Previously all queries execute on the Redis primary servers, since that gives us transactional safety and guaranteed data synchronisation.

In a heavily-loaded cluster, moving read queries out to the replicas can be an acceptable performance tradeoff. Any query which doesn't modify data will be directed to replicas (if any exist - the primary is used otherwise). Writes still go to the primary.

https://redis.io/docs/latest/commands/readonly/

This is disabled by default - you can opt in for the entire process by setting the transport_use_read_replica configuration value, e.g. export MYRIAD_TRANSPORT_USE_READ_REPLICA=1.