer with JSONStream + event-stream. I modified the JSONPath query to select only what I needed. This time the memory never went higher than 82MB, however, the whole thing now took 70 seconds to complete!

I wrote a module that can do this, called BFJ. Specifically, the method bfj.match can be used to break up a large stream into discrete chunks of JSON:

const bfj = require('bfj');
const fs = require('fs');

const stream = fs.createReadStream(filePath);

bfj.match(stream, (key, value, depth) => depth === 0, { ndjson: true })
  .on('data', object => {
    // do whatever you need to do with object
  })
  .on('dataError', error => {
    // a syntax error was found in the JSON
  })
  .on('error', error => {
    // some kind of operational error occurred
  })
  .on('end', error => {
    // finished processing the stream
  });

Here, bfj.match returns a readable, object-mode stream that will receive the parsed data items, and is passed 3 arguments:

1. A readable stream containing the input JSON.

2. A predicate that indicates which items from the parsed JSON will be pushed to the result stream.

3. An options object indicating that the input is newline-delimited JSON (this is to process format B from the question, it's not required for format A).

Upon being called, bfj.match will parse JSON from the input stream depth-first, calling the predicate with each value to determine whether or not to push that item to the result stream. The predicate is passed three arguments:

1. The property key or array index (this will be undefined for top-level items).

2. The value itself.

3. The depth of the item in the JSON structure (zero for top-level items).

Of course a more complex predicate can also be used as necessary according to requirements. You can also pass a string or a regular expression instead of a predicate function, if you want to perform simple matches against property keys.