[stock_toaster]

  > - poor single node performance

I am in the process of evaluating riak for an project, and have run across this too. I am not sure if it is specific to the eleveldb backend (I just used the default values so far), or just a bad config on my part. Even when I relaxed consistency down to r=1,w=1 and used the protobuf interface, riak was still performing rather slowly (compared to some other things I am testing) in single node performance.

[Ixiaus]

SERIOUSLY! Riak is NOT MEANT for single-node clusters!!! It gets faster THE MORE NODES YOU ADD.

The minimum recommended number of nodes is 3; the "ideal" minimum number of nodes is 5.

[stock_toaster]

yes, with n_val=3 (default) 3 nodes is the minimum. I was merely testing (actually had 3 riak instances on a single box as a "cluster") but with r,w=1 and n_val=1 it was about the same (slow), which was disconcerting. cpu was very high in one instance (others were nearly idle, and IO was low. I did not try dw=0 though (if that would even work). Might be worth a try.

I am going to try bitcask too, to see if eleveldb was the culprit and could be tuned.

[Ixiaus]

Your test isn't valid if it was all on one box - the disk I/O is going to be really nasty with that. Your test would be better on three different machines.

[stock_toaster]

The disk io was pretty low. It is serial get/update/delete (as I mentioned earlier). I even tried moving the data dirs to ramdisk (was plenty of ram), but it didn't help much at all. i/o was certainly not the problem.

It very well could be a configuration issue. I am still trying a few things to see if I can get an improvement.

[pharkmillups]

If you haven't done so already, I would highly-recommend taking a look at Basho Bench, our open source benchmarking tool for Riak.

http://wiki.basho.com/Benchmarking.html https://github.com/basho/basho_bench

[stock_toaster]

thanks. Will do!

[Ixiaus]

Interesting. I can vouch for the speed of leveldb on a five node cluster - I'm using it in our startup on 5 dedicated machines with around 12GB of RAM, running on FreeBSD. Riak version is 1.1.4; Riak 1.2 is supposed to have a lot of leveldb refinements.

[stock_toaster]

Good to hear. I will keep poking at things. I probably have something misconfigured or an issue with how I am using the python client.

Thanks.

[drsintoma]

What did you find slow? writes? retrieving by key? map-reduce? From what I've heard erlang map-reduce is much performant than the javascript one.

[stock_toaster]

It was a sequence simple operations (simple as in no 2i or mapred) of read an existing key A, check if a nonexistent key exists B, write a previously nonexistent key C, delete an existing key D, overwrite an existing key A.

All performed in serial over protobuf (python riak lib).

[Ixiaus]

LevelDB is known to have heavily degrading performance when you ask for a key that doesn't exist.