# `KV` Scan vs. `Raft Learner`
+ Old `PR POC` used `Raft` `learner`
+ Proven to be bad for backup:
  + Complicated implementation
  + `Learner` node is bottle neck, cannot easily scale out
  + Very slow (speed similar to **`mydumper`** @ `10 GB/min`)
+ Iteration uses more resources
  + Affect latency for "high thinking" workload
  + Affect latency + `OPS` for "high working" workload
------------------------------------------------------------
"high thinking" workload: Client side spend the most of the thinking time, less workload on the server side per request.

"high working" workload: Client side spend less time, such as batch processing, larger workload on the server side per request.