-- VACUUM seals the open write buffer of a BM25 index once it holds a quarter of write_buffer_limit (1 MB at the default), as it does for a sift_col index, and the
-- rows are promoted: a search reads every document of the open buffer, so the 4 MB a buffer may reach cost a search many times what a small index needs
-- (300,000 documents, 16 MB: 684 blocks and 1.6 ms for a search with the buffer left, 186 blocks and 0.54 ms when the VACUUM of each round seals it).
-- A limit written larger keeps its buffer for longer.
CREATE EXTENSION pg_sift;
SET client_min_messages = warning;
SET sift.async_maintenance = off;
CREATE TABLE docs (id int PRIMARY KEY, body text) WITH (autovacuum_enabled = off);
CREATE TABLE docs_kept (LIKE docs INCLUDING ALL) WITH (autovacuum_enabled = off);
INSERT INTO docs SELECT g, (SELECT string_agg('w' || floor(2000 * power(abs(hashint8(g * 131 + k)::float8 / 2147483648.0), 2.5))::int, ' ') FROM generate_series(1, 25) k)
	FROM generate_series(1, 10000) g;
INSERT INTO docs_kept SELECT * FROM docs;
CREATE INDEX docs_ix ON docs USING sift (body);
CREATE INDEX docs_kept_ix ON docs_kept USING sift (body) WITH (write_buffer_limit = 16384);

-- 8,000 more documents are about 2 MB of write buffer, nine times the index; 2% of the first ones are deleted
INSERT INTO docs SELECT g, (SELECT string_agg('w' || floor(2000 * power(abs(hashint8(g * 131 + k)::float8 / 2147483648.0), 2.5))::int, ' ') FROM generate_series(1, 25) k)
	FROM generate_series(10001, 18000) g;
INSERT INTO docs_kept SELECT * FROM docs WHERE id > 10000;
DELETE FROM docs WHERE id % 50 = 0;
DELETE FROM docs_kept WHERE id % 50 = 0;
SELECT max(bytes) > 1500000 AS buffer_before FROM sift.segments('docs_ix') WHERE kind <> 'immutable';
VACUUM docs, docs_kept;

-- the default limit: the buffer was sealed and promoted; the larger limit keeps its buffer
SELECT count(*) FILTER (WHERE kind <> 'immutable' AND bytes > 500000) AS big_buffers_left, sum(ndocs) AS docs FROM sift.segments('docs_ix');
SELECT max(bytes) > 1500000 AS larger_limit_kept FROM sift.segments('docs_kept_ix') WHERE kind <> 'immutable';
SELECT count(*) FROM docs WHERE body @@> 'w3';
SELECT count(*) FROM docs_kept WHERE body @@> 'w3';
SELECT sift.verify('docs_ix', true);
DROP TABLE docs, docs_kept;
DROP EXTENSION pg_sift;