Howard Chu writes:

Fixed now in mdb.master. (...)

The me_nodemax issue remains though. Should be fixed in my branch "mdb/its7589".

However, another issue:

mdb_node_shrink() can give the node containing the sub-page uneven size, if MDB_DUPFIXED keys have uneven size: mdb_dbi_open(txn, "db", MDB_DUPSORT|MDB_DUPFIXED, &dbi); mdb_put(txn, dbi, &(MDB_val){4, "key"}, &(MDB_val){1, "A"}, 0); mdb_put(txn, dbi, &(MDB_val){4, "key"}, &(MDB_val){1, "B"}, 0); mdb_del(txn, dbi, &(MDB_val){4, "key"}, &(MDB_val){1, "B"});

Fix: Do nothing if the node would be uneven-sized. Similar to how mdb_cursor_put() adds room for 4N nodes at a time, with MDB_DUPFIXED. Dunno if the cursor tracking code in mdb_cursor_del() must still run.

I don't know what to do about databases that already contain such uneven-sized nodes. Maybe support reading them (if that needs any changes), but try to assert when writing. "Non-LEAF2 nodes are even-sized" feels like a deeply embedded assumption, but I haven't really looked.