On the Consistability of Storage Systems

Modern Internet-scale storage systems typically provide different consistency guarantees under different operating conditions (i.e., failure scenarios). Under the faultfree condition, the system may provide strong guarantees such as atomic consistency. However, under failure conditions (e.g., network partitioning), the system often sacrifices consistency in exchange for higher availability. For example, under network partitioning, the system may only provide eventual consistency. After all, by Brewer’s famous theorem, if one has to live with network partitioning, then one has to sacrifice either consistency or availability. The current practice is to only describe the weakest consistency provided by a storage system, although for a considerable amount of time, the system may operate under good conditions and provide stronger consistency. This makes the specification imprecise and focuses exclusively on the worst-case. Such specifications make it hard to compare two systems in detail. To address this problem, we have introduced the concept of consistability [2]. Inspired by and analogous to the notion of performability, consistability tries to capture the fact that a storage system provides different kinds of consistency under different operating conditions. At a high level, the consistability of a system describes what kinds of consistency are achieved by the system under each operating condition. If one can come up with the portion of time each operating condition occurs (by measuring, modeling, or estimating), then one can calculate how often a system achieves a certain consistency. For example, suppose a system provides atomic consistency under fault-free conditions, which is estimated to be around 80% of the time, and provides regular consistency under all faulty conditions, which is estimated to be around 20% of the time. Then the system provides atomic consistency 80% of the time, but regular consistency 100% of the time. (It remains an open problem how to detect the transitions between different consistencies and how to reason about the system’s behavior during a transition.) Consistability makes the specification of a system more precise and fa-