As the GTIDs drift during upgrade process of a galera node (mysql_upgrade --skip-write-binlog --system-tables) and to resynchonize the nodes I ran:

SET SESSION gtid_seq_no=<a nice number above the current gtid>;

INSERT INTO hurfdurf (a) VALUES (1);

With 10.4.26 the gtid set event was recorded twice in the binlog causing binlog replication to stop. This is a new bug introduced after 10.4.21 as previous upgrades using the same method worked fine. 10.5.17 with set session wsrep_gtid_seq_no is not affected.

Workaround:
1. stop slave, switch to file+position after the duplicate event and start slave
2. flush binlog at source as you can't continue from a file which has the duplicate event even if you're past the duplicate event itself.
3. stop slave, flip back to gtid based replication, start slave