测试类型

功能测试

测试版本

openGauss 7.0.0-RC3 build 38d194a0 (B015)

问题描述

设置 query_dop=3 后，执行带 HashJoin 的列存表复杂查询时，gsql 立即报错 connection to server was lost，主节点 Primary Down。DN 日志显示 VecHashJoin 发生 Disk Spilled (early spilled)，随后 fatal 退出。query_dop=1 时相同查询正常执行。

已按文档《高级特性-并行查询》的配置步骤完整验证：系统负载确认充足（CPU 空闲 94.5%）、dop=1 EXPLAIN 确认计划适用并行、dop=3 EXPLAIN 确认并行计划已生成、执行查询即 crash。

操作系统和硬件信息

openEuler 24.03 LTS / aarch64 (Kunpeng 920) / 8 核 / 30 GB 内存 / 3节点 1主2备部署

测试环境

本地 3 节点 openGauss 7.0.0-RC3 (B015) 主备集群，DN 端口 15600，节点 20.20.20.224(主) / 20.20.20.226(备) / 20.20.20.228(备)。work_mem=64MB, shared_buffers=128MB。

被测功能

并行查询 (query_dop + SMP)

预置条件

数据库正常运行（cluster_state: Normal）

操作步骤

【第一步：观察系统负载 — 确认资源充足，符合文档要求（<50%）】

top - 15:08:00 up 11 days, load average: 1.68
%Cpu(s):  2.7 us, 94.5 id   ← CPU 空闲 94.5%，远低于 50% 阈值
MiB Mem: 30990.0 total, ~13GB avail Mem

结论：资源充足，适合使用 SMP 并行。

【第二步：创建列存表 + 100万行数据 + query_dop=1 EXPLAIN 验证计划符合并行场景】

drop table if exists column_table;
create table column_table (id int, a int, b char(20)) with (orientation = column);
drop table if exists column_table_01;
create table column_table_01 (id int, a int, b char(20)) with (orientation = column);

insert into column_table select id, id % 10, 'foo' from generate_series(1, 1000000) id;
insert into column_table_01 select id, id % 10, 'foo' from generate_series(1, 1000000) id;

analyze column_table;
analyze column_table_01;

SET query_dop = 1;
EXPLAIN (COSTS OFF)
select table_002.id column_006, table_002.b column_007,
       table_003.column_004 column_008, table_003.column_005 column_009
from column_table table_002
inner join (
    select table_001.id column_001, table_001.a column_002,
           table_002.a column_003, table_002.b column_004,
           table_001.b column_005
    from column_table table_002
    full outer join column_table_01 table_001
    on table_002.a = table_001.id
) table_003 on table_002.a = table_003.column_001
limit 10;

输出：计划含 Vector HashJoin + CStore Scan → 符合文档并行适用场景。

【第三步：query_dop=3 EXPLAIN 验证并行计划已生成】

SET query_dop = 3;
EXPLAIN (COSTS OFF) -- 同上查询

输出：

 id |                                  operation
----+------------------------------------------------------------------------------
  1 | ->  Row Adapter
  2 |    ->  Vector Limit
  3 |       ->  Vector Streaming(type: LOCAL GATHER dop: 1/3)
  4 |          ->  Vector Limit
  5 |             ->  Vector Sonic Hash Join (6,12)
  6 |                ->  Vector Streaming(type: BROADCAST dop: 3/3)
  7 |                   ->  Vector Hash Right Join (8, 10)
  8 |                      ->  Vector Streaming(type: LOCAL REDISTRIBUTE dop: 3/3)
  9 |                         ->  CStore Scan on column_table table_002
 10 |                      ->  Vector Streaming(type: LOCAL REDISTRIBUTE dop: 3/3)
 11 |                         ->  CStore Scan on column_table_01 table_001
 12 |                ->  CStore Scan on column_table table_002

结论：LOCAL GATHER(dop 1/3)、LOCAL BROADCAST、LOCAL REDISTRIBUTE 均已出现，并行计划正确生成。

【第四步：执行相同查询（非 EXPLAIN）→ crash】

SET query_dop = 3;
-- 直接执行查询
select table_002.id column_006, table_002.b column_007,
       table_003.column_004 column_008, table_003.column_005 column_009
from column_table table_002
inner join (
    select table_001.id column_001, table_001.a column_002,
           table_002.a column_003, table_002.b column_004,
           table_001.b column_005
    from column_table table_002
    full outer join column_table_01 table_001
    on table_002.a = table_001.id
) table_003 on table_002.a = table_003.column_001
limit 10;

输出：

SET
connection to server was lost

【第五步：清理】

drop table if exists column_table;
drop table if exists column_table_01;

注：crash 后清理无法执行，需先 gs_om -t restart 恢复集群。

预期输出

query_dop=3 时查询正常返回 10 rows，不 crash

实际输出

gsql 输出: SET
connection to server was lost
主节点状态变为 Primary Down，后续连接失败: failed to connect /home/wcl0623/cluster/tmp:15600

日志信息

DN pg_log 关键行（dn_6001 postgresql-*.log）:

• VecHashJoin(7) Disk Spilled
• Memory is not enough, early spilled
• FATAL: server process exited abnormally

crash 后需手动 gs_om -t restart 恢复集群

提单组织

测试团队

测试代码

Yat GUC 测试套: Opengauss_Function_Tools_Query_Dop_Case0038.py
路径: Yat/openGaussBase/testcase/GUC/VPC/