-- CREATE INDEX on a table of large documents in one heap page: the text of a page was copied into one string
-- of 1 GB at most, so three documents of 360 MB (out of line, compressed in the heap) failed the build.
-- Each document larger than what a build buffers (a quarter of its memory, at most 256 MB; sift.test_page_text
-- lowers it to 1 MB) is analyzed on its own: three documents of 3.7 MB take less memory than their text.
CREATE EXTENSION pg_sift;
SET max_parallel_maintenance_workers = 0;
SET sift.test_page_text = 1048576;

CREATE TABLE ld (id int, body text);
INSERT INTO ld SELECT i, repeat('internationalization localization ' || 'w' || i || ' ', 100000) FROM generate_series(1, 3) i;
SELECT sum(length(body)) > 3 * 1048576 AS over_the_limit, min(length(body)) > 1048576 AS each_over_it, count(DISTINCT (ctid::text::point)[0]) AS pages FROM ld;

SET sift.test_build_peak = 0;
CREATE INDEX ld_idx ON ld USING sift (body);
SELECT pg_size_bytes(current_setting('sift.test_build_peak')) < (SELECT sum(length(body)) FROM ld) AS peak_below_the_text;
SELECT sift.verify('ld_idx', true);
SET sift.enable_scan = off;
SELECT id FROM ld WHERE body @@> 'w2' ORDER BY id;
RESET sift.enable_scan;
SELECT id FROM ld WHERE body @@> 'w2' ORDER BY id;
SELECT count(*) AS docs FROM ld WHERE body @@> 'internationalization localization';

DROP TABLE ld;
DROP EXTENSION pg_sift;