Thank you. I was not aware of the terminology surrounding this technique or the current work on it.
On Fri, Feb 15, 2013 at 10:34 AM, J. Andrew Rogers <and...@jarbox.org>wrote: > > On Feb 15, 2013, at 1:41 AM, David Barbour <dmbarb...@gmail.com> wrote: > > > > A technique I've seen used (albeit, only once or twice) is to batch > updates from multiple connections, and run batches at 20+Hz while there is > demand for them. This only requires one connection from the server to its > database, and allows you to respond and close each HTTP connection in a > predictable amount of time, and allows you to ensure the updates are > committed before replying. The limitation of this technique is that you > can't have ad-hoc updates, since those would be too difficult to batch > consistently and sensibly. > > > > It's a bit specialized, but is able to handle a lot of updates. > > > Distributed database engines that do this natively are becoming popular. > They are usually referred to as "short transaction processing" databases > and workloads. By forcing all open transactions to commit or rollback > several times a second it is possible to guarantee strict serializability > in a distributed database if you can package your transactions within that > window. There are a few variations on the idea. > > Traditional transaction processing databases allowed transactions to be > open for an unbounded period of time. There are stories passed around among > DBAs of zombie transactions that had been open for months or years. For > large-scale distributed databases, the cost of coordinating the transaction > state is prohibitive and limits the total transaction throughput. Some > clever person noticed that if you can strictly bound the transaction length > then it makes the distributed transaction problem much simpler. > Conveniently, many transaction processing applications can be made to fit > within the STP database model. > > An STP database engine can support orders of magnitude more transaction > throughput than a conventional OLTP database. The tradeoff is that the > transaction latency is higher. 10-100 milliseconds seems to be the zone > where latency is acceptable but you can still fit most transactions into > that window. > > > _______________________________________________ > fonc mailing list > fonc@vpri.org > http://vpri.org/mailman/listinfo/fonc > -- bringing s-words to a pen fight
_______________________________________________ fonc mailing list fonc@vpri.org http://vpri.org/mailman/listinfo/fonc