# `TiKV` Client
+ `retry/backoff`
  + Retry mechanism is a redundant design in the perspective of time (replicas are from space perspective)
  + Higher level clients could only observe the latency increasing
+ `Region` cache is lazily updated, re-acquired from `PD` in case there are errors
  + `Regions` tends to be numerous and frequently updated
  + Some shortcuts are introduced
    + **`RegionCache`** stores all `peers` at the same time
    + `TiKV` returns **`RegionError`** along with latest `Region` information
+ Detect network status with `gRPC keepalive`
------------------------------------------------------------
Why 20s?

`leader` election timeout = 10s

`tidb-server` has a `TiKV` client