Comment by rubyrescue | Hacker News Reader

Comment by rubyrescue | Hacker News Reader

Parent Full thread

At my consulting company, Inaka, we built this originally as a side project when we were using Redis and thought we needed multi-master larger-than-RAM datasets, but then we ended up moving to CouchDB for that project.

It's currently a research project for us and we have an intern from Uppsala University working on the HA features including paxos leader election and multi-master conflict resolution. See the mutinode branch. Some people have added some additional database backends, but I don't know of any serious production users yet.

I gave a talk on it at the 2012 Erlang Factory as it's entirely Erlang-based. Here's the talk - http://vimeo.com/42630498

Benefits over Redis: Faster startup time over big datasets, pluggable backends, (eventually) Multi-master and HA (with a different approach than Redis is currently taking with Sentinel).

Drawbacks: 3x to 10x slower on all operations. We've done no work to optimize.

Have you seen Raft? It tries to be a much simpler and much more understandable alternative to Paxos.

https://ramcloud.stanford.edu/wiki/download/attachments/1137...

looks interesting, thanks!

You moved from Redis to CouchDB? That sounds odd.

Why? His data set is larger than ram, and couchdb is a really great disk based key/value store.

Last time I checked CouchDB was a document store with MapReduce queries. A very different beast from Redis.

It's really a much longer story. We use Redis and Couch together. Redis for caching, Couch for long-term storage.

Whether Couch has or doesn't have Map/Reduce is irrelevant, really. Couch Map/Reduce is a toy (just like Riak) and is not to be trusted. We avoid it as much as possible.

> Couch Map/Reduce is a toy (just like Riak) and is not to be trusted.

Really? How so?

I guess I don't see the point of a Redis clone in Erlang when it would be easier to tack use cases onto something that already exists, namely Riak. It's a key-value store, it's in Erlang, it's distributed, it's getting a full-text search and it has adjustable parameters so you can pick how you want to deal with the tyranny of the CAP theorem.

The only think Riak doesn't have is a high-speed RAM-based data structure capability. But hey, Redis does that. ;-)

> I guess I don't see the point of a Redis clone in Erlang

The only reason to do it if you already have something written that uses Redis and just want to build on that.

Or perhaps you're already committed to Erlang and want to benefit from a Redis-like system?

Reply on news.ycombinator.com