Index Scan using bigtable_y2020_cars_ref_idx on bigtable_y2020 bigtable (cost=0.57..2966738.51 rows=2873818 width=636) (actual time=0.026..0.026 rows=1 loops=1)
Given the system expects to get almost 3 million rows when it should be just 1, it seems like a stats problem to me. How is ndistinct on that column compared to reality?