@Zoran Radovanovic since you keep saying cluster, I guess you’re running an instance of this here for Redis OSS cluster:
https://redis.io/docs/management/scaling/
That would normally translate into a strategy of sharding your keyspace in an elegant way to distribute your data across nodes. A cluster-aware client will discover that topology automatically and communicate with the right nodes to access the data they contain. In that scenario the introduction of a LB makes almost no sense.
But what you’re stating as a test scenario makes me believe you’re running a Sentinel setup - which is no cluster. I am confused 🙂