[Bug] ParsePrepareRecord() parses XLOG_XACT_PREPARE with a stale xl_xact_prepare layout: gid decodes as empty in pg_waldump and two-phase logical decoding

src/backend/access/rmgrdesc/xactdesc.c の ParsePrepareRecord() から始め、src/include/access/xact.h と src/include/access/twophase_xlog.h の TwoPhaseFileHeader を比較します。twophase.c の StartPrepare()/EndPrepare() を追跡します。提供されている pg_waldump および logical-decoding コマンドで問題を再現します。PREPARE が実際の gid を表示して出力し、two-phase decoding が正しく行われ、関連するレコードレイアウトにリグレッションがなければ完了です。

Apache Cloudberry version

main (89baa48a4f4, 2026-09-07) and local 3.0.0-devel+dev.8551.gc781604c5ba (PostgreSQL 16.9 base). The mismatch has existed since the GPDB-era commits b0208894eb5 / 68c0ff2d4ec / b74f174b6cd (2019) that extended TwoPhaseFileHeader.

What happened

ParsePrepareRecord() in src/backend/access/rmgrdesc/xactdesc.c walks a XLOG_XACT_PREPARE record using the layout of xl_xact_prepare (src/include/access/xact.h). But the record is actually written by StartPrepare()/EndPrepare() (twophase.c) as a TwoPhaseFileHeader (src/include/access/twophase_xlog.h), and Cloudberry extended that struct with fields that xl_xact_prepare never got:

TwoPhaseFileHeader (what is written, 96 bytes)     xl_xact_prepare (what is parsed, 72 bytes)
  ...                                                ...
  int32  nabortrels;                                 int32  nabortrels;
  int32  ncommitdbs;        <- extra                 int32  ncommitstats;
  int32  nabortdbs;         <- extra                 int32  nabortstats;
  int32  ncommitstats;                               int32  ninvalmsgs;
  int32  nabortstats;                                bool   initfileinval;
  int32  ninvalmsgs;                                 uint16 gidlen;
  bool   initfileinval;                              XLogRecPtr origin_lsn;
  Oid    tablespace_oid_to_delete_on_abort;  <- extra TimestampTz origin_timestamp;
  Oid    tablespace_oid_to_delete_on_commit; <- extra
  uint16 gidlen;
  XLogRecPtr origin_lsn;
  TimestampTz origin_timestamp;

Consequences of reading through the wrong struct:

• xlrec->gidlen is read from offset 54, which is the upper half of the real nabortstats (normally 0), so parsed->twophase_gid is an empty string.
• xlrec->ncommitstats / nabortstats / ninvalmsgs actually read ncommitdbs / nabortdbs / ncommitstats; initfileinval and origin_lsn read from unrelated bytes.
• bufptr starts at offset 72 instead of 96, so parsed->subxacts, xlocators, abortlocators, stats, abortstats, msgs all point at the wrong bytes. The parser also never skips the commitdbs / abortdbs (DbDirNode) arrays that EndPrepare() writes between abortrels and the stats arrays.

Two user-visible symptoms:

1. pg_waldump prints an empty gid for every PREPARE record (every distributed transaction on a segment):

   rmgr: Transaction ... desc: PREPARE gid : 2026-09-09 05:06:49.215203 CST
   rmgr: Transaction ... desc: COMMIT_PREPARED 5256: 2026-09-09 05:06:49.216803 CST gxid = 34413
   

   (COMMIT_PREPARED is parsed by ParseCommitRecord() and is fine, which shows the gid really is 34413 in the WAL.)

2. Logical decoding with two_phase emits PREPARE TRANSACTION '' with an empty gid, while the matching COMMIT PREPARED carries the real gid. A pgoutput subscriber would try to PREPARE TRANSACTION ''. Worse, DecodePrepare() passes the shifted parsed->subxacts / nsubxacts into SnapBuildCommitTxn() and ReorderBufferPrepare(), so decoding can consume garbage subxact ids.

Crash recovery is not affected because xact_redo() for PREPARE goes through twophase.c's own TwoPhaseFileHeader-based parsing, not ParsePrepareRecord().

What you think should happen instead

xl_xact_prepare must have the same layout as TwoPhaseFileHeader (the upstream PostgreSQL comment in twophase.c says the two must stay in sync), and ParsePrepareRecord() must skip the commitdbs / abortdbs arrays. Then pg_waldump shows PREPARE gid 34413: ... and two-phase logical decoding emits PREPARE TRANSACTION '34413'.

A static assert comparing sizeof(xl_xact_prepare) and sizeof(TwoPhaseFileHeader) (or simply typedef-ing one to the other) would keep this from regressing.

How to reproduce

On a demo cluster with wal_level = logical (needed only for step 2; step 1 works with replica):

-- via coordinator
create table t2pc(id int primary key, v text) distributed by (id);
insert into t2pc select g, 'x' from generate_series(1,6) g;   -- touches all segments -> 2PC

1. pg_waldump on any primary segment:

   pg_waldump -r Transaction <segdatadir>/pg_wal/<current wal file> | grep -E 'PREPARE|COMMIT_PREPARED' | tail -2
   

   shows PREPARE gid : <timestamp> (empty gid) followed by COMMIT_PREPARED ... gxid = <n>.

2. two-phase logical decoding on a segment (utility mode):

   -- PGOPTIONS='-c gp_role=utility' psql -p <segport>
   select * from pg_create_logical_replication_slot('s2pc', 'test_decoding', false, true);
   -- run the INSERT above on the coordinator
   select data from pg_logical_slot_peek_changes('s2pc', NULL, NULL, 'include-xids', '1');
   

   output:

   BEGIN 5256
   table public.t2pc: INSERT: id[integer]:2 v[text]:'x'
   ...
   PREPARE TRANSACTION '', txid 5256
   COMMIT PREPARED '34413', txid 5256
   

Operating System

CentOS 7 (kernel 3.10.0-1160), gcc 12.2.1

Anything else

Found while prototyping an MPP logical-replication receiver, where the segment-side gid (which is the distributed xid) is the key for aligning per-segment streams. The fix is small and self-contained; happy to send a PR.

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