postgresql dont use index on timestamptz columns - postgresql

Question

Indexes on table:
create index shifts_start_at_idx
on shifts (start_at);
Query 1 with at time zone:
SELECT shifts.id
FROM shifts
JOIN stores ON shifts.store_id = stores.id AND stores.deleted_at IS NULL
JOIN cities ON stores.city_id = cities.id
WHERE TRUE
AND (shifts.start_at >= '2022-05-06 03:00:00'::timestamp AT TIME ZONE
(EXTRACT(timezone FROM cities.time_zone) * INTERVAL '1 second'))
ORDER BY shifts.start_at DESC, shifts.end_at DESC, shifts.id DESC
LIMIT 100;
Explain query 1:
Limit (cost=0.86..298.93 rows=100 width=24) (actual time=0.143..25.257 rows=100 loops=1)
-> Nested Loop (cost=0.86..1485256.59 rows=498300 width=24) (actual time=0.131..23.317 rows=100 loops=1)
" Join Filter: (shifts.start_at >= timezone((date_part('timezone'::text, cities.time_zone) * '00:00:01'::interval), '2022-05-06 03:00:00'::timestamp without time zone))"
-> Nested Loop (cost=0.72..1209695.67 rows=1494900 width=32) (actual time=0.096..17.621 rows=100 loops=1)
-> Index Scan Backward using shifts_admin_order_by_idx on shifts (cost=0.43..291132.79 rows=3000000 width=32) (actual time=0.036..6.780 rows=205 loops=1)
-> Index Scan using stores_id_deleted_at_null_idx on stores (cost=0.29..0.31 rows=1 width=16) (actual time=0.025..0.025 rows=0 loops=205)
Index Cond: (id = shifts.store_id)
-> Index Scan using cities_pkey on cities (cost=0.14..0.16 rows=1 width=20) (actual time=0.017..0.017 rows=1 loops=100)
Index Cond: (id = stores.city_id)
Planning Time: 0.632 ms
Execution Time: 26.436 ms
Postgres doesn't use index
Query 2 without at time zone:
SELECT shifts.id
FROM shifts
JOIN stores ON shifts.store_id = stores.id AND stores.deleted_at IS NULL
JOIN cities ON stores.city_id = cities.id
WHERE TRUE
AND (shifts.start_at >= '2022-05-06 03:00:00')
ORDER BY shifts.start_at DESC, shifts.end_at DESC, shifts.id DESC
LIMIT 100;
Explain query 2:
Limit (cost=0.86..108.84 rows=100 width=24) (actual time=0.125..8.866 rows=100 loops=1)
-> Nested Loop (cost=0.86..898691.17 rows=832261 width=24) (actual time=0.115..7.886 rows=100 loops=1)
-> Nested Loop (cost=0.72..761958.37 rows=832261 width=32) (actual time=0.066..5.570 rows=100 loops=1)
-> Index Scan Backward using shifts_admin_order_by_idx on shifts (cost=0.43..248984.02 rows=1670200 width=32) (actual time=0.014..1.380 rows=205 loops=1)
Index Cond: (start_at >= '2022-05-06 03:00:00+00'::timestamp with time zone)
-> Index Scan using stores_id_deleted_at_null_idx on stores (cost=0.29..0.31 rows=1 width=16) (actual time=0.008..0.008 rows=0 loops=205)
Index Cond: (id = shifts.store_id)
-> Index Only Scan using cities_pkey on cities (cost=0.14..0.16 rows=1 width=8) (actual time=0.008..0.008 rows=1 loops=100)
Index Cond: (id = stores.city_id)
Heap Fetches: 100
Planning Time: 0.327 ms
Execution Time: 9.394 ms
It is not entirely clear why it does not want to use the index when converting the time to a time format with a timezone