[taligent]

Why do you need me ? I am just that crazy troll who believes that no database is perfect for all use cases and that making blind generalisations like "all NoSQL is bad" is stupid. Especially since MongoDB is as different from Cassandra or Riak as it is from Spanner as that is from the other 120+ odd NoSQL databases.

And if benchmarks impress you I suggest you take a look at Cassandra sometime.

http://techblog.netflix.com/2011/11/benchmarking-cassandra-s...

[jchrisa]

I keep hearing Cassandra touted as fast, but I just don't see it. Instacluster brags about 17ms at the 99th percentile being fast, but where I come from that gets you thrown out of the running.[1] The Netflix benchmark linked above sees 10,000 ops/sec/node. Again, not impressive.

Skip to slide 22 of this comparative benchmark with Couchbase Server[2] and you'll see that <5ms latencies at 99% is totally achievable. We can't take too much credit for these stats, most of the blame goes to the fact we use memcached at the front end, and it's been optimized for consistent low-latency for a long time.

I don't want to make this a speed contest, as there are a lot of good things about Cassandra. It just strikes me as odd that people sell Cassandra based on low latency. If you want consistent high-performance, don't use a database that runs on a language virtual machine. There are lots of great reasons to use Cassandra, standout benchmarks aren't one of them.

[1] https://www.instaclustr.com/solution#tco [2] http://www.slideshare.net/renatko/couchbase-performance-benc...

[3amOpsGuy]

What else would you say is good about Cassandra?

I find the datamodel inflexible and often hard to work with for real world use cases. I think the tooling is immature and the hector API leaves me feeling depressed (pycassa and astyanx are quite a lot better).

All that said, I find cassandra to be fast, incredibly fast. 99th percentile on our 9Tb time series store is around 20ms on below par hardware.

Another thing i like is we've never had downtime, not even scheduled. We've had nodes fail, we've had datacentres isolated during disaster recovery tests, yet our cluster has continued on regardless (2 x sub clusters per region). Whatever way you stack it up, that's impressive. Remember with cassandra there's no load balancers or other shenanigans involved.

Garbage collection / virtual machine concerns are a red herring due to the mechanics of a cassandra query (certainly for CL < RF)

[jchrisa]

I'm glad you are happy with 20ms, to me it is slow...

Probably the best thing about Cassandra is that it is written in Java, so it is easy to fork and add custom behavior. Of course this is orthogonal to performance.

[neya]

I only tagged you because you seem to have experience running Mongodb in production (as claimed by you in one of the previous threads) and I honestly wanted your feedback about this. Also I was curious to see if you would evaluate this nifty PostgreSQL feature :)