Both GFS and Bigtable make the same core design choice – to have a single master. What are the repercussions of a failure of this single master in each case?

What will be an ideal response?

This decision was made, particularly in GFS, to simplify the design process and allow Google to get a production system in operation quickly. Although this is a single point of failure, failures can be detected and the master restored. This is supported by the fact that key metadata is stored in a persistent operations log and this is replicated on several machines, enabling relatively rapid recovery. Other data can be restored by polling the chunkservers. There will be a short delay from failure occurring to being detected through to the master being restored and this will disable operations that require the master, but not ongoing transfers as they bypass the master due to the separation of control and data (note that in Bigtable, fewer operations go through the master).