2 ms·
If its a problem with writes, then why not put the logging tables onto a different disk? Surely that would fix write contention?
by chris_wot 15y ago
If its a problem with writes, then why not put the logging tables onto a different disk? Surely that would fix write contention?
- nl 15y agoCassandra can do over a million writes per second[1] because of a combination of its horizontal scaling and eventual consistency. I expect HBase could scale similarly. By comparison, people write about scaling MySQL to 25000 writes per second[2] It's very hard to get a single server to scale multiple orders of magnitude higher (even with FusionIO or something). Getting around that single master server for writes is possible, but either complicated or expensive. Even then I doubt you will get to 1 million writes per second without sharding, and if you are going to shard you might as well go to a NoSQL solution built for the problem. [1] http://techblog.netflix.com/2011/11/benchmarking-cassandra-scalability-on.html http://techblog.netflix.com/2011/11/benchmarking-cassandra-s... [2] http://www.spoutingshite.com/2011/01/27/scaling-mysql-to-25000-inserts-per-second/ http://www.spoutingshite.com/2011/01/27/scaling-mysql-to-250...
- chris_wot 15y agoThat is indeed extremely impressive! I hope I didn't come across as meaning that NoSQL solutions aren't appropriate - it's very hard for me to say this anything about these solutions, as I've got next to no experience with this technology. I am very interested in the first article though. It mentions number of instances in an EC2 availability zone - what does this mean? Is this similar to a clustering/load balanced solutions in traditional RDBMSes? The automated tooling that Netflix has developed lets us quickly deploy large scale Cassandra clusters, in this case a few clicks on a web page and about an hour to go from nothing to a very large Cassandra cluster consisting of 288 medium sized instances, with 96 instances in each of three EC2 availability zones in the US-East region. Using an additional 60 instances as clients running the stress program we ran a workload of 1.1 million client writes per second. Edit: Argh! Sorry, I just didn't read far enough before I posted. They seem to be being powered by some fairly heavy duty hardware. Is there a comparitive study using Oracle RAC or something like this?