MongoDB unique value aggregation via map reduce

Question

I see plenty of questions on SO about aggregation in MongoDB, however, I have not found a complete solution to mine yet.

Here's an example of my data:

{
    "fruits" : {
        "apple" : "red",
        "orange" : "orange",
        "plum" : "purple"
    }
}
{
    "fruits" : {
        "apple" : "green",
        "plum" : "purple"
    }
}
{
    "fruits" : {
        "apple" : "red",
        "orange" : "yellow",
        "plum" : "purple"
    }
}

Now, my goal is to determine the popularity of each color for each fruit, so something like this would be the output collection:

{
    "_id" : "apple"
    "values" : {
        "red" : 2,
        "green" : 1
    }
}
{
    "_id" : "orange"
    "values" : {
        "orange" : 1,
        "yellow" : 1
    }
}
{
    "_id" : "plum"
    "values" : {
        "purple" : 3
    }
}

I have tried various M/R functions, and in the end they either don't work, or they take exponentially long. In the context of the example (fruit), I have about 1,000 different fruits and 100,000 colors over about 10,000,000 total documents. My current working M/R is this:

map = function() {
    if (!this.fruits) return;
    for (var fruit in this.fruits) {
        emit(fruit, {
            val_array: [
                {value: this.fruits[fruit], count: 1}
            ]
        });
    }
};

reduce = function(key, values) {
    var collection = {
        val_array: []
    };
    var found = false;
    values.forEach(function(map_obj) {
        map_obj.val_array.forEach(function(value_obj) {
            found = false;
            // if exists in collection, inc, else add
            collection.val_array.forEach(function(coll_obj) {
                if (coll_obj.value == value_obj.value) {
                    // the collection already has this object, increment it
                    coll_obj.count += value_obj.count;
                    found = true;
                    return;
                }
            });
            if (!found) {
                // the collection doesn't have this obj yet, push it
                collection.val_array.push(value_obj);
            }
        });
    });
    return collection;
};

Now, this does work, and for 100 records, it takes just a second or so, but the time increases non-linearly, so 100M records would take a very long time. The problem is that I'm doing a poor-mans sub-aggregation in the reduce function with the collection array, thus requiring me to iterate over both collection and the values from my map function. Now I just need to figure out how to do this efficiently (even if it requires multiple reductions). Any suggestions are welcome!

---

EDIT For lack of a better place to post it, here's my solution.

---

First, I created a file called mr.js:

map = function() {
    if (!this.fruits) return;
    var skip_fruits = {
        'Watermelon':1,
        'Grapefruit':1,
        'Tomato':1 // yes, a tomato is a fruit
    }
    for (var fruit in this.fruits) {
        if (skip_fruits[fruit]) continue;
        var obj = {};
        obj[this.fruits[fruit]] = 1;
        emit(fruit, obj);
    }
};

reduce = function(key, values) {
    var out_values = {};
    values.forEach(function(v) {
        for(var k in v) { // iterate values
            if (!out_values[k]) {
                out_values[k] = v[k]; // init missing counter
            } else {
                out_values[k] += v[k];
            }
        }
    });
    return out_values;
};

var in_coll = "fruit_repo";
var out_coll = "fruit_agg_so";
var total_docs = db[in_coll].count();
var page_size = 100000;
var pages = Math.floor(total_docs / page_size);
print('Starting incremental MR job with '+pages+' pages');
db[out_coll].drop();
for (var i=0; i<pages; i++) {
    var skip = page_size * i;
    print("Calculating page limits for "+skip+" - "+(skip+page_size-1)+"...");
    var start_date = db[in_coll].find({},{date:1}).sort({date:1}).skip(skip).limit(1)[0].date;
    var end_date = db[in_coll].find({},{date:1}).sort({date:1}).skip(skip+page_size-1).limit(1)[0].date;
    var mr_command = {
        mapreduce: in_coll,
        map: map,
        reduce: reduce,
        out: {reduce: out_coll},
        sort: {date: 1},
        query: {
            date: {
                $gte: start_date,
                $lt: end_date
            }
        },
        limit: (page_size - 1)
    };
    print("Running mapreduce for "+skip+" - "+(skip+page_size-1));
    db[in_coll].runCommand(mr_command);
}

That file iterates over my entire collection, incrementally map/reducing 100k docs (sorted by date which MUST have an index!) at a time, and reducing them into a single output collection. It's used like this: mongo db_name mr.js.

Then, after a couple hours, I've got a collection with all the info. To figure out which fruits have the most colors, I use this from the mongo shell to print out the top 25:

// Show number of number of possible values per key
var keys = [];
for (var c = db.fruit_agg_so.find(); c.hasNext();) {
    var obj = c.next();
    if (!obj.value) break;
    var len=0;for(var l in obj.value){len++;}
    keys.push({key: obj['_id'], value: len});
}
keys.sort(function(a, b){
    if (a.value == b.value) return 0;
    return (a.value > b.value)? -1: 1;
});
for (var i=0; i<20; i++) {
    print(keys[i].key+':'+keys[i].value);
}

The really cool thing about this approach is that since it's incremental, I can work with the output data while the mapreduce is running.

Answer 1

It seems that you don't really need val_array. Why not use a simple hash? Try this:

map = function() {
    if (!this.fruits) return;
    for (var fruit in this.fruits) {
        emit(fruit, 
             {this.fruits[fruit]: 1});
    }
};

reduce = function(key, values) {
  var colors = {};

  values.forEach(function(v) {
    for(var k in v) { // iterate colors
      if(!colors[k]) // init missing counter
        colors[k] = 0

      color[k] += v[k];
    }
  });

  return colors;
}