This paper makes the point that implementing Paxos is difficult to do, even though the concept is not all that complex in theory. Many systems claim to use Paxos, but their precise implementation details are often left undescribed. The Viewstamped Replication paper makes an effort to explain how to use Paxos in the real world, but uses distributed transactions (that add more complexity than all applications necessarily require) and doesn’t really address group membership changes.
The paper then goes on to describe the protocol using a C/++ interface and state-machine replication (a deterministic service that accepts requests & produces replies). The level of detail of the implementation is very thorough and certainly sets the paper apart from the work on Paxos that had been previously published. As far as trade-offs, the implementation can be made more efficient by broadcasting requests to backups. Likewise, there are hardware tradeoffs: the system requires at least three full replicas to survive a failure (so that the two remaining ones constitute a majority). The author proposes having the third machine possibly act as a “witness” that can participate to help replicas recover after a failure or partition.
The potential influence of this paper is hard to gauge: given the specificity of the implementation, the paper should always be helpful to people trying to implement Paxos, but I’m not sure that it presents any real new ideas that will sway the direction of further research on Paxos and consensus protocols.
The paper then goes on to describe the protocol using a C/++ interface and state-machine replication (a deterministic service that accepts requests & produces replies). The level of detail of the implementation is very thorough and certainly sets the paper apart from the work on Paxos that had been previously published. As far as trade-offs, the implementation can be made more efficient by broadcasting requests to backups. Likewise, there are hardware tradeoffs: the system requires at least three full replicas to survive a failure (so that the two remaining ones constitute a majority). The author proposes having the third machine possibly act as a “witness” that can participate to help replicas recover after a failure or partition.
The potential influence of this paper is hard to gauge: given the specificity of the implementation, the paper should always be helpful to people trying to implement Paxos, but I’m not sure that it presents any real new ideas that will sway the direction of further research on Paxos and consensus protocols.
No comments:
Post a Comment