Write a Reduce Function

Role of the Reduce Function in MapReduce

mapreduce requires both an input map function that receives blocks of data and that outputs intermediate results, and an input reduce function that reads the intermediate results and produces a final result. Thus, it is normal to break up a calculation into two related pieces for the map and reduce functions to fulfill separately. For example, to find the maximum value in a data set, the map function can find the maximum value in each block of input data, and then the reduce function can find the single maximum value among all of the intermediate maxima.

This figure shows the Reduce phase of the mapreduce algorithm.

The Reduce phase of the mapreduce algorithm has the following steps:

The result of the Map phase of the mapreduce algorithm is an intermediate KeyValueStore object that contains all of the key-value pairs added by the map function. Before calling the reduce function, mapreduce groups the values in the intermediate KeyValueStore object by unique key. Each unique key in the intermediate KeyValueStore object results in a single call to the reduce function.
For each key, mapreduce creates a ValueIterator object that contains all of the values associated with that key.
The reduce function scrolls through the values from the ValueIterator object using the hasnext and getnext functions, which are typically used in a while loop.
After performing a summary calculation, the reduce function adds one or more key-value pairs to the final KeyValueStore object using the add and addmulti functions.

The Reduce phase of the mapreduce algorithm is complete when the reduce function processes all of the unique intermediate keys and their associated values. The result of this phase of the mapreduce algorithm (similar to the Map phase) is a KeyValueStore object containing all of the final key-value pairs added by the reduce function. After the Reduce phase, mapreduce pulls the key-value pairs from the KeyValueStore and returns them in a datastore (a KeyValueDatastore object by default). The key-value pairs in the output datastore are not in sorted order; they appear in the same order as they were added by the reduce function.

Requirements for Reduce Function

mapreduce automatically calls the reduce function for each unique key in the intermediate KeyValueStore object, so the reduce function must meet certain basic requirements to run properly during these automatic calls. These requirements collectively ensure the proper movement of data through the Reduce phase of the mapreduce algorithm.

The inputs to the reduce function are intermKey, intermValIter, and outKVStore:

intermKey is one of the unique keys added by the map function. Each call to the reduce function by mapreduce specifies a new unique key from the keys in the intermediate KeyValueStore object.
intermValIter is the ValueIterator object associated with the active key, intermKey. This ValueIterator object contains all of the values associated with the active key. Scroll through the values using the hasnext and getnext functions.
outKVStore is the name for the final KeyValueStore object to which the reduce function needs to add key-value pairs. The add and addmulti functions use this object name to add key-value pairs to the output. mapreduce takes the output key-value pairs from outKVStore and returns them in the output datastore, which is a KeyValueDatastore object by default. If the reduce function does not add any key-value pairs to outKVStore, then mapreduce returns an empty datastore.

In addition to these basic requirements for the reduce function, the key-value pairs added by the reduce function must also meet these conditions:

Keys must be numeric scalars, character vectors, or strings. Numeric keys cannot be NaN, logical, complex, or sparse.
All keys added by the reduce function must have the same class, but that class may differ from the class of the keys added by the map function.
If the OutputType argument of mapreduce is 'Binary' (the default), then a value added by the reduce function can be any MATLAB^® object, including all valid MATLAB data types.
If the OutputType argument of mapreduce is 'TabularText', then a value added by the reduce function can be a numeric scalar, character vector, or string. In this case, the value cannot be NaN, complex, logical, or sparse.

Note

The above key-value pair requirements may differ when using other products with mapreduce. See the documentation for the appropriate product to get product-specific key-value pair requirements.

Sample Reduce Functions

Here are a few illustrative reduce functions used in mapreduce examples.

Simple Reduce Function

One of the simplest examples of a reducer is maxArrivalDelayReducer, which is the reducer for the example Find Maximum Value with MapReduce. The map function in this example finds the maximum arrival delay in each chunk of input data. Then the reduce function finishes the task by finding the single maximum value among all of the intermediate maxima. To find the maximum value, the reducer scrolls through the values in the ValueIterator object and compares each value to the current maximum. mapreduce only calls this reducer function once, since the mapper adds a single unique key to the intermediate KeyValueStore object. The reduce function adds a single key-value pair to the output.

function maxArrivalDelayReducer(intermKey, intermValIter, outKVStore)
  % intermKey is 'PartialMaxArrivalDelay'. intermValIter is an iterator of
  % all values that has the key 'PartialMaxArrivalDelay'.
  maxVal = -Inf;
  while hasnext(intermValIter)
    maxVal = max(getnext(intermValIter), maxVal);
  end
  % The key-value pair added to outKVStore will become the output of mapreduce 
  add(outKVStore,'MaxArrivalDelay',maxVal);
end

Advanced Reduce Function

A more advanced example of a reducer is statsByGroupReducer, which is the reducer for the example Compute Summary Statistics by Group Using MapReduce. The map function in this example groups the data in each input using an extra parameter (airline carrier, month, and so on), and then calculates several statistical quantities for each group of data. The reduce function finishes the task by retrieving the statistical quantities and concatenating them into long vectors, and then using the vectors to calculate the final statistical quantities for count, mean, variance, skewness, and kurtosis. The reducer stores these values as fields in a structure, so that each unique key has a structure of statistical quantities in the output.

function statsByGroupReducer(intermKey, intermValIter, outKVStore)
% Reducer function for the StatisticsByGroupMapReduceExample.

% Copyright 2014 The MathWorks, Inc.

n = [];
m = [];
v = [];
s = [];
k = [];

% get all sets of intermediate statistics
while hasnext(intermValIter)
    value = getnext(intermValIter);
    n = [n; value(1)];
    m = [m; value(2)];
    v = [v; value(3)];
    s = [s; value(4)];
    k = [k; value(5)];
end
% Note that this approach assumes the concatenated intermediate values fit
% in memory. Refer to the reducer function, covarianceReducer,  of the
% CovarianceMapReduceExample for an alternative pairwise reduction approach

% combine the intermediate results
count = sum(n);
meanVal = sum(n.*m)/count;
d = m - meanVal;
variance = (sum(n.*v) + sum(n.*d.^2))/count;
skewnessVal = (sum(n.*s) + sum(n.*d.*(3*v + d.^2)))./(count*variance^(1.5));
kurtosisVal = (sum(n.*k) + sum(n.*d.*(4*s + 6.*v.*d +d.^3)))./(count*variance^2);

outValue = struct('Count',count, 'Mean',meanVal, 'Variance',variance,...
                 'Skewness',skewnessVal, 'Kurtosis',kurtosisVal);

% add results to the output datastore
add(outKVStore,intermKey,outValue);

More Reduce Functions

For more information about common programming patterns in map or reduce functions, see Build Effective Algorithms with MapReduce.

Documentation