Several cloud providers provide storage in many data centers globally, and customers can use simple PUTs and GETs to store and retrieve data without dealing with the complexities of the storage infrastructure. However, in reality, every storage system leaves replication across data centers to the application, and although replication across all data centers provides low latency, it is expensive.
SPANStore: Cost-Effective Geo-Replicated Storage Spanning Multiple Cloud Services is the first system that tries to solve this automatically, by minimizing the cost incurred by latency-sensitive application providers.
What is SPANStore?
SPANStore is a key-value store that provides a unified view of storage services present in several geographically distributed data centers. It spans data centers across multiple cloud providers and determines where to replicate every object and how to perform this replication. Finally, it reduces cost by minimizing the computing resources necessary to offer a global view of storage.
SPANStore uses multiple cloud providers to offer lower GET/PUT latencies. Also, this allows for lower cost by exploiting price discrepancies across providers to meet latency SLOs.
PMan determines the replication policies in SPANStore. It takes a three-part specification as input:
- a characterization of SPANStore’s deployment
- the application’s latency, fault tolerance, and consistency requirements
- a specification of the application’s workload as inputs
and in return, it provides the set of data centers that maintain copies of all objects with that access set, and, at each data center in the access set, which of these copies SPANStore should read from and write to when an application VM issues a GET or PUT.
SPANStore can trade-off costs for storage, PUT/GET requests, and network transfers if the application requires only eventual consistency. SPANStore replicates objects at fewer data centers to reduce storage costs and PUT request costs. PMan address this trade-off between storage, networking, and PUT/GET request costs using a replication policy as a mixed integer program.
They rely on quorum consistency for strong consistency. They use asymmetric quorum sets and require an intersection of at least data centers with the PUT replica set of every other data center in the access set.
The main goal of this paper seems to be to minimize costs of deploying an application by trading off replication, network costs, storage costs, and latency. It seems that there is still a huge burden on developer to provide the correct inputs to PMan so that PMan can provide the best replication policy. This doesn’t seem to reduce the complexity involved. A lot of the paper relies on the objective function, and there are not many new distributed system concepts.