查询只在leader有数据，fellow查询是空

💬 StarRocks 用户问答 功能使用相关
#1

为了更快的定位您的问题，请提供以下信息，谢谢
【详述】3.5.11存算分离版本，同一条sql在不同fe节点查询，只有leader查询有数据，其余节点是NULL
这里触发ALTER SYSTEM CREATE IMAGE后还是一样是NULL,具体sql如下

with ct001 as 
(
select 
            t1.p_dt
            ,t1.case_id
            ,t1.begin_date
            ,t1.end_date
            ,t1.company_name
            ,t1.department
            ,t1.post
            ,t1.work_place
            ,t1.work_record
       from
       (
            select 
                    t1.p_dt
                    ,t1.case_id
                    ,get_json_object(t2.item,'$.beginDate') as begin_date
                    ,get_json_object(t2.item,'$.endDate') as end_date
                    ,get_json_object(t2.item,'$.companyName') as company_name
                    ,get_json_object(t2.item,'$.department') as department
                    ,get_json_object(t2.item,'$.post') as post
                    ,get_json_object(t2.item,'$.workPlace') as work_place
                    ,t1.work_record
                    ,row_number() over(partition by t1.p_dt,t1.case_id order by get_json_object(t2.item,'$.beginDate') desc) as rn
            from hrdw_dwd.t_dwd_dw_hrm_register_flow_staff_info_df t1, 
            lateral json_each(t1.work_record) AS t2(item_index,item)
            where t1.p_dt =20260423 and case_id='RZ2026032500163'
       ) t1 
       where rn = 1
)
select t3.case_id ,t6.*
from hrdw_dwd.t_dwd_dw_hrm_register_flow_inaugural_form_df t1
left join hrdw_dwd.t_dwd_dw_hrm_register_flow_main_df t2 
       on t1.p_dt = t2.p_dt
      and t1.flow_main_id = t2.flow_main_id
left join hrdw_dwd.t_dwd_dw_hrm_register_flow_staff_info_df t3 
       on t1.p_dt = t3.p_dt 
      and t1.flow_main_id = t3.flow_main_id
left join hrdw_dim.t_dim_dw_core_location_df t4 
       on t1.p_dt = t4.p_dt 
      and t1.work_location_id = t4.location_id 
      and t4.tenant_id='tencent'
left join hrdw_dim.t_dim_md_core_manage_unit_df t5 
       on t3.p_dt = t5.p_dt
      and t3.manage_unit_id = t5.manage_unit_id
      and t5.tenant_id='tencent'
      and t5.data_status = 1
left join ct001 t6 
       on t1.p_dt = t6.p_dt
      and t3.case_id = t6.case_id
where t1.p_dt = 20260423
      and t1.enable_flag = 1
      and t2.flow_status_code in ('flowProcessing','flowComplete')
      and t1.staff_id =502439

【背景】无
【业务影响】
【是否存算分离】是
【StarRocks版本】3.5.11
【集群规模】例如：3fe（2 follower+1leader）+5cn（fe与cn混部）
【机器信息】CPU虚拟核/内存/网卡，例如：48C/64G/万兆
【联系方式】联系方式：社区群28-无痕

#2

看执行计划，fellow额外做了低基数字典优化，SET cbo_enable_low_cardinality_optimize = false;
后在fellow查出来了，这是版本bug吗

#3

StarRocks 3.5.11 在 Follower FE 上，当 SQL 同时包含以下要素时会触发 Bug ：

  1. 低基数字典优化（ cbo_enable_low_cardinality_optimize = true ，默认开启）
  2. Table Function（ json_each / lateral view）
  3. Window Function（ row_number() over(partition by ...) ）
  4. Partition by 的列被字典编码了（ p_dt ）

Follower 上 partition by 错误地使用了字典编码列（978）而不是原始列（586），导致数据窗口划分错误，最终 row_number() = 1 过滤掉所有数据，返回 NULL。

京ICP备2022026421号-1