Distributed object management in large-scale distributed systems Marc Shapiro Thursday 12 December 13:00 University Paris 6, Building 41, Room 203/205 You are kindly invited to my "Habilitation à diriger les recherches" thesis defense at Université Paris 6, 4 place Jussieu, Paris 5th arrondissement (métro Jussieu). Follow directions at . There will be drinks and food after the defense. Abstract: Despite being widespread, distributed systems remain hard to understand and to program. Our research aims at easing information sharing in large scale systems, without giving up their power. The first part of the talk will provide some historical perspective. We first used a software engineering angle, structuring distributed systems as distributed shared objects called Fragmented Objects. Next, we focused on distributed referencing and garbage collection with our Stub-Scion Pair Chains mechanism. This was later generalised, as Larchant and PerDiS, to garbage collection and persistence, in replicated shared memories. Most recently we have been working on the IceCube reconciliation engine. The main part of the talk focuses on consistency of mutable replicated shared data. In optimistic replication, concurrency control is lazy and consistency is restored after the fact. We present a new theory of eventual consistency based on two simple primitives that can encode the usual relations such as transactions, conflicts, and causal dependence. We exhibit local and global correctness invariants for eventual consistency. The global invariant is quite strong and hard to ensure. Finally we show a new distributed reconciliation protocol that integrates IceCube and guarantees eventual consistency. This work was performed at INRIA Rocquencourt and Microsoft Research, Cambridge, United Kingdom.