Summary

The MapReduce paper is one of the pioneers in the field of large scale data processing. The paper describes a new programming model based on functional programming's map and reduce primitives for processing large data set, thereby providing an abstraction that hides many system-level details from the programmer. This paper describes the various error scenarios that happen in a large cluster of commodity hardware machines as well as various performance improvements that are added to the framework.

What I liked about this paper

a. The innovate concept of using pure functions (functions without any side-effects) from the functional programming world for fault tolerance (running mappers, reducers multiple times on the same data).
b. Instead on moving data from multiple machine to be processed by a function, the application master moves the mapper functions to the data nodes. This is a vibrant example of “moving code to data” and shows the tremendous performance benefits.
c. The biggest contribution of the paper is that it provides an abstraction that hides many system-level details from the programmer/data scientist. So a ML practitioner with limited knowledge of parallel programming, doesn’t have to learn the internals to implement his/her solution.

Critical comments

a. Application master is still the single point of failure.
b. Each stage spits data into GFS and reads from GFS, thereby imposing a tremendous disk/network bottleneck.
c. MapReduce API although succinct, is very limited in power. Functional programming primitives like flatMap, groupByKey, reduceByKey, etc. have to be written by the programmer from scratch.
d. Mapreduce doesn’t talk about how it can handle security- that is, work on a cluster of untrusted or possibly malicious computer systems.

Questions

a. How can we run Map reduce in a cluster on untrusted/ unreliable computers? It would be great if we come up with a service that takes some computational power/ space in user’s mobile phones to assign mapper/reducer tasks in them. What can be the challenges in such an approach?
b. In figure 3c, we see that the input, shuffle and output data rate in MBps drops below zero. How is that possible?
c. Active-active Masters: As How can we change map-reduce cluster to have multiple masters? Maybe one can be active while the other passively copies state from the main master. Or can we also have an Active-Active master scenario?