[Bug] GPORCA ignores foreign segments of a postgres_fdw table and returns duplicated rows

未关闭
#2,341 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 2 天内回复

还没有人认领这个 Issue。

评估

难度
4/5
预计耗时
3-5 天
新手友好度
35/100
Issue 类型
缺陷
描述清晰度
描述清楚
活跃度
活跃
技术栈
c, postgresql
领域
databases

调研方向

Start by comparing GPORCA's handling of postgres_fdw foreign segments with the Postgres-based planner path, focusing on the foreign scan and gp_foreign_server behavior described in the issue. Reproduce with the provided SQL on a three-primary demo cluster, then inspect the optimizer code and tests for foreign tables with mpp_execute 'all segments'. Done means GPORCA returns one scan per foreign segment with correct server OIDs, or falls back to the Postgres planner.

由索引模型根据 Issue 内容生成。

描述

type: Bug
Apache Cloudberry version

Reproduced on:

  • REL_2_STABLE at 91582af9baf (PostgreSQL 14.12, 2.2.0-incubating build dev)
  • main at eb9bf338855 (PostgreSQL 16.9, 3.0.0-devel)
What happened

A postgres_fdw foreign table with mpp_execute 'all segments' that has several foreign
segments (ADD FOREIGN SEGMENT FROM SERVER ... INTO ..., added in #320) is scanned correctly by
the Postgres-based planner: one executor per foreign server. With GPORCA, which is the default
optimizer, the foreign segments are ignored. The table is scanned on all segments of the
cluster, so some remote servers are read more than once and rows are duplicated.
gp_foreign_server is 0 for every row.

What you think should happen instead

GPORCA should produce the same plan shape as the Postgres-based planner: one scan per foreign
segment, each reading its own server, with gp_foreign_server set to that server's OID. If
GPORCA cannot support foreign segments, it should fall back to the Postgres-based planner
instead of returning wrong results.

How to reproduce

Demo cluster with 3 primaries. The remote servers are two databases on the same cluster.

CREATE DATABASE bugfdw1;
CREATE DATABASE bugfdw2;
\c bugfdw1
CREATE TABLE t1 (a int, b text) DISTRIBUTED BY (a);
INSERT INTO t1 SELECT g, 'bugfdw1' FROM generate_series(1, 10) g;
\c bugfdw2
CREATE TABLE t1 (a int, b text) DISTRIBUTED BY (a);
INSERT INTO t1 SELECT g, 'bugfdw2' FROM generate_series(1, 10) g;

CREATE DATABASE bugrepro;
\c bugrepro
CREATE EXTENSION postgres_fdw;
CREATE SERVER mpps1 FOREIGN DATA WRAPPER postgres_fdw OPTIONS (host '127.0.0.1', port '7000', dbname 'bugfdw1');
CREATE SERVER mpps2 FOREIGN DATA WRAPPER postgres_fdw OPTIONS (host '127.0.0.1', port '7000', dbname 'bugfdw2');
CREATE USER MAPPING FOR CURRENT_USER SERVER mpps1;
CREATE USER MAPPING FOR CURRENT_USER SERVER mpps2;
CREATE FOREIGN TABLE fs1 (a int, b text) SERVER mpps1
  OPTIONS (table_name 't1', mpp_execute 'all segments');
ADD FOREIGN SEGMENT FROM SERVER mpps1 INTO fs1;
ADD FOREIGN SEGMENT FROM SERVER mpps2 INTO fs1;

SET optimizer = off;
EXPLAIN (COSTS OFF) SELECT * FROM fs1;
SELECT gp_foreign_server, b, count(*) FROM fs1 GROUP BY 1, 2 ORDER BY 1, 2;

SET optimizer = on;
EXPLAIN (COSTS OFF) SELECT * FROM fs1;
SELECT gp_foreign_server, b, count(*) FROM fs1 GROUP BY 1, 2 ORDER BY 1, 2;

Postgres-based planner (correct):

 Gather Motion 2:1  (slice1; segments: 2)
   ->  Foreign Scan on fs1
 Optimizer: Postgres query optimizer

 gp_foreign_server |    b    | count
-------------------+---------+-------
             17040 | bugfdw1 |    10
             17041 | bugfdw2 |    10

GPORCA (wrong):

 Gather Motion 3:1  (slice1; segments: 3)
   ->  Foreign Scan on fs1
 Optimizer: GPORCA

 gp_foreign_server |    b    | count
-------------------+---------+-------
                 0 | bugfdw1 |    20
                 0 | bugfdw2 |    10

The results are identical on REL_2_STABLE and main. With 3 segments and 2 foreign servers,
segment 2 wraps around to the first server, so bugfdw1 is read twice.

Operating System

Rocky Linux 9.7 (apache/incubator-cloudberry:cbdb-build-rocky9-latest), built with devops/build/automation/cloudberry/scripts.

Anything else

Joins on gp_foreign_server (a.gp_foreign_server = b.gp_foreign_server), which the feature
documents for co-located joins, also return wrong results under GPORCA because the column is
always 0. Workaround: SET optimizer = off before querying such tables.

Bug found by Claude Code. Please review it again.

主要语言
C
星标
1.4k
派生
273
平均合并
4 天 6 小时
30 天内合并 PR
105

环境准备

从这里开始

  1. 先读完整个 Issue,再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库,在一个分支上完成修改。
  4. 提交 Pull Request,并在描述里引用这个 Issue 编号。

apache/cloudberry 的其他 Issue

查看 apache/cloudberry 的全部 Issue

相似的 Issue

更多 C Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。