[midenok]

This is 11.8 part of the fix, on top of 10.11 (#5881) and 11.4 (#5930).
The merge of 11.4 branch after #5930 pushed can be resolved by 10.11 and 11.4 commits from this PR.

Window_funcs_sort failed accessing win_func->window_spec as window spec was not defined in the query. Usually this is checked by fixing item win_func, but it was not done by setup_conds.

Actually, earlier check by vers_setup_conds() must fail on DELETE HISTORY from non-versioned table, but TABLE_LIST for t1 has no versioning conditions.

The cause was the parser assigning versioning conditions to wrong table pointed by last_table() which was already switched to another table from SYSTEM_TIME expression (cs1).

The fix assigns versioning conditions to the correct table stored to correspondent_table by delete_single_table branch of the parser.

[Copilot]

Copilot review overview

🔵 Needs a closer look

Add coverage for the analogous FOR PORTION OF path and address the two style nits.

Review effort: Lite
Findings: None

What changed in this PR

Fixes MDEV-40821 by attaching versioning conditions to the intended DELETE target, preventing a SIGSEGV when expressions introduce another table.

Changes:

• Uses correspondent_table for condition assignment.
• Adds regression coverage for sequence and window-function expressions.
• Updates expected results.

File	Summary
sql/​sql_yacc.yy	Corrects DELETE condition association. A moderate regression case remains missing for FOR PORTION OF with setval(), along with two line-length nits.
mysql-test/​suite/​versioning/​t/​delete_history.test	Adds MDEV-40821 regression cases.
mysql-test/​suite/​versioning/​r/​delete_history.result	Records expected test results.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

Copilot review overview

🟢 Approval recommended

The fix and regression coverage are acceptable; the remaining formatting nit is non-blocking.

Review effort: Lite
Findings: 1

Open (1)

• Break declaration exceeding 80-character limit · New

[CLAassistant]

All committers have signed the CLA.

[sanja-byelkin]

The fix is in the right place: the parser attached vers_conditions / period_conditions to last_table(), which becomes a table referenced from the BEFORE SYSTEM_TIME / FOR PORTION OF expression (here cs1 via setval()) instead of the DELETE target. Fixing that lets the existing fix_fields() / vers_setup_conds() checks do their job, with no guard needed in Window_funcs_sort::setup(). The vers_switch_partition assertion fix is also right. A few things should change before merge, though.

Should change

1. Target branch. The same last_table()->vers_conditions / ->period_conditions assignments and the assertion exist unchanged in 10.11 (sql_yacc.yy:13428/13454, sql_base.cc:1845) and 11.4 (sql_yacc.yy:13622/13668, sql_base.cc:1882). Please target 10.11 and merge up.

2. The DELETE target is spelled three ways (sql_yacc.yy:13963, 14011, sql_base.cc:1876): auxiliary_table_list.first->correspondent_table twice in the grammar and query_tables in the assertion. The sibling update_table_list just uses $$. Either make delete_single_table return the TABLE_LIST* and use $2 / $1, or use Lex->query_tables in both actions. The latter also backports to 10.11, where correspondent_table isn't set here.

3. The bug is also silent data loss, not only a crash. Without a window function there is no crash, but the conditions still go to the wrong table. Checked on an unpatched 11.8 debug build:

   • DELETE FROM t1 FOR PORTION OF app FROM (SELECT max(d) FROM t2) TO ... deleted every row of t1, with no error.
   • DELETE HISTORY FROM t1 BEFORE SYSTEM_TIME (SELECT ... FROM t2 LIMIT 1) deleted the current row and kept the history row. From reading the code, only DELETE_ACL is checked there, not DELETE_HISTORY_ACL.

   Please mention this in the commit message, because it affects how severe the bug is and where to backport.

4. Tests only check error codes, so they don't cover the case above. Please add:

   • a FOR PORTION OF case whose bound is a subquery on another table, with no window function, followed by select * from t1 to show the rows were not wholly deleted;
   • a successful DELETE HISTORY with a subquery in BEFORE SYSTEM_TIME, followed by select count(*) from t1 and ... for system_time all, to show current rows survive and only history is purged;
   • optionally, a user who has DELETE but not DELETE HISTORY, to cover the privilege check.

5. FOR PORTION OF via the multi-table path (sql_delete.cc:1961, 2090). A single-table DELETE is converted to multi-table when it has index hints (the shape used in this PR's test) or a subquery that references the target table. On that path FOR PORTION OF still deletes whole overlapping rows instead of splitting them, and skips the ER_NOT_CONSTANT_EXPRESSION check. Confirmed on unpatched 11.8 with a constant range; the patch doesn't touch it. Options:

   • add has_period() to processing_as_multitable_delete_prohibited();
   • reject it with ER_NOT_SUPPORTED_YET, as UPDATE already does;
   • at least file a follow-up MDEV and link it here.

Nits

• sql_yacc.yy:13963 and 14011 are over 80 columns; break after =.
• The tests hard-code engine innodb / engine myisam, which overrides the suites' engine combinations; please drop them. Keep ignore index (primary) only if it's needed to reach the crash, which ties in with point 5.
• sql_base.cc:1876: if you keep two spellings of the target, add a short comment saying they're the same table for single-table DELETE.

Note: the 10.11/11.4 and multi-table points come from reading the code. The patched build was not run; the runtime results above are from the unpatched 11.8 build.