365 karma · joined August 24, 2007
You don't need to think about servers at SaaS (Software as a Service. e.g, Google Docs/Apps) and PaaS (Platform as a Service. e.g. Google Apple Engine, Windows Azure; Amazon S3, Elastic MapReduce etc.) level, but you do have (virtualized) servers when you need IaaS (Infrastructure as a Service. e.g. Amazon EC2, VMware infrastructure) where you need to put existing/legacy applications that have clear dependencies on explicit servers in the cloud.
Cheap commodity disks however have an annualized failure rate of 4% in Google's datacenter (according to their disk analysis paper.)
NoSQL is ironic word in that both Hive and Pig are gonna introduce an ANSI SQL dialect for scalable data queries :)
The NoSQL movement is really a NoBlindJoinsViaRDBMS movement, but the latter is less catchy.
It's quite strange that they didn't reference Bigtable paper at all, while saying "to the best of our knowledge, there exists no published deployment of a parallel database with nodes numbering into the thousands". Google had a dozen bigtable clusters with more than 500 nodes and at least one cluster with a few thousand nodes (for the main crawl db), more than 3 years ago.
http://www.stanford.edu/~engler/
Coverity is a startup from such research. The new Klee paper is a good read as well.
Engler (whose students founded Coverity) et al. had an OSDI "best paper" last year on using static analysis (with constraint solvers etc.) to automatically generate test cases, which actually beats hand written test cases for glibc with years of development.
While I like the the shared/synced document object model, I'm not sure about the value added via a new/unproven protocol stack just for the per character gimicks.
"Real time" is overrated and not needed in many cases.
I'd like to see median and 99 percentile latency for any throughput numbers.
Give it up. "Doing" instead.
MurmurHash2 is the winner in all cases I've tested.
Regarding your 2nd point, latency != 1/throughput when you have parallel queries.
Another way for front-page ordering would be (total karma * exp(-beta * age)), you can tweak beta for time decay.
works and more up to date (updated on May 14, 2009)
Looks pretty good.
The only way to overcome that is to do more rewrites :)