# MapReduce and Spark Implementation
## Overview
1. Given m documents, compute the term-term relevance using MapReduce algorithm and Spark implementation.
2. Input a text file, each line represents a document.
3. Poutput a list of term-term paris sorted by their similarity descending.
4. Compute Term Frequency – inverse Document Frenquency (TF-IDF) for each term.
5. Compute and sort term-term relevance between a query term and all terms associated with the TF-IDF matrix.
## Run
### Step 1:
– In line 22, change the file name ‘project2_test.txt’ to the file name that you named.
– OR just using the ‘project2_test.txt’ file that I attached in this zip file.
### Step 2:
– Install numpy package by using the command ‘pip install numpy’.
– Install pyspark package by using the command ‘pip install pyspark’.
– Install psutil package by using the command ‘pip install psutil’.
### Step 3:
– Put the code into a python ide and run the code.
– OR run the code in the command line interface.
## Potential Improvements
1. Try to reduce operations like gropByKey(), reducebyKey(), join().
– The function groupBykey must hold all the key-value pair in memory and if a key
has too many values, it can cause an out of memory error.
2. Reduce shuffling
– Spark uses shuffling to redistribute data.
– Shuffling is an expensive operation.
3. Caching
– Spark will store the dataset in memory which allows for faster access and
retrieval.
4. Dynamic allocation
– Scaling up or down based number of executors based on workload.
5. Data Skewing
– There might be uneven distribution of data which reduces utilization.
6. Optimize the amount of Spark partitions
– Too much or too little spark partitions could mean some executors are idle or
scheduling overhead.
7. Use mapPartitions() over map()
– Using mapPartitions provides initialization for many RDD elements rather once
per RDD element.
8. Check for memory leaks
– Unchecked memory leaks can cause a host of memory issues and slow data
processing.
9. Check for bottlenecks
– Bottlenecks can occur in any stage of our algorithm which can often slow data
processing.
10. Improve queries
– Instead of returning every row or column we should only return the ones we are
looking for.
