On 10/30/2012 5:15 PM, paulruizendaal wrote:

>
>
> > VoltDB takes that approach.
>

I think we're actually in agreement. A transaction in Volt is a single
round trip from the client to server and back, and once started, runs to
completion.

Incidentally, to get particularly good TPCC numbers (un-audited, of
course) Volt takes certain liberties to check for a possible unique
constraint violation so they can optimize out transaction backout.

>
> No. VoltDB runs each transaction to completion. The design pattern I'm
> talking about runs non-blocking tasklets concurrently. Non-blocking as
> in yielding the thread when the time slot is up, or when encountering
> a dependency on another tasklet. Translating that pattern to databases
> requires multiversion non-blocking datastructures (i.e. lockfree with
> dependencies handled by the tasklet scheduler). Although different,
> NuoDB is much closer to the pattern than VoltDB.
>
> Actually, I think some of the NuoDB ideas could have very useful
> application in the distributed version of the erlang VM.
>
>

[Non-text portions of this message have been removed]