Excerpts from Jon Foster's message of 2016-10-14 09:33:59 -0700:

I have a DB scenario that is very write intensive. Essentially its a large scale hit counter of sorts. Currently we're running on a single 12core server with 6 SSDs in a RAID6 array. But we're looking for a way to scale out write volume by adding more servers, hopefully as conveniently as I might add Apache servers to a website. We see two servers laboring so we add a third to the mix and so on.

I'm still looking at the various technologies available to me and was wondering if someone out there had some suggestions on this front. Although Galera Cluster says it scales for write loads too it also says that all servers write the same data and make sure its committed by the time the transaction is completed. So it seems like write performance would never scale because all servers are doing the same write.

Maybe I'm missing something.

Anyhow some pointers would be appreciated.

If you're overwhelming one server (you haven't even begun to scale up btw) you will have to shard. However, before you do that.. consider how cheap 24 core servers with 10x FusionIO or similar SSDs would be versus the cost in complexity of sharding.

Maybe take a look at Vitess, which helps shard things but keeps all the good stuff you like about MariaDB/MySQL:

http://vitess.io/

On 15/10/16 03:33, Jon Foster wrote:

I have a DB scenario that is very write intensive. Essentially its a large scale hit counter of sorts. Currently we're running on a single 12core server with 6 SSDs in a RAID6 array.

Not using RAID6 would be a good start.

But we're looking for a way to scale out write volume by adding more servers, hopefully as conveniently as I might add Apache servers to a website. We see two servers laboring so we add a third to the mix and so on.

I'm still looking at the various technologies available to me and was wondering if someone out there had some suggestions on this front. Although Galera Cluster says it scales for write loads too

Where? This is dubious for the reasons below (which are the same as why raid6 is poor).

it also says that all servers write the same data

(well - *have* the same data rather than have it written)

and make sure its committed by the time the transaction is completed. So it seems like write performance would never scale because all servers are doing the same write.

Maybe I'm missing something.

I don't think so. You seem to have that worked out.

Anyhow some pointers would be appreciated.

MyRocks will be interesting to keep an eye on.

Answering Justin's question and details to what component in hardware/MariaDB is the limiting factor is probably critical before you go much further. Perhaps even getting professional help.