-- CREATE INDEX at the smallest memory: a build participant spills runs of a few documents and
-- merges them in passes (a bounded fan-in), whatever the table size, serial or parallel. The index
-- is the one a build with plenty of memory writes, but for its segment layout.
CREATE EXTENSION pg_sift;
SET max_parallel_maintenance_workers = 2;

CREATE TABLE br (id serial, body text) WITH (parallel_workers = 2);
INSERT INTO br (body)
SELECT 'k' || (i % 7) || ' g' || (i % 1000) || ' common word' || (i % 50) || ' ' || repeat('filler ', i % 5)
FROM generate_series(1, 30000) i;
UPDATE br SET body = NULL WHERE id % 1000 = 0;

CREATE FUNCTION answers(OUT n bigint, OUT ids text) AS $$
	SELECT (SELECT count(*) FROM br WHERE body @@> 'common'),
	       (SELECT md5(string_agg(id::text, ',' ORDER BY id)) FROM br WHERE body @@> 'g5 OR g7 OR k3 AND word9')
$$ LANGUAGE sql;

CREATE FUNCTION ranked() RETURNS text AS $$
	SELECT string_agg(id::text || ':' || round(s::numeric, 5), ' ' ORDER BY s DESC, id) FROM
	(SELECT id, sift.score(ctid) AS s FROM br WHERE body @@> 'g5 OR g7 OR k3 AND word9'
	 ORDER BY s DESC, id LIMIT 30) t
$$ LANGUAGE sql;

-- the reference: serial, memory to spare
SET max_parallel_maintenance_workers = 0;
-- no dead version is left for the index build (a session of another test with an older snapshot makes the planner leave the index out: indcheckxmin)
CREATE TEMP TABLE br_rows AS SELECT * FROM br ORDER BY ctid;
TRUNCATE br;
INSERT INTO br SELECT * FROM br_rows;
DROP TABLE br_rows;
CREATE INDEX br_ref ON br USING sift (body);
CREATE TABLE ref AS SELECT a.*, ranked() AS top FROM answers() a;
SELECT n, top IS NOT NULL AS ranked FROM ref;

-- the answers of the index in use equal the reference's (a seq scan has no scores to rank by)
CREATE FUNCTION same(seq boolean DEFAULT false) RETURNS boolean AS $$
	SELECT a.n = r.n AND a.ids = r.ids AND CASE WHEN seq THEN true ELSE ranked() = r.top END
	FROM answers() a, ref r
$$ LANGUAGE sql;
DROP INDEX br_ref;

-- a seq scan agrees with the reference
SET sift.enable_scan = off;
SELECT same(true) AS seqscan_same;
RESET sift.enable_scan;

-- 1MB, three documents a run, merges of three runs: hundreds of runs a participant in the old build
SET maintenance_work_mem = '1MB';
SET sift.test_run_docs = 3;
SET sift.test_merge_fanin = 3;
SET max_parallel_maintenance_workers = 2;
SET sift.test_par_workers = 0;
CREATE INDEX br_par ON br USING sift (body);
SELECT current_setting('sift.test_par_workers')::int AS workers;
SELECT sift.verify('br_par', true);
SELECT same() AS parallel_same;
DROP INDEX br_par;

SET max_parallel_maintenance_workers = 0;
CREATE INDEX br_ser ON br USING sift (body);
SELECT sift.verify('br_ser', true);
SELECT same() AS serial_same;
DROP INDEX br_ser;

-- more runs than fit the fan-in at the minimum: the 1MB floor of the real build
RESET sift.test_merge_fanin;
SET max_parallel_maintenance_workers = 2;
CREATE INDEX br_par ON br USING sift (body);
SELECT sift.verify('br_par', true);
SELECT same() AS default_fanin_same;
DROP INDEX br_par;

-- a directory of 128 segments holds what the memory budget would spread over more: the build writes
-- larger segments instead of failing
SET sift.test_run_docs = 20;
SET sift.test_segment_postings = 5;
SET max_parallel_maintenance_workers = 0;
CREATE INDEX br_seg ON br USING sift (body);
SELECT count(*) BETWEEN 2 AND 128 AS within_directory FROM sift.segments('br_seg');
SELECT sift.verify('br_seg', true);
SELECT same() AS segments_same;
DROP INDEX br_seg;

-- the same with forced segment boundaries
CREATE INDEX br_bound ON br USING sift (body) WITH (build_segments = 60);
SELECT count(*) BETWEEN 60 AND 128 AS within_directory FROM sift.segments('br_bound');
SELECT sift.verify('br_bound', true);
SELECT same() AS bounded_same;

DROP TABLE br, ref;
DROP FUNCTION same(boolean), ranked(), answers();
DROP EXTENSION pg_sift;