A solution for MDEV-26345 works only in 10.5+, because a change in a statement in do_select() seems to be needed. And that change is part of the fix for MDEV-20468 which was fixed for 10.5+ despite affecting 10.1+. So we may want to backport MDEV-20468 to 10.4. I already have a draft backport patch (04899dbba585aff687860a7d193202f3218b51bb) and the new buildbot seems to be happy so far. See also comment https://jira.mariadb.org/browse/MDEV-26345?focusedCommentId=277134&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-277134

The MDEV-20468 commit:

8e92d5e5e3842fbb805304f1d0e4f5b1def9a145
MDEV-20468: Allocating more space than required for JOIN_TAB array for a query with SJM table

The change that is needed for MDEV-26345:

@@ -19772,8 +19774,7 @@ do_select(JOIN *join, Procedure *procedure)

     if (join->pushdown_query->store_data_in_temp_table)

     {

-      JOIN_TAB *last_tab= join->join_tab + join->table_count -

-                          join->exec_join_tab_cnt();      

+      JOIN_TAB *last_tab= join->join_tab + join->exec_join_tab_cnt();

       last_tab->next_select= end_send;

       enum_nested_loop_state state= last_tab->aggr->end_send();