replica behavior

[travisghansen]

With this configuration what should the behavior be?

apiVersion: hyperspike.io/v1
kind: Valkey
  name: valkey
spec:
  nodes: 1
  replicas: 2

I have 3 pods (0-2), all of them show role:master and I don't see any replication happening among members.

[dmolik]

So there is some history here, but suffice to say, this is a bug that I was going to fix with or soon after I implemented zero downtime upgrades, can you hang tight for a few days?

[travisghansen]

Of course! Just wasn't sure what to expect. Been a while since I dug into the depths of redis semantics.

Assuming replication is working as desired, how does the Service ensure writes are sent to the master? Or does the protocol have all the brains to understand that and make sure that after initial connection reads/writes go where they are supposed to?

[dmolik]

okay so, I'm happy to get on a video call and explain this, but it all starts with the concept of SLOTS, which is like a logical partition a Redis now Valkey Cluster gets about 16 and a half thousand slots, a Master Node is the primary owner of some number of slots, additionally all slots must be 'owned' by some combination of masters. All that said a slave node is a read only replica of a master node. In the CRD, nodes are master nodes and replicas are slaves of a master, so in theory you should have nodes x replicas pods. Funny note, in theory you could have 16k master nodes.

When you connect to a Valkey cluster, the node will check if the slot is 'owned' locally and will serve the value, if not the client gets a redirect message, and will then connect to the node with the next slot, in Valkey Clusters with many slave nodes there is a concept of read-only operations being redirected to slave nodes.

I'm no expert when it comes to Valkey networking topology so my suggestion is to checkout the docs: https://valkey.io/docs/ or ask a question in the discussions: https://github.com/orgs/valkey-io/discussions

[travisghansen]

OK that sounds sane. Thanks for the refresher.

I have 1 small change I would make to the above statement: number of pods will be (<nodes> + (<nodes> * <replicas>)) vs nodes x replicas.

Thanks!

[dmolik]

Correct, thank you.