[BUG] Row binlog tablet schema missing __DORIS_BINLOG_LSN__ column index causing coredump when creating rowset writer

[heguanhui]

Search before asking

• I had searched in the issues and found no similar issues.

Version

newest master branch

What's Wrong?

Problem Description

When creating a row binlog table and writing data, a coredump occurs in RowBinlogSegmentWriter::init() with the following CHECK failure:
[----------] 2 tests from GroupRowsetWriterTest
19:13:27 [ RUN ] GroupRowsetWriterTest.sub_writer_rollback
19:13:27 F20260613 19:13:27.347458 362558 row_binlog_segment_writer.cpp:69] Check failed: lsn_col_id >= 0 binlog schema missing DORIS_BINLOG_LSN
19:13:27 *** Check failure stack trace: ***
19:13:27 @ 0x55812f5cd876 google::LogMessageFatal::~LogMessageFatal()
19:13:27 @ 0x55812bcd1f9e doris::segment_v2::RowBinlogSegmentWriter::init()
19:13:27 @ 0x55812b79d73e doris::SegmentFlusher::_create_segment_writer()
19:13:27 @ 0x55812b798030 doris::SegmentFlusher::flush_single_block()
19:13:27 @ 0x55812b7a18f1 doris::SegmentCreator::flush_single_block()
19:13:27 @ 0x55812b6dfaae doris::SegmentCreator::flush_single_block()
19:13:27 @ 0x55812b69e591 doris::BaseBetaRowsetWriter::flush_single_block()
19:13:28 @ 0x55812b6f4e2f doris::GroupRowsetWriter::flush_single_block()
19:13:28 @ 0x558114c3a9f7 doris::GroupRowsetWriterTest_sub_writer_rollback_Test::TestBody()
19:13:28 @ 0x55812fcc2f94 testing::internal::HandleSehExceptionsInMethodIfSupported<>()
19:13:28 @ 0x55812fcb1db6 testing::internal::HandleExceptionsInMethodIfSupported<>()
19:13:28 @ 0x55812fc941c7 testing::Test::Run()
19:13:28 @ 0x55812fc94cec testing::TestInfo::Run()
19:13:28 @ 0x55812fc954b2 testing::TestSuite::Run()
19:13:28 @ 0x55812fca6ccc testing::internal::UnitTestImpl::RunAllTests()
19:13:28 @ 0x55812fcc4dc4 testing::internal::HandleSehExceptionsInMethodIfSupported<>()
19:13:28 @ 0x55812fcb3e46 testing::internal::HandleExceptionsInMethodIfSupported<>()
19:13:28 @ 0x55812fca684b testing::UnitTest::Run()
19:13:29 @ 0x558117a84401 RUN_ALL_TESTS()
19:13:29 @ 0x558117a79b0e main
19:13:29 @ 0x7f419b4b3555 __libc_start_main
19:13:29 @ 0x55810c4be02a _start
19:13:29 @ (nil) (unknown)
19:13:29 run-be-ut.sh: line 595: 362558 Aborted (core dumped) LLVM_PROFILE_FILE="${profraw}" "${test}" --gtest_output="xml:${GTEST_OUTPUT_DIR}/${file_name}.xml" --gtest_print_time=true "${FILTER}" "${GEN_OUT}" "${GEN_REGRESSION_CASE}"

Steps to Reproduce

1. Run BE unit test:

   ./run-be-ut.sh --run GroupRowsetWriterTest.sub_writer_rollback
   

GDB Debug Info

(gdb) p this->_tablet_schema->_binlog_lsn_col_idx
$24 = -1
(gdb) p this->_tablet_schema->_binlog_timestamp_col_idx
$25 = -1
(gdb) p this->_tablet_schema->_num_columns
$26 = 5

The schema contains 5 columns (including 3 binlog special columns), but _binlog_lsn_col_idx is -1.

Steps to Reproduce

cd be/ut_build_dir
./run-be-ut.sh --run GroupRowsetWriterTest.sub_writer_rollback

### Root Cause
When converting TTabletSchema (Thrift) to TabletSchemaPB (Protobuf) via TabletMeta::init_schema_from_thrift(), the indices of binlog special columns (__DORIS_BINLOG_LSN__ and __DORIS_BINLOG_TIMESTAMP__) are not set in the protobuf.

As a result, when TabletSchema::init_from_pb() is called, _binlog_lsn_col_idx remains -1, causing the CHECK failure in RowBinlogSegmentWriter::init().

### Expected Behavior
_binlog_lsn_col_idx should be correctly set to the column index of __DORIS_BINLOG_LSN__ (which is 2 in the test case).


### What You Expected?

successfully executed run-be-ut.sh

### How to Reproduce?

execute run-be-ut.sh will certainly reproduce this problem

### Anything Else?

_No response_

### Are you willing to submit PR?

- [x] Yes I am willing to submit a PR!

### Code of Conduct

- [x] I agree to follow this project's [Code of Conduct](https://www.apache.org/foundation/policies/conduct)

[zclllyybb]

Breakwater-GitHub-Analysis-Slot: slot_c4e837598f90

Initial analysis: this looks like a real row-binlog metadata construction bug on current upstream/master. I checked the refreshed upstream/master ref at 4dda859061c, and the proposed root cause is consistent with the code path.

Evidence:

• testutil::enable_row_binlog() builds the row-binlog schema by appending __DORIS_BINLOG_LSN__, __DORIS_BINLOG_OP__, and __DORIS_BINLOG_TIMESTAMP__ after the normal columns. In the reported test case, the LSN column should therefore be column id 2, and the timestamp column should be column id 4.
• TabletMeta::init_schema_from_thrift() converts TTabletSchema directly into TabletSchemaPB by adding ColumnPBs. It propagates special indexes that exist on TTabletSchema, such as delete_sign_idx, but there is no thrift field for the row-binlog LSN/timestamp indexes, and this function does not derive those indexes from the column names.
• TabletMeta::init_from_pb() then calls TabletSchema::init_from_pb() for row_binlog_schema. That path assigns _binlog_lsn_col_idx and _binlog_timestamp_col_idx from the protobuf fields, whose defaults are -1.
• RowBinlogSegmentWriter::init() hard-checks _tablet_schema->binlog_lsn_col_idx() >= 0, so it aborts even though the __DORIS_BINLOG_LSN__ column exists in the schema's column list.

Recommended fix:

• In TabletMeta::init_schema_from_thrift(), while iterating tablet_schema.columns, set tablet_schema_pb->set_binlog_lsn_col_idx(...) and tablet_schema_pb->set_binlog_timestamp_col_idx(...) when tcolumn.column_name matches BINLOG_LSN_COL and BINLOG_TIMESTAMP_COL. Alternatively, route this conversion through the normal TabletSchema append/to-pb path so all special-column indexes are derived consistently.
• Add a regression assertion that the created row-binlog tablet schema has binlog_lsn_col_idx() == 2 and binlog_timestamp_col_idx() == 4, not only that field_index("__DORIS_BINLOG_LSN__") >= 0. field_index can pass while the persisted special-column index remains -1.
• Keep GroupRowsetWriterTest.sub_writer_rollback as the end-to-end guard; after the metadata fix it should no longer abort in RowBinlogSegmentWriter::init().

Missing information:

• For this BE unit-test failure, the issue already contains enough information to identify the bug.
• If the same failure is observed outside BE UT, please also provide the exact Doris commit SHA, the row-binlog table DDL/binlog properties, and the BE fatal log/core stack from that environment. If any tablet meta was created by the broken code before the fix, maintainers should also consider whether TabletSchema::init_from_pb() needs a compatibility fallback that derives missing row-binlog indexes from column names when the protobuf fields are still -1.