There are several methods of performing aggregations in MongoDB. These examples cover the new aggregation framework, using map reduce and using the group method.

.. testsetup::

  from pymongo import MongoClient
  client = MongoClient()
  client.drop_database('aggregation_example')

To start, we'll insert some example data which we can perform aggregations on:

>>> from pymongo import MongoClient
>>> db = MongoClient().aggregation_example
>>> result = db.things.insert_many([{"x": 1, "tags": ["dog", "cat"]},
...                                 {"x": 2, "tags": ["cat"]},
...                                 {"x": 2, "tags": ["mouse", "cat", "dog"]},
...                                 {"x": 3, "tags": []}])
>>> result.inserted_ids
[ObjectId('...'), ObjectId('...'), ObjectId('...'), ObjectId('...')]

This example shows how to use the :meth:`~pymongo.collection.Collection.aggregate` method to use the aggregation framework. We'll perform a simple aggregation to count the number of occurrences for each tag in the tags array, across the entire collection. To achieve this we need to pass in three operations to the pipeline. First, we need to unwind the tags array, then group by the tags and sum them up, finally we sort by count.

As python dictionaries don't maintain order you should use :class:`~bson.son.SON` or :class:`collections.OrderedDict` where explicit ordering is required eg "$sort":

>>> from bson.son import SON
>>> pipeline = [
...     {"$unwind": "$tags"},
...     {"$group": {"_id": "$tags", "count": {"$sum": 1}}},
...     {"$sort": SON([("count", -1), ("_id", -1)])}
... ]
>>> import pprint
>>> pprint.pprint(list(db.things.aggregate(pipeline)))
[{u'_id': u'cat', u'count': 3},
 {u'_id': u'dog', u'count': 2},
 {u'_id': u'mouse', u'count': 1}]

To run an explain plan for this aggregation use the :meth:`~pymongo.database.Database.command` method:

>>> db.command('aggregate', 'things', pipeline=pipeline, explain=True)
{u'ok': 1.0, u'stages': [...]}

As well as simple aggregations the aggregation framework provides projection capabilities to reshape the returned data. Using projections and aggregation, you can add computed fields, create new virtual sub-objects, and extract sub-fields into the top-level of results.

.. seealso:: The full documentation for MongoDB's `aggregation framework
    <http://docs.mongodb.org/manual/applications/aggregation>`_

Another option for aggregation is to use the map reduce framework. Here we will define map and reduce functions to also count the number of occurrences for each tag in the tags array, across the entire collection.

Our map function just emits a single (key, 1) pair for each tag in the array:

>>> from bson.code import Code
>>> mapper = Code("""
...               function () {
...                 this.tags.forEach(function(z) {
...                   emit(z, 1);
...                 });
...               }
...               """)

The reduce function sums over all of the emitted values for a given key:

>>> reducer = Code("""
...                function (key, values) {
...                  var total = 0;
...                  for (var i = 0; i < values.length; i++) {
...                    total += values[i];
...                  }
...                  return total;
...                }
...                """)

Finally, we call :meth:`~pymongo.collection.Collection.map_reduce` and iterate over the result collection:

>>> result = db.things.map_reduce(mapper, reducer, "myresults")
>>> for doc in result.find():
...   pprint.pprint(doc)
...
{u'_id': u'cat', u'value': 3.0}
{u'_id': u'dog', u'value': 2.0}
{u'_id': u'mouse', u'value': 1.0}

PyMongo's API supports all of the features of MongoDB's map/reduce engine. One interesting feature is the ability to get more detailed results when desired, by passing full_response=True to :meth:`~pymongo.collection.Collection.map_reduce`. This returns the full response to the map/reduce command, rather than just the result collection:

>>> pprint.pprint(
...     db.things.map_reduce(mapper, reducer, "myresults", full_response=True))
{...u'counts': {u'emit': 6, u'input': 4, u'output': 3, u'reduce': 2},
 u'ok': ...,
 u'result': u'...',
 u'timeMillis': ...}

All of the optional map/reduce parameters are also supported, simply pass them as keyword arguments. In this example we use the query parameter to limit the documents that will be mapped over:

>>> results = db.things.map_reduce(
...     mapper, reducer, "myresults", query={"x": {"$lt": 2}})
>>> for doc in results.find():
...   pprint.pprint(doc)
...
{u'_id': u'cat', u'value': 1.0}
{u'_id': u'dog', u'value': 1.0}

You can use :class:`~bson.son.SON` or :class:`collections.OrderedDict` to specify a different database to store the result collection:

>>> from bson.son import SON
>>> pprint.pprint(
...     db.things.map_reduce(
...         mapper,
...         reducer,
...         out=SON([("replace", "results"), ("db", "outdb")]),
...         full_response=True))
{...u'counts': {u'emit': 6, u'input': 4, u'output': 3, u'reduce': 2},
 u'ok': ...,
 u'result': {u'collection': ..., u'db': ...},
 u'timeMillis': ...}

.. seealso:: The full list of options for MongoDB's `map reduce engine <http://www.mongodb.org/display/DOCS/MapReduce>`_