[Bug] Concurrent update cause select data error

[yanboer]

Apache Cloudberry version

gpadmin=# select version();
                                                                                                    version
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
 PostgreSQL 14.4 (Apache Cloudberry 2.0.0-incubating build 1) on x86_64-pc-linux-gnu, compiled by gcc (GCC) 11.5.0 20240719 (Red Hat 11.5.0-11), 64-bit compiled on Dec  2 2025 19:35:44 (with assert checking)
(1 row)

[gpadmin@cdw ~]$ pg_config
BINDIR = /usr/local/cloudberry-db/bin
DOCDIR = /usr/local/cloudberry-db/share/doc/postgresql
HTMLDIR = /usr/local/cloudberry-db/share/doc/postgresql
INCLUDEDIR = /usr/local/cloudberry-db/include
PKGINCLUDEDIR = /usr/local/cloudberry-db/include/postgresql
INCLUDEDIR-SERVER = /usr/local/cloudberry-db/include/postgresql/server
LIBDIR = /usr/local/cloudberry-db/lib
PKGLIBDIR = /usr/local/cloudberry-db/lib/postgresql
LOCALEDIR = /usr/local/cloudberry-db/share/locale
MANDIR = /usr/local/cloudberry-db/share/man
SHAREDIR = /usr/local/cloudberry-db/share/postgresql
SYSCONFDIR = /usr/local/cloudberry-db/etc/postgresql
PGXS = /usr/local/cloudberry-db/lib/postgresql/pgxs/src/makefiles/pgxs.mk
CONFIGURE =  '--prefix=/usr/local/cloudberry-db' '--enable-cassert' '--enable-debug-extensions' '--enable-ic-proxy' '--enable-mapreduce' '--enable-orafce' '--enable-orca' '--enable-pxf' '--enable-tap-tests' '--with-gssapi' '--with-ldap' '--with-libxml' '--with-openssl' '--with-pam' '--with-perl' '--with-pgport=5432' '--with-python' '--with-pythonsrc-ext'
CC = gcc
CPPFLAGS = -D_GNU_SOURCE -I/usr/include/libxml2
CFLAGS = -Wall -Wmissing-prototypes -Wpointer-arith -Werror=vla -Wendif-labels -Wmissing-format-attribute -Wimplicit-fallthrough=3 -Wcast-function-type -Wformat-security -fno-strict-aliasing -fwrapv -fexcess-precision=standard -Wno-unused-but-set-variable -Werror=implicit-fallthrough=3 -Wno-format-truncation -Wno-stringop-truncation -O3 -fPIC  -DUSE_INTERNAL_FTS=1  -Werror=uninitialized -Werror=implicit-function-declaration -Werror
CFLAGS_SL = -fPIC
LDFLAGS = -Wl,--as-needed -Wl,-rpath,'/usr/local/cloudberry-db/lib',--enable-new-dtags -Werror
LDFLAGS_EX =
LDFLAGS_SL =
LIBS = -lpgcommon -lpgport -lxerces-c -lbz2 -lxml2 -lpam -lrt -lssl -lcrypto -lgssapi_krb5 -luv -lz -lreadline -lm  -lcurl -lzstd
VERSION = PostgreSQL 14.4

What happened

select data error

## real result（multi rows）
gpadmin=*# select * from public.pgbench_branches where bid = 71;
 bid | bbalance | filler
-----+----------+--------
  71 |   868076 |
  71 |   785976 |
(2 rows)

gpadmin=*# \d public.pgbench_branches
s              Table "public.pgbench_branches"
  Column  |     Type      | Collation | Nullable | Default
----------+---------------+-----------+----------+---------
 bid      | integer       |           | not null |
 bbalance | integer       |           |          |
 filler   | character(88) |           |          |
Distributed by: (bid)

What you think should happen instead

select only 1 rows data

## expect result（only 1 rows）
gpadmin=*# select * from public.pgbench_branches where bid = 71;
 bid | bbalance | filler
-----+----------+--------
  71 |   xxx |
(1 rows)

How to reproduce

1、set param

gpconfig -c gp_enable_global_deadlock_detector -v on
gpconfig -c max_connections -v 4096 -m 2048
gpconfig -c gp_vmem_protect_limit -v 16384
gpstop -r -M fast -a

2、prepare sql

• error_snapshot.sql

begin;
SET TRANSACTION ISOLATION LEVEL repeatable read;
SELECT pg_catalog.pg_export_snapshot();
\copy public.pgbench_branches to '/home/gpadmin/pgbench_branches.csv';
select count(1) from public.pgbench_branches;

DO $$
DECLARE
    user_count INTEGER;
    max_users INTEGER := 1000;
BEGIN
    select count(1) into user_count from public.pgbench_branches;

    IF user_count > max_users THEN
        perform pg_sleep(10000000);
    END IF;
END 
$$;

COMMIT;

• pgbench.sql

\set bid random(1, 100 * :scale)
\set delta random(-50000, 50000)
BEGIN;
UPDATE pgbench_branches SET bbalance = bbalance + :delta WHERE bid = :bid;
END;

3、reproduce

### init data
pgbench -i -s 1000 -h localhost -p 5432

## concurrent run pgbench and error_snapshot.sql
pgbench -M prepared -h localhost -p 5432 -T 30000 --rate 100000 -j 32 -c 512 -P 1 -f pgbench.sql
for i in $(seq 1 100000); do psql -f error_snapshot.sql; done

Operating System

NAME="Rocky Linux" VERSION="9.7 (Blue Onyx)" ID="rocky" ID_LIKE="rhel centos fedora" VERSION_ID="9.7" PLATFORM_ID="platform:el9" PRETTY_NAME="Rocky Linux 9.7 (Blue Onyx)" ANSI_COLOR="0;32" LOGO="fedora-logo-icon" CPE_NAME="cpe:/o:rocky:rocky:9::baseos" HOME_URL="https://rockylinux.org/" VENDOR_NAME="RESF" VENDOR_URL="https://resf.org/" BUG_REPORT_URL="https://bugs.rockylinux.org/" SUPPORT_END="2032-05-31" ROCKY_SUPPORT_PRODUCT="Rocky-Linux-9" ROCKY_SUPPORT_PRODUCT_VERSION="9.7" REDHAT_SUPPORT_PRODUCT="Rocky Linux" REDHAT_SUPPORT_PRODUCT_VERSION="9.7"

Anything else

If the problem cannot be reproduced, try increasing the concurrency.

Are you willing to submit PR?

• Yes, I am willing to submit a PR!

Code of Conduct

• I agree to follow this project's Code of Conduct.

[github-actions]

Hi, @yanboer welcome!🎊 Thanks for taking the time to point this out.🙌

[yanboer]

problem reason

The issue arises from the time difference between QE releasing the lock and writing to DistributedLog, and QD removing itself from allTmGxact.

When committing a transaction, the QE first calls notifyCommittedDtxTransaction xact.c#L2956 to commit the transaction in QE

• QE will commit its transaction, release the lock, and write to the DistributedLog.

QD will remove itself from allTmGxact after procarray.c#L746 (Other transactions can still see the GXID when GetSnapshotData procarray.c#L2665).

During this period, other transactions can see the data deleted by the current transaction (the gxid corresponding to xmax is still active).

[HuSen8891]

problem reason

The issue arises from the time difference between QE releasing the lock and writing to DistributedLog, and QD removing itself from allTmGxact.

When committing a transaction, the QE first calls notifyCommittedDtxTransaction xact.c#L2956 to commit the transaction in QE

• QE will commit its transaction, release the lock, and write to the DistributedLog.

QD will remove itself from allTmGxact after procarray.c#L746 (Other transactions can still see the GXID when GetSnapshotData procarray.c#L2665).

During this period, other transactions can see the data deleted by the current transaction (the gxid corresponding to xmax is still active).

Thanks! I'll reproduce this issue, investigate the root cause.

[HuSen8891]

@yanboer Thank you very much! I have already reproduced this issue and identified the root cause. We will discuss this problem and the corresponding solutions, and then attempt to fix it.

[HuSen8891]

@yanboer Thank you very much! I have already reproduced this issue and identified the root cause. We will discuss this problem and the corresponding solutions, and then attempt to fix it.

This issue is relatively complex, we are planning complete solution, and this will take a significant amount of time to fix.