Mapreduce

December 24,2022

last update: 2024-11-19

MapReduce is a programming model published in 2004 by google, to process large data in clusters

in general, it involves 2 events

map(String key, String value):
// key: document name
// value: document contents
    for word in value:
        EmitIntermediate(w, "1"); // Write to intermediate file store

reduce(String key, Iterator values):
// key: a word
// values: a list of counts

    int result = 0 
    for each v in values:
        result += ParseInt(v);
    Emit(AsString(result)); //Write to output file

Workflow

• MapReduce library in the user program splits the input files into M pieces. Then it starts up copies of the program in a cluster of machines.
• One copy is the master, everything else are workers. There are M map tasks and R reduce tasks to assign. Master assigns M/R to idle workers.
• worker parses the input data and passes each pair to user-defined map function. intermediate values are buffered in memory.
• Periodically, buffered values => written to disk => forwarded to master => forwarded to reduce workers
• Reduce workers => notified by master, of the disk location => r-workers make rpc call to fetch file from m-worker => sort based on keys (so that same intermediate keys are grouped together) => passes key and correspond intermediaet value to the user-defined reduce function. output is appended ot a final output file for this partition.
• when all map-reduce tasks are done. master wakes up the user program.

Master Data Structures

• State of tasks
• intermediate data location