Thanks for the report! I repeated as described on 10.4-10.11, InnoDB. Probably caused by MDEV-371

--source include/have_innodb.inc 

CREATE TABLE t (s BLOB, n INT, UNIQUE (s)) engine=innodb;

INSERT INTO t VALUES ('Hrecvx_0004ln-00',1), ('Hrecvx_0004mm-00',1);

REPLACE INTO t VALUES ('Hrecvx_0004mm-00',2);

CURRENT_TEST: main.1_my

mysqltest: At line 9: query 'REPLACE INTO t VALUES ('Hrecvx_0004mm-00',2)' failed: 1062: Duplicate entry 'Hrecvx_0004mm-00' for key 's'

CREATE TABLE t (s BLOB, n INT, UNIQUE (s)) engine=innodb;

INSERT INTO t VALUES ('Hrecvx_0004ln-00',1), ('Hrecvx_0004mm-00',1);

INSERT INTO t VALUES ('Hrecvx_0004mm-00',2) ON DUPLICATE KEY UPDATE n = VALUES(n) ;

SELECT * FROM t;

SELECT * FROM t;

s	n

Hrecvx_0004ln-00	2

Hrecvx_0004mm-00	1

Thanks bar for the pointers, the bug with IDEMPOTENT replication was found with long uniques, which uses the same logic as REPLACE, and even contains the copy-paste from there. I guess it was made so to allocate and reuse the key buffer memory on the stack, saving from extra malloc. I think our priority now is to minimize stack usage vs extra mallocs, so it's not the point anymore.

I made some refactoring and extracted common code. Also moved handler's RND search initialization to Write_rows_log_event::do_before_row_operations. This fixed many long unique bugs in relplication and optimized the use a little bit.

https://github.com/MariaDB/server/commit/703e73e221a42638f2f05379124b35c57482da93

One more refactoring should be done to generalize handler initialization (and de-initialization) across REPLACE, LOAD DATA, IDEMPOTENT replication, and improve memory usage.

The commit https://github.com/MariaDB/server/commit/703e73e221a42638f2f05379124b35c57482da93 was never in a pull request nor included on any branch, and `git log -S MDEV-30046` does not yield any results from 10.5 branch, so I assume this issue is still open (and thus will also continue to keep https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=1015293 open).

The refactoring took a significant amount of time, but I think it will be very beneficial.

We had two implementations of the replace behavior, one is usual, and another one is an idempotent insert. The latter was many fixes behind the primary one.

First, the implementations should have been joined. I didn't stop only on that, though.
I also reworked the implementation to a better readability, as well as, as I believe, maintainability.
Also I have optimized it in several places.

The commit message will bring the better detail.

serg, please review commits the following commits

013fbe19 MDEV-30046 Refactor write_record and fix idempotent replication

aae554b1 MDEV-30046 wrong row targeted with "insert ... on duplicate" and "replace"

Or take a look at the same-titled commits at branch bb-10.5-nikita-MDEV-30046.

They are best to view with commit

9f2ef5d4 MDEV-15990 REPLACE on a precise-versioned table returns ER_DUP_ENTRY

as it further simplifies the REPLACE implementation, however it goes for review to midenok. Anyway, comments are welcome.

https://github.com/MariaDB/server/commit/972f6941df8b09b523f7f14d2465cee39a98fecb is ok to push, preferably with a clearer comment in sql_insert.cc.

The refactoring commit is generally moving in the right direction, but it won't be ready before the release, so I don't think the actual bug fix should wait for it.

For the record, this was merged in as https://github.com/MariaDB/server/commit/72429cad7f5448ad680d8df7b55aa055462d5964 (not visible in this Jira)