mirror of
https://github.com/scylladb/scylladb.git
synced 2026-05-22 07:42:16 +00:00
38d4015619
In scylla, the replacing node is set as HIBERNATE status. It is the only place we use HIBERNATE status. The replacing node is supposed to be alive and updating its heartbeat, so it is not supposed to be in dead state. This patch fixes the following problem in replacing. 1) start n1, n2 2) n2 is down 3) start n3 to replace n2, but kill n3 in the middle of the replace 4) start n4 to replace n2 After step 3 and step 4, the old n3 will stay in gossip forever until a full cluster shutdown. Note n3 will only stay in gossip but in system.peers table. User will see the annoying and infinite logs like on all the nodes rpc - client $ip_of_n3:7000: fail to connect: Connection refused Fixes: #5449 Tests: replace_address_test.py + manual test