-- The key of a wildcard (`scores.*`, the keys of a JSON object) is a field of the object, a handful of names: the planner prices the dictionaries of
-- the units for it from the keys the object has, not from the statistics of the jsonb column (every document differs: every value distinct), which priced
-- the scan of this table at several times its cost. The scan stays the cheaper plan by a wide margin.
CREATE EXTENSION pg_sift;
SET max_parallel_workers_per_gather = 0;
CREATE TABLE wk_facts (id int, d date NOT NULL, metrics_a jsonb) WITH (autovacuum_enabled = false);
INSERT INTO wk_facts SELECT i, date '2020-01-01' + i % 1000,
	(SELECT jsonb_object_agg(k, 1 + (hashint4(i * 16 + n::int) & 7))
	   FROM unnest(ARRAY['a_1', 'a_2', 'b_1', 'b_2', 'c_1', 'c_2', 'd_1', 'd_2', 'e_1', 'e_2', 'f_1', 'f_2']) WITH ORDINALITY t(k, n))
	FROM generate_series(1, 100000) i;
SET client_min_messages = warning;
CREATE INDEX wk_facts_col ON wk_facts USING sift_col (id, d, metrics_a);
RESET client_min_messages;
VACUUM ANALYZE wk_facts;
SELECT used, scan_cost * 4 < normal_cost AS scan_far_cheaper
FROM sift.explain_why($$SELECT split_part(e.key, '_', 1), avg((e.value #>> '{}')::int) FROM wk_facts, jsonb_each(wk_facts.metrics_a) e
	WHERE d >= date '2021-01-01' AND d < date '2021-02-01' GROUP BY 1$$);
 used | scan_far_cheaper 
------+------------------
 t    | t
(1 row)

DROP TABLE wk_facts;
DROP EXTENSION pg_sift;