thangTang commented on issue #1071: HBASE-23693 Split failure may cause region hole and data loss when use zk assign URL: https://github.com/apache/hbase/pull/1071#issuecomment-576352379 > This seems to solve the main issue, which is the parent and its interim daughters been wiped off. However, I'm wondering if the problem happens because we are setting parent state offline and split=true too soon in the split operation. What would you think, @thangTang ? > > Also, is it possible to provide a UT to reproduce? Might be too complex, though, since it involves some sort of race condition. I think the direct reason for this problem is that the information in the meta table has not been updated after the daughter region is cleaned up. At this time, the sub daughter no longer exists and the master will also try to reassign the parent region, so updating the information in the meta table should not have a negative impact. As for whether to adjust the timing of setting the parent region state in the meta table, we may need to sort out the split process as a whole. After all, when ZK assign is used, split transaction is complex. I've considered the problem of unit testing, and it's difficult to mock this case --- I met this problem in an extreme scenario --- RS crash when split step after PONR but not entirely complete, no other RS is available, and the master switch occurs.
---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services
