Uh oh!
There was an error while loading. Please reload this page.
[improvement](tablet clone) fix balanced new replica will be removed when load txn continuously - #25061
Conversation
yujun777
commented
Oct 7, 2023
run buildall |
hello-stephen
commented
Oct 7, 2023
(From new machine)TeamCity pipeline, clickbench performance test result: |
de210b0 to
e9e3d69Compareyujun777
commented
Oct 8, 2023
run buildall |
hello-stephen
commented
Oct 8, 2023
(From new machine)TeamCity pipeline, clickbench performance test result: |
Proposed changes
pick: #24845
For a new clone replica, its init state is CLONE. During CLONE state, the replica can not load data.
When cloning finishing, check this replica is catchup with the partition. Check new replica's version is not enough. Because replica not write txn during CLONE, when these txns change from prepare to commit or visible, the new replica will be fallbehind (last failed version > 0). Later check this tablet is reduant and the new replica will be removed cause its last failed version > 0.
To fix this, when clone a new replica finished, we should check the txns during CLONE should be finished.
TEST: Let write txn continuously and none-stop, and make be clone a full snapshot cost 20s. Then let balance occur, later the new replica can success cloning after two or three further repair, and the old replica will be removed later.
Further comments
If this is a relatively large or complex change, kick off the discussion at dev@doris.apache.org by explaining why you chose the solution you did and what alternatives you considered, etc...