FogStore

Abstract

We design Fogstore, a key-value store for event-based systems, that exploits the concept of relevance to guarantee low-latency access to relevant data with strong consistency guarantees, while providing tolerance from geographically correlated failures. Distributed event-based processing pipelines are envisioned to utilize the resources of densely geo-distributed infrastructures for low-latency responses - enabling real-time applications. Increasing complexity of such applications results in higher dependence on state, which has driven the incorporation of state-management as a core functionality of contemporary stream processing engines a la Apache Flink and Samza. Processing components executing under the same context (like location) often produce information that may be relevant to others, thereby necessitating shared state and an out-of-band globally-accessible data-store. Efficient access to application state is critical for overall performance, thus centralized data-stores are not a viable option due to the high-latency of network traversals. On the other hand, a highly geo-distributed datastore with low-latency implemented with current key-value stores would necessitate degrading client expectation of consistency as per the PACELC theorem. In this paper we exploit the notion of contextual relevance of events (data) in situation-awareness applications - and offer differential consistency guarantees for clients based on their context. We highlight important systems concerns that may arise with a highly geo-distributed system and show how Fogstore's design tackles them. We present, in detail, a prototype implementation of Fogstore's mechanisms on Apache Cassandra and a performance evaluation. Our evaluations show that Fogstore is able to achieve the throughput of eventually consistent configurations while serving data with strong consistency to the contextually relevant clients.