At last, I solved this problem: I suspected the HDP version stored in postgres is wrong. So I modified these items: delete from servicedesiredstate where service_name='SPARK2'; delete from clusterservices where service_name='SPARK2'; delete from alert_grouping where definition_id = 201; delete from alert_grouping where definition_id = 202; delete from alert_grouping where definition_id = 203; delete from alert_definition where service_name = 'SPARK2'; select * from stack; stack_id | stack_name | stack_version ----------+-------------+----------------- 1 | HDP | 2.0.6.GlusterFS 2 | HDP | 2.5 3 | HDP | 2.4 4 | HDP | 2.3.ECS 5 | HDP | 2.0.6 6 | HDP | 2.1.GlusterFS 7 | HDP | 2.3 8 | HDP | 2.3.GlusterFS 9 | HDP | 2.2 10 | HDP | 2.1 11 | HDP | 2.0 51 | HDP | 2.6 101 | BigInsights | 4.2.5 102 | BigInsights | 4.1 103 | BigInsights | 4.0 104 | BigInsights | 4.2 update clusters set desired_stack_id = 51 where cluster_id = 2; Note that the digits above are referred to table alert_grouping and stack. So in my case, the desired_stack_id in table clusters was 3 (means HDP2.4). I updated it to 51(means HDP2.6) and corrected the undefined alert definition. At last the ambari server startup and recognize service Spark2 in HDP2.6. If someone can offer more details in ambari-server metadata in database, it will be easier to locate a problem like this. ... View more