-- The groups of a bulk load are written by the workers of a parallel context when the load has more than one: a group is 8 units of 2 MB write
-- buffers (6500 rows of about 330 bytes each), the backend and its workers take a group each, and the rows and the TIDs are those of a load alone (which keeps one segment writer for all its units, so it leaves no more segments)
CREATE EXTENSION pg_sift;
SET sift.table_write_buffer_limit = 2048;
CREATE TABLE twin AS SELECT g AS id, g % 7 AS grp, 'row ' || g AS v, lpad(g::text, 300, '0') AS pad FROM generate_series(1, 150000) g;
CREATE FUNCTION same_rows(a regclass, b regclass) RETURNS boolean LANGUAGE plpgsql AS $$
DECLARE r boolean;
BEGIN
	EXECUTE format('SELECT (SELECT md5(string_agg(a::text, $q$,$q$ ORDER BY id)) FROM %1$s a) = (SELECT md5(string_agg(b::text, $q$,$q$ ORDER BY id)) FROM %2$s b)', a, b) INTO r;
	RETURN r;
END $$;
SET max_parallel_maintenance_workers = 2;
SET sift.test_par_workers = 0;
CREATE TABLE par USING sift_table AS SELECT * FROM twin;
SELECT current_setting('sift.test_par_workers')::int AS workers;
 workers 
---------
       2
(1 row)

SELECT count(*) AS segments FROM sift.table_segments('par') WHERE kind = 'columns';
 segments 
----------
        4
(1 row)

SELECT same_rows('par', 'twin');
 same_rows 
-----------
 t
(1 row)

-- alone: the same rows at the same TIDs, in no more segments
SET max_parallel_maintenance_workers = 0;
SET sift.test_par_workers = 0;
CREATE TABLE alone USING sift_table AS SELECT * FROM twin;
SELECT current_setting('sift.test_par_workers')::int AS workers;
 workers 
---------
       0
(1 row)

SELECT (SELECT array_agg(ctid ORDER BY id) FROM par) = (SELECT array_agg(ctid ORDER BY id) FROM alone) AS same_tids;
 same_tids 
-----------
 t
(1 row)

SELECT (SELECT count(*) FROM sift.table_segments('par') WHERE kind = 'columns') >= (SELECT count(*) FROM sift.table_segments('alone') WHERE kind = 'columns') AS no_more_segments;
 no_more_segments 
------------------
 t
(1 row)

-- a load that rolls back leaves nothing, and the refresh of a materialized view takes the same path
SET max_parallel_maintenance_workers = 2;
BEGIN;
CREATE TABLE gone USING sift_table AS SELECT * FROM twin;
ROLLBACK;
SELECT to_regclass('gone') IS NULL AS rolled_back;
 rolled_back 
-------------
 t
(1 row)

SET sift.test_par_workers = 0;
CREATE MATERIALIZED VIEW mv USING sift_table AS SELECT * FROM twin;
REFRESH MATERIALIZED VIEW mv;
SELECT current_setting('sift.test_par_workers')::int AS workers;
 workers 
---------
       2
(1 row)

SELECT same_rows('mv', 'twin');
 same_rows 
-----------
 t
(1 row)

-- values stored out of line are read in by the backend, the workers have no snapshot to read the toast table with
CREATE TABLE wide (id int, doc text);
ALTER TABLE wide ALTER doc SET STORAGE EXTERNAL;
INSERT INTO wide SELECT g, (SELECT string_agg(md5(g::text || i), '') FROM generate_series(1, 250) i) FROM generate_series(1, 2500) g;
SET sift.test_par_workers = 0;
CREATE TABLE wide_par USING sift_table AS SELECT * FROM wide;
SELECT current_setting('sift.test_par_workers')::int AS workers;
 workers 
---------
       2
(1 row)

SELECT same_rows('wide_par', 'wide');
 same_rows 
-----------
 t
(1 row)

DROP MATERIALIZED VIEW mv;
DROP TABLE twin, par, alone, wide, wide_par;
DROP FUNCTION same_rows(regclass, regclass);
DROP EXTENSION pg_sift;