-- A wide prefix (more terms than sift.max_expansions) is built from the postings of its terms in windows of
-- superblocks that fit sift.test_prefix_gather, and a superblock that alone does not fit sums its postings
-- on the spot: the memory of a pass is bounded by the budget, the matches and the scores are those of the
-- unbounded build.
CREATE EXTENSION pg_sift;
CREATE EXTENSION pg_sift_test;
-- 40k documents on about 200 pages per 256: several superblocks; every document has 3 words of the prefix
CREATE TABLE rb (id int PRIMARY KEY, body text);
INSERT INTO rb SELECT i, 'tk' || lpad((i % 700)::text, 3, '0') || ' tk' || lpad((i % 311)::text, 3, '0') || ' tkz' || (i % 3) ||
	' w' || i % 7 || repeat(' pad filler words', i % 5) FROM generate_series(1, 40000) i;
CREATE INDEX rb_idx ON rb USING sift (body);
ANALYZE rb;
SELECT sum(sb_hi - sb_lo + 1) > 1 AS several_superblocks FROM sift.segments('rb_idx'::regclass);
 several_superblocks 
---------------------
 t
(1 row)

CREATE FUNCTION rb_answer(q text) RETURNS text LANGUAGE plpgsql AS $$
DECLARE r text;
BEGIN
	EXECUTE format('SELECT (SELECT count(*) FROM rb WHERE body @@> %1$L) || '' '' ||
		(SELECT string_agg(id || '':'' || s, '','' ORDER BY n) FROM (SELECT row_number() OVER () AS n, id, s
			FROM (SELECT id, sift.score(ctid) s FROM rb WHERE body @@> %1$L ORDER BY sift.score(ctid) DESC, id LIMIT 30) x) y)', q) INTO r;
	RETURN r;
END $$;
CREATE TEMP TABLE rb_q (q text);
INSERT INTO rb_q VALUES ('tk*'), ('tk0*'), ('tk00*'), ('tkz*'), ('tk* w1'), ('tk* AND NOT w3'), ('tk0* OR w2^3'), ('t*');
SET sift.max_expansions = 1;
CREATE TEMP TABLE rb_ref AS SELECT q, rb_answer(q) AS a FROM rb_q;
SELECT count(*) AS queries, count(DISTINCT a) AS distinct_answers FROM rb_ref;
 queries | distinct_answers 
---------+------------------
       8 |                7
(1 row)

SELECT key, value > 0 AS nonzero FROM sift._test_read_stats() WHERE key IN ('prefix_passes', 'prefix_windows', 'prefix_direct') ORDER BY key;
      key       | nonzero 
----------------+---------
 prefix_direct  | f
 prefix_passes  | t
 prefix_windows | f
(3 rows)

-- the unbounded build gathers every posting of the expansion at once: more than the budget used below
SELECT value * 4 > 64 * 1024 AS unbounded_gathers_more FROM sift._test_read_stats(true) WHERE key = 'prefix_peak_postings';
 unbounded_gathers_more 
------------------------
 t
(1 row)

-- 1 kB = 256 postings: every pass of the segment is over budget, the windows are one superblock each
SET sift.test_prefix_gather = 1;
SET sift.prefix_cache_size = 0;
SELECT q FROM rb_ref WHERE rb_answer(q) IS DISTINCT FROM a;
 q 
---
(0 rows)

SELECT key, value > 0 AS nonzero FROM sift._test_read_stats(true) WHERE key IN ('prefix_windows', 'prefix_direct') ORDER BY key;
      key       | nonzero 
----------------+---------
 prefix_direct  | t
 prefix_windows | t
(2 rows)

-- a budget of several superblocks: windows that gather, none that sums on the spot
SET sift.test_prefix_gather = 64;
SELECT q FROM rb_ref WHERE rb_answer(q) IS DISTINCT FROM a;
 q 
---
(0 rows)

SELECT key, value > 0 AS nonzero, value * 4 <= 64 * 1024 AS within_budget FROM sift._test_read_stats(true) WHERE key IN ('prefix_windows', 'prefix_peak_postings') ORDER BY key;
         key          | nonzero | within_budget 
----------------------+---------+---------------
 prefix_peak_postings | t       | t
 prefix_windows       | t       | t
(2 rows)

DROP FUNCTION rb_answer(text);
DROP TABLE rb;
DROP EXTENSION pg_sift_test;
DROP EXTENSION pg_sift;