AlterTable does not appear to anchor its WAL ops, meaning there is nothing preventing Kudu from GCing a WAL segment including an AlterTable that is running very slowly for some reason. If that happens and then the tserver is killed, it's possible for that replica to fail to start back up later. We should anchor alter ops in the same way we anchor write operations.
- is related to
KUDU-860 AlterSchema should log its success or failure in the COMMIT message