News for hackers. Stuff that matters.
Re:Copy-on-write B-tree finally beaten. by justincormack on Monday April 11, 2011 @06:00PM (#2434606)
Acunu (http://acunu.com) are a UK startup and have an implementation of Cassandra, it supports versions presumably based on this.
Re:Copy-on-write B-tree finally beaten. by dchest on Monday April 11, 2011 @06:48PM (#2434788)
Correct, here's a post by one of the paper authors: http://www.acunu.com/2011/03/big-dictionaries-ii-versioning/
Re:Copy-on-write B-tree finally beaten. by cbetz on Monday April 11, 2011 @08:19PM (#2435126)
this is actually very exciting. I decided to explore the site a little and found this quote:

"most ambitious storage project undertaken since ZFS" (http://www.acunu.com/2011/03/why-is-acunu-in-kernel/)

it looks like they put a key/value store in the kernel and they came up with a new userspace API for it. i can see also see how getting something like this into the mainline kernel is going to be a big uphill battle, but it might actually be a really big win.

Re:Copy-on-write B-tree finally beaten. by tdmackey on Tuesday April 12, 2011 @12:47AM (#2435940)
Part of my job is working on a in-kernel key/value store for a data center network operating system. The problem it is the context switching between user and kernel space kills your performance if you're targeting sub-millisecond read/writes. That may not matter for something like acunu when you factor in network latency but when you're using the database as part of the packet path it does. In addition, under a high system load your user space process has a high likelihood of getting scheduled out at the ioctl call which makes latency even worse. Although being in the kernel allows you a little leeway in terms durability constraints and all that because if you screw up the entire system comes crashing down anyway.

It will never be in the mainline kernel. Also, although I haven't actually looked at what they did yet, I assume they're just loading a regular old kernel module instead of actually really messing with a lot of the mainline code.