Study of skew mitigation techniques in MapReduce applications

Data skew is one of the reasons due to which MapReduce has been criticised for years. Skew occurs as a result of uneven assignment of workload to computational nodes. Many real world applications such as PageRank, CloudBurst, etc. severely suffer from the problem of skew which occurs either at map side or at reduce side in MapReduce model. Unfair task distribution in such applications shows the negative impact of skew on overall job execution and its performance. This study attempts to explore various types of skew, their causes and existing solutions for skew mitigation. The study observed that unfair task distribution in distributed environment leaves the potential parallelism unexploited. The paper also presents few applications which show the presence of skew and possible improvements.