UPDATE FOR PORTION OF leaves stale stored generated values after BEFORE INSERT trigger changes

UPDATE FOR PORTION OF creates leftover rows whose stored generated column does not reflect changes made by a BEFORE INSERT trigger. The statement succeeds, but both leftovers retain g='g-old' after the trigger changes a to 'trigger'.

This occurs when an application-time period update splits a row in a table with a stored generated column and a BEFORE INSERT trigger that changes its input.

Reproduction

Execute the following SQL in one MariaDB client session with permission to create a database, table and trigger. The database name must be unused. The initial row is inserted before the trigger is created; the subsequent ordinary INSERT provides a control for generated-column calculation after the trigger.

CREATE DATABASE period_demo;

USE period_demo;

CREATE TABLE t1 (id INT, a VARCHAR(20), s DATE, e DATE,

PERIOD FOR app_time(s,e),

g TEXT AS (CONCAT('g-',a)) STORED) ENGINE=InnoDB;

INSERT INTO t1 (id,a,s,e) VALUES (1,'old','2000-01-01','2020-01-01');

CREATE TRIGGER bi BEFORE INSERT ON t1 FOR EACH ROW SET NEW.a='trigger';

INSERT INTO t1 (id,a,s,e) VALUES (2,'control','2000-01-01','2020-01-01');

SELECT 'control',a,g,g=CONCAT('g-',a) AS ok FROM t1 WHERE id=2;

UPDATE t1 FOR PORTION OF app_time FROM '2005-01-01' TO '2010-01-01' SET a='new' WHERE id=1;

SELECT id,a,s,e,g,g=CONCAT('g-',a) AS ok FROM t1 ORDER BY id,s;

SELECT 'oracle',COUNT(*) AS rows_seen,SUM(NOT(g <=> CONCAT('g-',a))) AS mismatches FROM t1;

Actual result

The SQL completes without errors and returns:

control	a	g	ok

control	trigger	g-trigger	1

id	a	s	e	g	ok

1	trigger	2000-01-01	2005-01-01	g-old	0

1	new	2005-01-01	2010-01-01	g-new	1

1	trigger	2010-01-01	2020-01-01	g-old	0

2	trigger	2000-01-01	2020-01-01	g-trigger	1

oracle	rows_seen	mismatches

oracle	4	2

Expected result

Every stored g must equal CONCAT('g-',a), its declared generation expression. Both leftover rows should therefore contain a='trigger', g='g-trigger', and ok=1. The middle portion should remain a='new', g='g-new', and the ordinary INSERT control should remain a='trigger', g='g-trigger'. The final aggregate should report four rows and zero mismatches.

The trigger changing a in the leftover rows is expected; retaining a generated value calculated from the previous a is the defect.