On Sun, Sep 2, 2018 at 2:30 AM, James Smith <js4625074@xxxxxxxxx> wrote: > Hi, > > When pg goto peered state, it will wait the failed osd been mark out. > Reading(without modify) on peered pg will be blocked. The reason of > this restrict is for consistency or something else? Have anyone ever > talked about letting go of this restriction? When a PG isn't active it has no ability to do anything with clients. If it's peered, it *might* possess enough information to safely serve reads, but changing the code to do so would be a big job and would probably involve protocol changes these days. (And it's not guaranteed to be able to do so, anyway.) How are you seeing this scenario? Is it theoretical or is there some case where you've got a lot of PGs peered for a long time without serving IO? -Greg
