I presented most of the Gray "Dangers of Replication" paper [GHOS96]. I don't believe the arguments in detail. But I do believe the basic contention: a system that attempts to provide the strong consistency guarantee that is traditional in database systems -- one-copy serializability (1SR) -- is unlikely to scale acceptably.
5 September 00:
Please look at [WPS+99], [WPS+00] and [WPS99]. These are pretty easy overview papers; I don't think we want to present them in class, but they provide a bit of common background.
I would like us to spend a few lectures on techniques that guarantee 1SR without imposing semantic restrictions on transactions. (A syntactic restriction -- e.g. "all the data updated by a transaction must be mastered at the same node" is acceptable here; a semantic restriction like "the following operations must commute" is not. So "syntactic" means, roughly, we get to analyze dependencies between operations that are classified as reads or writes, but are otherwise uninterpreted.)
This may be a slightly unnatural order, but I'd like to start discussion of some simple lazy schemes (as opposed to eager schemes, in the classification from [GHOS96]). I'll make some general observations from Oracle's lazy replication system, then move to [CRR96], which uses a combination of syntactic restrictions and data placement to guarantee 1SR in a lazy-master system. The approach in this paper is quite restrictive
There is a cluster of recent papers about lazy group protocols, leading up to [BKR+99]. I'll continue this thread on Thursday, and would like a volunteer for next week.
6 September 00
Here are some long-promised proposals for topics/papers to present:
Eager schemes using group communication protocols.
The argument is that the traditional 2PC distributed commitment protocol is inefficient at least for the special case of a fully-replicated database, and an optimistic protocol using some variant of atomic broadcast will give better performance. Relevant papers are [AAE+96] ("seminal"), [PGS98b] (probably the best single paper to present, and there is a lot of related stuff by the same authors in the bibliography). The [KA98, KA00] sequence also might be worth looking at.
11 September 00
Okay, since I'm still so far behind in the outline, I'm going to start the "exploiting atomic broadcast" subject myself. Tuesday will cover [PGS98b] or, equivalently, [P99] through Section 3.4.