Hacker Timesnew | past | comments | ask | show | jobs | submitlogin

Author somewhat intelligently argues that Dynamo sacrifices more consistency than it gains availability. Part of his argument is the difficulty of truly leveraging partitioning to gain availability, mostly the practical issues of users finding the currently available nodes in such a system. Another is issues arising from nodes re-joining a cluster without first performing a synchronization. He seems to favor the BigTable family, or simple master-slave replication (assuming that the write-load is low enough for this to be acceptable).


Except that he leaves out vector clocks and read repair which are mechanisms critical to fixing consistency issues in a dynamo style system.

Also his treatment of partition tolerance assumes that operations set up the Dynamo cluster machines incorrectly; all in the same rack.

He also completely glosses over merkle trees, assuming that the sync they help perform must be a heavy operation.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: