Dimitar Dimitrov - GUEST LECTURE
יום שלישי, 27.3.2018, 14:30
High-availability requirements in modern software triggered the
widespread adoption of eventually consistent data stores.
Unfortunately, reasoning about the correctness of programs running under
eventual consistency is a challenging task, largely due to the
possibility of very weak system behaviors.
In this talk, I will present new automated techniques that help
developers precisely identify problematic weak behaviors. Our starting
point will be to adopt conflict serializability as a correctness
criterion. We then give a new generalized version that can reason
about abstract data types under eventual consistency. This enables
precise reasoning about data store clients that use commutative
replicated data types. The generalization is founded in an abstract
notion of operation dependency, which permits us to reason about
dependencies between high-level operations.
We show how to compute dependencies in practice by using two algebraic
properties, namely, commutativity and absorption. This results in an
effective dynamic serializability analysis based on cycle detection. In
addition, we briefly discuss a static serializability analysis approach,
but for a slightly stronger consistency property - causal consistency.
Dimitar Dimitrov is a PhD student at ETH Zurich working on analysis
methods for concurrent and distributed programs. His research interests
span program analysis, theory of concurrency, distributed computing,