The function row_sel_get_clust_rec_for_mysql() is invoking row_sel_sec_rec_is_for_clust_rec() even at the READ UNCOMMITTED isolation level, where we should just assume that a non-delete-marked secondary index record does exist.

One of the two calls is preceded by a condition that refers to the isolation level, but that condition seems wrong:

		if (clust_rec

		    && (old_vers

			|| trx->isolation_level <= TRX_ISO_READ_UNCOMMITTED

			|| rec_get_deleted_flag(rec, dict_table_is_comp(

							sec_index->table)))

		    && !row_sel_sec_rec_is_for_clust_rec(

			    rec, sec_index, clust_rec, clust_index)) {

			clust_rec = NULL;

Because the lowest value for trx->isolation_level is

#define TRX_ISO_READ_UNCOMMITTED	0

the condition <= actually is equivalent to ==.

I think that for READ UNCOMMITTED, we should only filter secondary index records by the delete-mark flag. If a record is delete-marked, pretend that it does not exist. This is exactly what we should already be doing for clustered index records.