The lack of horizontal scalability for compute-streaming nodes is due to the fact that each source's number of active source actors won't exceed the number of splits in the upstream data source. Since your tables are singleton, only one source actor on one node will ingest data to maintain data consistency. The scheduling strategy might be causing all active actors to be on the same node. For faster failover, you should use risectl to scale the horizon and unregister the compute node gracefully before killing it. This process is currently manual but will be automated in future releases. The long failover time you're experiencing is likely due to the meta node waiting to see if the killed node comes back before reassigning tasks, which is a safety feature to handle unexpected failures.