Marko Mäkelä added a comment - 2023-09-04 05:59 There is a rather similar test case in MySQL Bug #86485 : set sql_mode= "" ; drop table if exists t; create table t ( `a` int , `b` int not null , `c` char (2) generated always as ( 'aa' ) virtual, unique key (`b`), unique key (`c`(1)) ) engine=innodb default charset=utf8 row_format=compact;   drop procedure if exists p; delimiter $ create procedure p() begin declare continue handler for sqlexception begin end ; repeat if rand()>0.5 then insert into t(a,b) values (floor(rand()*10),floor(rand()*10)); end if ; if rand()>0.5 then delete from t where a<floor(rand()*10); end if ; if rand()>0.5 then update t set b=concat(a,a); end if ; if rand()>0.5 then replace into t(a,b) values (floor(rand()*10),floor(rand()*10)); end if ; until 1=2 end repeat; end $ delimiter ;   call p(); As far as I can tell, the column a as well as the specification row_format=compact should be irrelevant. The column b could be defined as a PRIMARY KEY . Here is a slightly simplified mtr test: --source include/have_innodb.inc   set sql_mode= "" ; create table t ( b int primary key , c char (2) generated always as ( 'aa' ) virtual, unique key (c(1)) ) engine=innodb default charset=utf8;   delimiter $; create procedure p() begin declare continue handler for sqlexception begin end ; repeat if rand()>0.5 then insert into t(b) values (floor(rand()*10)); end if ; if rand()>0.5 then delete from t; end if ; if rand()>0.5 then update t set b=floor(rand()*10); end if ; if rand()>0.5 then replace into t(b) values (floor(rand()*10)); end if ; until 1=2 end repeat; end $ delimiter ;$   call p(); This will eventually crash due to a debug assertion failure: 10.6 2325f8f33971c29503639126c69e549aa53b5585 mysqltest: At line 23: query 'call p()' failed: <Unknown> (2013): Lost connection to server during query … Version: '10.6.16-MariaDB-debug-log' socket: '/dev/shm/10.6/mysql-test/var/tmp/4/mysqld.1.sock' port: 16060 Source distribution 2023-09-04 8:48:30 0 [ERROR] InnoDB: tried to purge non-delete-marked record in index `c` of table `test`.`t`: tuple: TUPLE (info_bits=0, 2 fields): {[1]a(0x61),[4] (0x80000000)}, record: COMPACT RECORD(info_bits=0, 2 fields): {[1]a(0x61),[4] (0x80000000)} mariadbd: /mariadb/10.6/storage/innobase/btr/btr0cur.cc:1011: dberr_t btr_cur_t::search_leaf(const dtuple_t*, page_cur_mode_t, btr_latch_mode, mtr_t*): Assertion `index()->is_btree() || index()->is_ibuf()' failed. 2023-09-04 8:48:30 0 [ERROR] InnoDB: Flagged corruption of `c` in table `test`.`t` in purge There was an attempt to fix this bug in MySQL 8.0.11 but never in MySQL 5.7 or MariaDB Server. I think that the fix may be incorrect, because it is dealing with bytes and not characters. A test case that used non-ASCII data should reveal it. With this port of the suspected incorrect patch and the above test case, 10.6 does not crash for me: diff --git a/storage/innobase/row/row0vers.cc b/storage/innobase/row/row0vers.cc index a4fc32cc5a8..de1586ac9af 100644 --- a/storage/innobase/row/row0vers.cc +++ b/storage/innobase/row/row0vers.cc @@ -734,9 +734,15 @@ row_vers_vc_matches_cluster( && (!compare[v_col->v_pos])) { if (ind_field->prefix_len != 0 - && !dfield_is_null(field2) - && field2->len > ind_field->prefix_len) { - field2->len = ind_field->prefix_len; + && !dfield_is_null(field2)) { + if (field2->len > ind_field->prefix_len) { + field2->len = ind_field->prefix_len; + } + if (field2->type.mbmaxlen > 1) { + /* FIXME: extract the correct number of characters + (not bytes)! */ + field2->len = field1->len; + } } /* The index field mismatch */ I interrupted the test after 143 seconds (23 concurrent instances). The last run without the patch would end in a crash in 26 seconds.

Marko Mäkelä added a comment - 2023-09-18 06:40 I think that the correct approach is to invoke dtype_get_at_most_n_mbchars() , like row_sel_sec_rec_is_for_clust_rec() and the MDEV-24402 row_check_index_match() do.

Marko Mäkelä added a comment - 2023-09-18 06:48 I tested an improved version of the test case in the Description: --source include/have_innodb.inc CREATE TABLE t (a BINARY (8) NOT NULL DEFAULT '' , b CHAR (8) AS (a) VIRTUAL, KEY (b(4))) CHARACTER SET utf8 ENGINE=InnoDB; INSERT INTO t (a) VALUES ( '' ),( '' ); UPDATE t SET a = 'x' ; UPDATE t SET a = '' ;   SET @save_freq=@@ GLOBAL .innodb_purge_rseg_truncate_frequency; SET GLOBAL innodb_purge_rseg_truncate_frequency=1; SET GLOBAL innodb_max_purge_lag_wait=0; SET GLOBAL innodb_purge_rseg_truncate_frequency=@save_freq;   CHECK TABLE t EXTENDED; DROP TABLE t; This test case passed 5×100 runs when the patch was applied. When not, it would fail reliably thanks to MDEV-24402 : 10.6 0f870914d410e8a34ea99b1e0ea7eb71094606f9 @@ -8,5 +8,6 @@ SET GLOBAL innodb_purge_rseg_truncate_frequency=@save_freq; CHECK TABLE t EXTENDED; Table Op Msg_type Msg_text -test.t check status OK +test.t check Warning InnoDB: Index b is marked as corrupted +test.t check error Corrupt DROP TABLE t;   mysqltest: Result length mismatch