On Tue, 28 Mar 2017, Xinze Chi (信泽) wrote: > hi, all: > > Recently, I read some qa case. I confuse about pg move case. > > In teuthology log, I see the log "Moving pg 1.3 from osd.1 to > osd.5". In this case, maybe the lastest data is in osd.1. After move > the pg to osd.5, the lastest data is in osd.5. > > But in peering process, we get pg info from osd.1 instead of > osd.5 base on osdmap. It would be wrong, right? Do I ignore some > important info? > > Could anyone explain the pg move case in detail? You're exactly right. And there is a hard-to-hit race where peering will make a bad inference because of that; see http://tracker.ceph.com/issues/16003 Most of the time, though, the OSD with the imported PG copy comes up and notifies the new primary that it has a copy before the new primary concludes that the recent PG events don't exist. Or the test model in ceph_test_rados just doesn't notice the resulting inconsistency. sage
