February 17th – February 22nd 2013, Dagstuhl Seminar 13081
Consistency In Distributed Systems
Kapil Vaswani (Microsoft Research India - Bangalore, IN)
1 / 2 >
For support, please contact
In distributed systems, there exists a fundamental trade-off between data consistency, availability, and the ability to tolerate failures. This trade-off has significant implications on the design of the entire distributed computing infrastructure such as storage systems, compilers and runtimes, application development frameworks and programming languages. Unfortunately, it also has significant, and poorly understood, implications for the designers and developers of end applications. As distributed computing become mainstream, we need to enable programmers who are not experts to build and understand distributed applications.
A seminar on "Consistency in Distributed Systems" was held from 18th to 22nd, February, 2013 at Dagstuhl. This seminar brought together researchers and practitioners in the areas of distributed systems, programming languages, databases and concurrent programming, to make progress towards the above mentioned goal. Specifically, the aim was to understand lessons learnt in building scalable and correct distributed systems, the design patterns that have emerged, and explore opportunities for distilling these into programming methodologies, programming tools, and languages to help make distributed computing easier and more accessible.
We may classify current approaches to deal with the challenges of building distributed applications into the following three categories:
- Strong Consistency and Transactions: Strong consistency means that shared state behaves like on a centralised system, and programs (and users) cannot observe any anomalies caused by concurrent execution, distribution, or failures. From a correctness perspective, this is a most desirable property. For instance, a database management system protects the integrity of shared state with transactions, which provide the so-called ACID guarantees: atomicity (all-or-nothing), consistency (no transaction in isolation violates database integrity), isolation (intermediate states of a transaction cannot be observed by another one), and durability (a transaction's effects are visible to all later ones).
- Weak Consistency: Unfortunately strong consistency severely impacts performance and availability. As applications executing in the cloud serve larger workloads, providing the abstraction of a single shared state becomes increasingly difficult. Scaling requires idioms such as replication and partitioning, for which strongly-consistent protocols such as 2-Phase Commit are expensive and hard to scale. Thus, contemporary cloud-based storage systems, such as Amazon's Dynamo or Windows Azure Tables, provide only provide weak forms of consistency (such as eventual consistency) across replicas or partitions. Weakly consistent systems permit anomalous reads, which complicates reasoning about correctness. For example, application designers must now ascertain if the application can tolerate stale reads and/or delayed updates. More parallelism allows better performance at lower cost, but at the cost of high complexity for the application programmer.
- Principled Approaches to Consistency: A number of approaches and tools have been developed for reasoning about concurrently-accessed shared mutable data. The concept of linearizability has become the central correctness notion for concurrent data structures and libraries. This has led to significant advances in verification, testing and debugging methodologies and tools. Transactional memory provides a higher-level, less error-prone programming paradigm.
- Principles for weak consistency: More recently, a number of principles have emerged for dealing with weak consistency. For example, if all operations in a program are monotonic, strong correctness guarantees can be provided without the use of expensive global synchronization. Similarly, certain data structures such as sets and sequences can be replicated in a correct way without synchronisation.
These developments illustrate the benefits of cross-fertilization of ideas between these different communities, focused on the topic of concurrency. We believe that such principled approaches will become increasingly critical to the design of scalable and correct distributed applications. The time is ripe for the development of new ideas by cross-fertilisation between the different research communities.
It is crucial for researchers from different communities working in this same space to meet and share ideas about what they believe are the right approaches to address these issues. The questions posed for the seminar include:
- Application writers are constantly having to make trade-offs between consistency and scalability. What kinds of tools and methodologies can we provide to help this decision? How does one understand the implications of a design choice?
- Weakly consistent systems are hard to design, test and debug. Do existing testing and debugging tools suffice for identifying and isolating bugs due to weak consistency?
- Can we formalize commonly desired (generic) correctness (or performance) properties?
- Can we build verification or testing tools to check that systems have these desired correctness properties?
- How do applications achieve the required properties, while ensuring adequate performance, in practice? What design patterns and idioms work well?
- To what degree can these properties be guaranteed by the platform (programming language, libraries, and runtime system)? What are the performance tradeoffs (when one moves the responsibility for correctness between the platform and application)?
In order to ensure a common understanding between the different research communities that the workshop brings together, the seminar started with a few tutorials from the perspective of each community. Other presentations presented a specific piece of research or a research question. Participants brain-stormed on a specific issue during each of the two break-out sessions.
This report is the compilation of notes taken by several note-takers, rotating at each session. The majority of the participants served as scribe for some session.
It will be helpful to refer to the abstracts and slides of the different presentations, which are available at http://www.dagstuhl.de/mat/index.en.phtml?13081.
- Daniel J. Abadi. Consistency tradeoffs in modern distributed database system design. Computer, 45(2):37–42, February 2012.
- Eric Brewer. CAP twelve years later: How the "rules" have changed. IEEE Computer, 45(2):23–29, February 2012.
- Maurice Herlihy and Jeannette Wing. Linearizability: a correcteness condition for concurrent objects. ACM Transactions on Programming Languages and Systems, 12(3):463–492, July 1990.
- Maurice Herlihy and J. Eliot B. Moss. Transactional memory: Architectural support for lock-free data structures. In Int. Conf. on Comp. Arch. (ISCA), pages 289–300, San Diego CA, USA, May 1993.
Creative Commons BY 3.0 Unported license
Bettina Kemme and Ganesan Ramalingam and André Schiper and Marc Shapiro and Kapil Vaswani
- Data Bases / Information Retrieval
- Programming Languages / Compiler
- Semantics / Formal Methods
- Distributed Computing
- Weak and Strong Consistency