changing pool crush rule may lead to IO stop
How to reproduce:
1. create some OSDs
2. change their class to, say, "xxx"
3. create replicated crush rule referring to OSD class xxx.
4. remove OSDs created on step 1 (or change their class to other value).
5. change crush rule for a working pool to one created at step 3.
6. All IO in that pool will be stopped, although I expect its PG to stuck in clean+active+remapped state.