Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                
Skip to content

Commit 8225c2f

Browse files
committed
Add regression test for BRIN parallel builds
Adds a regression test for parallel CREATE INDEX for BRIN indexes, to improve coverage for BRIN code, particularly code to allow parallel index builds introduced by b437571. The test is added to pageinspect, as that allows comparing the index to one built without parallelism. Another option would be to just build the index with parallelism and then check it produces correct results. But checking the index is exactly as if built without parallelism makes these query checks unnecessary. Discussion: https://postgr.es/m/1df00a66-db5a-4e66-809a-99b386a06d86%40enterprisedb.com
1 parent bb616ed commit 8225c2f

File tree

2 files changed

+192
-0
lines changed

2 files changed

+192
-0
lines changed

contrib/pageinspect/expected/brin.out

+101
Original file line numberDiff line numberDiff line change
@@ -108,3 +108,104 @@ SELECT (COUNT(*) = (SELECT relpages FROM pg_class WHERE relname = 'test2')) AS r
108108

109109
DROP TABLE test1;
110110
DROP TABLE test2;
111+
-- Test that parallel index build produces the same BRIN index as serial build.
112+
CREATE TABLE brin_parallel_test (a int, b text, c bigint) WITH (fillfactor=40);
113+
-- Generate a table with a mix of NULLs and non-NULL values (and data suitable
114+
-- for the different opclasses we build later).
115+
INSERT INTO brin_parallel_test
116+
SELECT (CASE WHEN (mod(i,231) = 0) OR (i BETWEEN 3500 AND 4000) THEN NULL ELSE i END),
117+
(CASE WHEN (mod(i,233) = 0) OR (i BETWEEN 3750 AND 4250) THEN NULL ELSE md5(i::text) END),
118+
(CASE WHEN (mod(i,233) = 0) OR (i BETWEEN 3850 AND 4500) THEN NULL ELSE (i/100) + mod(i,8) END)
119+
FROM generate_series(1,5000) S(i);
120+
-- Delete a couple pages, to make the ranges empty.
121+
DELETE FROM brin_parallel_test WHERE a BETWEEN 1000 and 1500;
122+
-- Vacuum to remove the tuples and make the ranges actually empty.
123+
VACUUM brin_parallel_test;
124+
-- Build an index with different opclasses - minmax, bloom and minmax-multi.
125+
--
126+
-- For minmax and opclass this is simple, but for minmax-multi we need to be
127+
-- careful, because the result depends on the order in which values are added
128+
-- to the summary, which in turn affects how are values merged etc. The order
129+
-- of merging results from workers has similar effect. All those summaries
130+
-- should produce correct query results, but it means we can't compare them
131+
-- using equality (which is what EXCEPT does). To work around this issue, we
132+
-- generated the data to only have very small number of distinct values per
133+
-- range, so that no merging is needed. This makes the results deterministic.
134+
-- build index without parallelism
135+
SET max_parallel_maintenance_workers = 0;
136+
CREATE INDEX brin_test_serial_idx ON brin_parallel_test
137+
USING brin (a int4_minmax_ops, a int4_bloom_ops, b, c int8_minmax_multi_ops)
138+
WITH (pages_per_range=7);
139+
-- build index using parallelism
140+
--
141+
-- Set a couple parameters to force parallel build for small table. There's a
142+
-- requirement for table size, so disable that. Also, plan_create_index_workers
143+
-- assumes each worker will use work_mem=32MB for sorting (which works for btree,
144+
-- but not really for BRIN), so we set maintenance_work_mem for 4 workers.
145+
SET min_parallel_table_scan_size = 0;
146+
SET max_parallel_maintenance_workers = 4;
147+
SET maintenance_work_mem = '128MB';
148+
CREATE INDEX brin_test_parallel_idx ON brin_parallel_test
149+
USING brin (a int4_minmax_ops, a int4_bloom_ops, b, c int8_minmax_multi_ops)
150+
WITH (pages_per_range=7);
151+
SELECT relname, relpages
152+
FROM pg_class
153+
WHERE relname IN ('brin_test_serial_idx', 'brin_test_parallel_idx')
154+
ORDER BY relname;
155+
relname | relpages
156+
------------------------+----------
157+
brin_test_parallel_idx | 3
158+
brin_test_serial_idx | 3
159+
(2 rows)
160+
161+
-- Check that (A except B) and (B except A) is empty, which means the indexes
162+
-- are the same.
163+
SELECT * FROM brin_page_items(get_raw_page('brin_test_parallel_idx', 2), 'brin_test_parallel_idx')
164+
EXCEPT
165+
SELECT * FROM brin_page_items(get_raw_page('brin_test_serial_idx', 2), 'brin_test_serial_idx');
166+
itemoffset | blknum | attnum | allnulls | hasnulls | placeholder | empty | value
167+
------------+--------+--------+----------+----------+-------------+-------+-------
168+
(0 rows)
169+
170+
SELECT * FROM brin_page_items(get_raw_page('brin_test_serial_idx', 2), 'brin_test_serial_idx')
171+
EXCEPT
172+
SELECT * FROM brin_page_items(get_raw_page('brin_test_parallel_idx', 2), 'brin_test_parallel_idx');
173+
itemoffset | blknum | attnum | allnulls | hasnulls | placeholder | empty | value
174+
------------+--------+--------+----------+----------+-------------+-------+-------
175+
(0 rows)
176+
177+
DROP INDEX brin_test_parallel_idx;
178+
-- force parallel build, but don't allow starting parallel workers to force
179+
-- fallback to serial build, and repeat the checks
180+
SET max_parallel_workers = 0;
181+
CREATE INDEX brin_test_parallel_idx ON brin_parallel_test
182+
USING brin (a int4_minmax_ops, a int4_bloom_ops, b, c int8_minmax_multi_ops)
183+
WITH (pages_per_range=7);
184+
SELECT relname, relpages
185+
FROM pg_class
186+
WHERE relname IN ('brin_test_serial_idx', 'brin_test_parallel_idx')
187+
ORDER BY relname;
188+
relname | relpages
189+
------------------------+----------
190+
brin_test_parallel_idx | 3
191+
brin_test_serial_idx | 3
192+
(2 rows)
193+
194+
SELECT * FROM brin_page_items(get_raw_page('brin_test_parallel_idx', 2), 'brin_test_parallel_idx')
195+
EXCEPT
196+
SELECT * FROM brin_page_items(get_raw_page('brin_test_serial_idx', 2), 'brin_test_serial_idx');
197+
itemoffset | blknum | attnum | allnulls | hasnulls | placeholder | empty | value
198+
------------+--------+--------+----------+----------+-------------+-------+-------
199+
(0 rows)
200+
201+
SELECT * FROM brin_page_items(get_raw_page('brin_test_serial_idx', 2), 'brin_test_serial_idx')
202+
EXCEPT
203+
SELECT * FROM brin_page_items(get_raw_page('brin_test_parallel_idx', 2), 'brin_test_parallel_idx');
204+
itemoffset | blknum | attnum | allnulls | hasnulls | placeholder | empty | value
205+
------------+--------+--------+----------+----------+-------------+-------+-------
206+
(0 rows)
207+
208+
DROP TABLE brin_parallel_test;
209+
RESET min_parallel_table_scan_size;
210+
RESET max_parallel_maintenance_workers;
211+
RESET maintenance_work_mem;

contrib/pageinspect/sql/brin.sql

+91
Original file line numberDiff line numberDiff line change
@@ -54,3 +54,94 @@ SELECT (COUNT(*) = (SELECT relpages FROM pg_class WHERE relname = 'test2')) AS r
5454

5555
DROP TABLE test1;
5656
DROP TABLE test2;
57+
58+
-- Test that parallel index build produces the same BRIN index as serial build.
59+
CREATE TABLE brin_parallel_test (a int, b text, c bigint) WITH (fillfactor=40);
60+
61+
-- Generate a table with a mix of NULLs and non-NULL values (and data suitable
62+
-- for the different opclasses we build later).
63+
INSERT INTO brin_parallel_test
64+
SELECT (CASE WHEN (mod(i,231) = 0) OR (i BETWEEN 3500 AND 4000) THEN NULL ELSE i END),
65+
(CASE WHEN (mod(i,233) = 0) OR (i BETWEEN 3750 AND 4250) THEN NULL ELSE md5(i::text) END),
66+
(CASE WHEN (mod(i,233) = 0) OR (i BETWEEN 3850 AND 4500) THEN NULL ELSE (i/100) + mod(i,8) END)
67+
FROM generate_series(1,5000) S(i);
68+
69+
-- Delete a couple pages, to make the ranges empty.
70+
DELETE FROM brin_parallel_test WHERE a BETWEEN 1000 and 1500;
71+
72+
-- Vacuum to remove the tuples and make the ranges actually empty.
73+
VACUUM brin_parallel_test;
74+
75+
-- Build an index with different opclasses - minmax, bloom and minmax-multi.
76+
--
77+
-- For minmax and opclass this is simple, but for minmax-multi we need to be
78+
-- careful, because the result depends on the order in which values are added
79+
-- to the summary, which in turn affects how are values merged etc. The order
80+
-- of merging results from workers has similar effect. All those summaries
81+
-- should produce correct query results, but it means we can't compare them
82+
-- using equality (which is what EXCEPT does). To work around this issue, we
83+
-- generated the data to only have very small number of distinct values per
84+
-- range, so that no merging is needed. This makes the results deterministic.
85+
86+
-- build index without parallelism
87+
SET max_parallel_maintenance_workers = 0;
88+
CREATE INDEX brin_test_serial_idx ON brin_parallel_test
89+
USING brin (a int4_minmax_ops, a int4_bloom_ops, b, c int8_minmax_multi_ops)
90+
WITH (pages_per_range=7);
91+
92+
-- build index using parallelism
93+
--
94+
-- Set a couple parameters to force parallel build for small table. There's a
95+
-- requirement for table size, so disable that. Also, plan_create_index_workers
96+
-- assumes each worker will use work_mem=32MB for sorting (which works for btree,
97+
-- but not really for BRIN), so we set maintenance_work_mem for 4 workers.
98+
SET min_parallel_table_scan_size = 0;
99+
SET max_parallel_maintenance_workers = 4;
100+
SET maintenance_work_mem = '128MB';
101+
CREATE INDEX brin_test_parallel_idx ON brin_parallel_test
102+
USING brin (a int4_minmax_ops, a int4_bloom_ops, b, c int8_minmax_multi_ops)
103+
WITH (pages_per_range=7);
104+
105+
SELECT relname, relpages
106+
FROM pg_class
107+
WHERE relname IN ('brin_test_serial_idx', 'brin_test_parallel_idx')
108+
ORDER BY relname;
109+
110+
-- Check that (A except B) and (B except A) is empty, which means the indexes
111+
-- are the same.
112+
113+
SELECT * FROM brin_page_items(get_raw_page('brin_test_parallel_idx', 2), 'brin_test_parallel_idx')
114+
EXCEPT
115+
SELECT * FROM brin_page_items(get_raw_page('brin_test_serial_idx', 2), 'brin_test_serial_idx');
116+
117+
SELECT * FROM brin_page_items(get_raw_page('brin_test_serial_idx', 2), 'brin_test_serial_idx')
118+
EXCEPT
119+
SELECT * FROM brin_page_items(get_raw_page('brin_test_parallel_idx', 2), 'brin_test_parallel_idx');
120+
121+
DROP INDEX brin_test_parallel_idx;
122+
123+
-- force parallel build, but don't allow starting parallel workers to force
124+
-- fallback to serial build, and repeat the checks
125+
126+
SET max_parallel_workers = 0;
127+
CREATE INDEX brin_test_parallel_idx ON brin_parallel_test
128+
USING brin (a int4_minmax_ops, a int4_bloom_ops, b, c int8_minmax_multi_ops)
129+
WITH (pages_per_range=7);
130+
131+
SELECT relname, relpages
132+
FROM pg_class
133+
WHERE relname IN ('brin_test_serial_idx', 'brin_test_parallel_idx')
134+
ORDER BY relname;
135+
136+
SELECT * FROM brin_page_items(get_raw_page('brin_test_parallel_idx', 2), 'brin_test_parallel_idx')
137+
EXCEPT
138+
SELECT * FROM brin_page_items(get_raw_page('brin_test_serial_idx', 2), 'brin_test_serial_idx');
139+
140+
SELECT * FROM brin_page_items(get_raw_page('brin_test_serial_idx', 2), 'brin_test_serial_idx')
141+
EXCEPT
142+
SELECT * FROM brin_page_items(get_raw_page('brin_test_parallel_idx', 2), 'brin_test_parallel_idx');
143+
144+
DROP TABLE brin_parallel_test;
145+
RESET min_parallel_table_scan_size;
146+
RESET max_parallel_maintenance_workers;
147+
RESET maintenance_work_mem;

0 commit comments

Comments
 (0)