On Wed, Jan 28, 2009 at 3:03 AM, Noah Slater <[email protected]> wrote: > On Wed, Jan 28, 2009 at 09:30:19PM +1030, Antony Blakey wrote: >> >> On 28/01/2009, at 9:10 PM, Jan Lehnardt wrote: >> >>> >>> On 28 Jan 2009, at 11:31, Brian Candler wrote: >>> >>>> BTW, I do think the atomic nature of bulk_docs is useful and should >>>> be kept, >>>> as it's the only way to get "transaction" semantics at the moment. >>> >>> We won't be able to guarantee transactions in a multi-node setup. >> >> And there's a universe of single-node applications. > > I would prefer a predictable interface over single-node special-casing. >
We had a book comment mentioning the usefulness of giving application-level control over whether to run a particular computation on a cluster, or on a single node, for the purpose of taking advantage of the different side of the CAP theorem. It made me rethink the idea that we should eschew single-node transactions. For instance, an application could replicate a set of docs to a single node, run a transaction across them, and then replicate back. Or you might put some types of data on a big fast machine, but spread most data around commodity cluster machines. It's worth thinking about - there are things that get much harder to do, without multi-doc transactions... Chris -- Chris Anderson http://jchris.mfdz.com
