Support standalone Redis (non-cluster)

[zuiderkwast]

It is useful for developers if the same client can connect to a standalone node, not only a cluster. For this, we need to handle an error response form CLUSTER SLOTS and handle this as if the whole cluster is handled by a single node.

The requirement that the cluster needs to have at least two masters for 'cluster_ok' may need to be reconsidered.

[ioolkos]

@zuiderkwast is it possible to connect to a standalone node, with PR #41 ?

[zuiderkwast]

Even without #41, it's possible to connect to a single node using the ered_client module. #41 does make things a little more clear though with functions like ered:connect_client for connecting to a single instance.

It's not possible to use a standalone instance as if it were a cluster though, with the same functions, but actually I don't think we want to do this automatically. Cluster and standalone deployments are very different. Standalone mode allows SELECT and cross-slot commands, for example, which are forbidden in cluster mode.

[ioolkos]

@zuiderkwast Thanks a lot for your useful information! Currently we're using eredis in VerneMQ so we cannot connect to a Redis cluster yet. And ered's stated goal is to replace both previous libraries, isn't it. In any case, I'll try to do some testing based on your suggestions.

[zuiderkwast]

Yes, ered implements the functionality for single instance and cluster, but the API has a preference for cluster. This was done before I was the maintainer of this project. Now, I've been trying to make the API look nice for single instance too.

An idea: We use a tuple like {cluster, Pid} or {standalone, Pid} as the ered instance reference, returned by ered:connect functions. Then we can use the same functions like ered:command/N for both of them and it can dispatch to the right module based on the tuple. That's future work, if you like the idea. (ssl:connect and gen_tcp:connect return tuples like this, so it's not that strange.)

I've been busy with other work though.

[ioolkos]

@zuiderkwast I suppose the client always knows whether it's connecting to a standalone or a cluster. Your suggestion sounds great to me, in any case!

[ioolkos]

I've now written my plugin using ered:start_link/2 as main entry point. As expected, this only works for a cluster, not a standalone (standalone gets the unmapped_slots error). A question, though: with the idea proposed above, will ered be able to figure out by itself whether it connects to a standalone or a cluster?

[zuiderkwast]

A question, though: with the idea proposed above, will ered be able to figure out by itself whether it connects to a standalone or a cluster?

It would be possible to auto-detect, but I think it's better that users connect explicitly to cluster or standalone. I'm thinking ered:connect(Host, Port, Opts) for standalone and ered:connect_cluster(AddressesAndPorts, Opts) for cluster. After that, ered:command(EredRef, Command) and similar can be used for both kinds of instances.

[ioolkos]

Sounds great. That interface is perfect at least for me, as the client (the plugin) always knows whether it connects to a cluster or not.
Let me know if I can test (or do anything helpful). No worry & no hurry for anyone :)
I think it'll be a cool library to use and like the unified approach in it.