I love simple, lightweight, small, minimal tools that just do the bare minimum. Based on that, does anyone have any good recommendations for a key value store that is:

• lightweight
  
• clustered
  
• sharded (so if I have 5 instances and 100 keys, each node will roughly have 20 keys on it).
  
• easy to join nodes: as in kv-server --join somehost:1111
  

For reference, I think Consul is too heavy (and not sharded I believe).

It would be great to have a small go executable, that I can run on 10 servers, all connected up, that exposes a redis like api. Simple GET, PUT and STREAM would be great.

@prologic@twtxt.net anyone else here I can ping?

Thanks guys. Bitraft is awesome @prologic@twtxt.net but yeah, not sharded :( I did try etcd before @abucci@anthony.buc.ci but I did find it tricker to setup than Bitraft. But again, it’s not sharded :(

@prologic@twtxt.net ever done any stress testing on bitraft? In a cluster, do you know that the throughput would be? Like, PUT’s per second and GET’s per second?

@prologic@twtxt.net I’m happy to do it. Might try now actually. It was just incase you knew. I’ll post in the README if I get it working. I’m hoping redis-benchmark will work since it’s got the same api as redis.

I wonder if sharding could be implemented by:

Presumptions:

• redis can broadcast to all nodes in the cluster
  
• REPLICA_COUNT is 3
  

PUT workflow:

• a PUT get’s forwarded to REPLICA_COUNT random nodes in the cluster
  

GET workflow:

• a broadcast is made to the cluster saying “I NEED A VALUE FOR KEY ‘TEST’”
  
• all nodes that contain that value reply to the server
  
• the first response get’s forwarded to the client
  
• the other responses are discarded
  

I’m sure there would be some edges cases, like syncing.

• What if 1 of the random node’s is full and therefore only REPLICA_COUNT-1 nodes received the document
  
• This could me 2 nodes have the new value, but the 3rd has the old value
  

Maybe it could be solved by only committing once REPLICA_COUNT nodes successfully receive the message.

But in Bitraft every node contains every key + value, right? I probably wasn’t clear above, but in my idea REPLICA_COUNT would be 3 but the NODE_COUNT may be 10. So a put would go to 3 of 10 of the nodes.

Did a quick benchmark:
https://git.mills.io/prologic/bitraft/issues/58

Seems the summary benchmark of a 5node cluster on my laptop is:

GET: 1165.64 requests per second
SET: 1061.80 requests per second

Or memcached with mcrouter.