Write a Map Function

Role of Map Function in MapReduce

mapreduce requires both an input map function that receives chunks 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 chunk of input data, and then the reduce function can find the single maximum value among all of the intermediate maxima.

This figure shows the Map phase of the mapreduce algorithm.

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

  1. mapreduce reads a single chunk of data using the read function on the input datastore, then calls the map function to work on the chunk.

  2. The map function then works on the individual chunk of data and adds one or more key-value pairs to the intermediate KeyValueStore object using the add or addmulti functions.

  3. mapreduce repeats this process for each of the chunks of data in the input datastore, so that the total number of calls to the map function is equal to the number of chunks of data. The ReadSize property of the datastore determines the number of data chunks.

The Map phase of the mapreduce algorithm is complete when the map function processes each of the chunks of data in the input datastore. The result of this phase of the mapreduce algorithm is a KeyValueStore object that contains all of the key-value pairs added by the map function. After the Map phase, mapreduce prepares for the Reduce phase by grouping all the values in the KeyValueStore object by unique key.

Requirements for Map Function

mapreduce automatically calls the map function for each chunk of data in the input datastore. The map function must meet certain basic requirements to run properly during these automatic calls. These requirements collectively ensure the proper movement of data through the Map phase of the mapreduce algorithm.

The inputs to the map function are data, info, and intermKVStore:

  • data and info are the result of a call to the read function on the input datastore, which mapreduce executes automatically before each call to the map function.

  • intermKVStore is the name of the intermediate KeyValueStore object to which the map function needs to add key-value pairs. The add and addmulti functions use this object name to add key-value pairs. If the map function does not add any key-value pairs to the intermKVStore object, then mapreduce does not call the reduce function and the resulting datastore is empty.

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

  1. Keys must be numeric scalars, character vectors, or strings. Numeric keys cannot be NaN, complex, logical, or sparse.

  2. All keys added by the map function must have the same class.

  3. Values can be any MATLAB® object, including all valid MATLAB data types.

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 Map Functions

These examples contain some map functions used by the mapreduce examples in the toolbox/matlab/demos folder.

Identity Map Function

A map function that simply returns what mapreduce passes to it is called an identity mapper. An identity mapper is useful to take advantage of the grouping of values by unique key before doing calculations in the reduce function. The identityMapper.m mapper file is one of the mappers used in the example file TSQRMapReduceExample.m.

type identityMapper.m
function identityMapper(data, info, intermKVStore)
% Mapper function for the MapReduce TSQR example.
%
% This mapper function simply copies the data and add them to the
% intermKVStore as intermediate values.

% Copyright 2014 The MathWorks, Inc.

x = data.Value{:,:};
add(intermKVStore,'Identity', x);

Simple Map Function

One of the simplest examples of a nonidentity mapper is maxArrivalDelayMapper.m, which is the mapper for the example file MaxMapReduceExample.m. For each chunk of input data, this mapper calculates the maximum arrival delay and adds a key-value pair to the intermediate KeyValueStore.

type maxArrivalDelayMapper.m
function maxArrivalDelayMapper (data, info, intermKVStore)
% Mapper function for the MaxMapreduceExample.

% Copyright 1984-2014 The MathWorks, Inc. 

% Data is an n-by-1 table of the ArrDelay. As the data source is tabular,
% the return of read is a table object.
partMax = max(data.ArrDelay);
add(intermKVStore, 'PartialMaxArrivalDelay',partMax);

Advanced Map Function

A more advanced example of a mapper is statsByGroupMapper.m, which is the mapper for the example file StatisticsByGroupMapReduceExample.m. This mapper uses a nested function to calculate several statistical quantities (count, mean, variance, and so on) for each chunk of input data, and then adds several key-value pairs to the intermediate KeyValueStore object. Also, this mapper uses four input arguments, whereas mapreduce only accepts a map function with three input arguments. To get around this, pass in the extra parameter using an anonymous function during the call to mapreduce, as outlined in the example.

type statsByGroupMapper.m
function statsByGroupMapper(data, ~, intermKVStore, groupVarName)
% Mapper function for the StatisticsByGroupMapReduceExample.

% Copyright 2014 The MathWorks, Inc.

% Data is a n-by-3 table. Remove missing values first
delays = data.ArrDelay;
groups = data.(groupVarName);
notNaN =~isnan(delays);
groups = groups(notNaN);
delays = delays(notNaN);

% find the unique group levels in this chunk
[intermKeys,~,idx] = unique(groups, 'stable');

% group delays by idx and apply @grpstatsfun function to each group
intermVals = accumarray(idx,delays,size(intermKeys),@grpstatsfun);
addmulti(intermKVStore,intermKeys,intermVals);

function out = grpstatsfun(x)
n = length(x); % count
m = sum(x)/n; % mean
v = sum((x-m).^2)/n; % variance
s = sum((x-m).^3)/n; % skewness without normalization
k = sum((x-m).^4)/n; % kurtosis without normalization
out = {[n, m, v, s, k]};

More Map Functions

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

See Also

| | |

Related Topics