select version()：PostgreSQL 9.6.2 on x86_64-pc-linux-gnu, compiled by gcc (GCC) 4.8.2 20140120 (Red Hat 4.8.2-16), 64-bit
我有两张桌子：
create table if not exists cover.covering_s2_cell_ids (
covering_id int references cover.coverings(id) on delete cascade,
s2_cell_id bigint,
s2_cell_min bigint,
s2_cell_max bigint,
primary key (covering_id, s2_cell_id)
);
create table if not exists tiles.tileset_segment_counts (
tileset_id int references tiles.tilesets(id) on delete cascade,
s2_cell_id bigint not null,
segments jsonb not null,
num int not null,
primary key (tileset_id, s2_cell_id)
);
此外，我在tiles.tileset_segment_counts(tileset_id, s2_cell_id, num)上有另一个索引。我想要运行的查询是
SELECT tsc.s2_cell_id, tsc.segments
from cover.covering_s2_cell_ids cs
JOIN tiles.tileset_segment_counts tsc on tsc.s2_cell_id BETWEEN cs.s2_cell_min AND cs.s2_cell_max
WHERE cs.covering_id = 2
and tsc.tileset_id = 1
and num > 100
运行速度相对较快，似乎正在按照我的预期行事，利用三重索引来过滤来自tiles.tileset_segment_counts的行：
Nested Loop (cost=0.71..1285012.81 rows=7919778 width=544) (actual time=0.039..20.139 rows=19305 loops=1)
Output: tsc.s2_cell_id, tsc.segments
Buffers: shared hit=5150
-> Index Scan using covering_s2_cell_ids_covering_id_idx on cover.covering_s2_cell_ids cs (cost=0.29..12.04 rows=157 width=16) (actual time=0.018..0.088 rows=157 loops=1)
Output: cs.covering_id, cs.s2_cell_id, cs.s2_cell_min, cs.s2_cell_max
Index Cond: (cs.covering_id = 2)
Buffers: shared hit=4
-> Index Scan using tileset_segment_counts_tileset_id_s2_cell_id_num_idx on tiles.tileset_segment_counts tsc (cost=0.43..7680.28 rows=50444 width=544) (actual time=0.006..0.075 rows=123 loops=157)
Output: tsc.tileset_id, tsc.s2_cell_id, tsc.segments, tsc.num
Index Cond: ((tsc.tileset_id = 1) AND (tsc.s2_cell_id >= cs.s2_cell_min) AND (tsc.s2_cell_id <= cs.s2_cell_max) AND (tsc.num > 100))
Buffers: shared hit=5146
Planning time: 0.156 ms
Execution time: 23.760 ms
如果我将num增加为更大的值，则会出现问题。假设我使用大于1000，这是对该表的限制性更强的查询，查询计划程序选择使用大型tiles.tileset_segment_counts表的顺序扫描和过滤：
Nested Loop (cost=0.29..405447.61 rows=1624688 width=1111) (actual time=4656.731..6367.624 rows=6185 loops=1)
Output: tsc.s2_cell_id, tsc.segments
Join Filter: ((tsc.s2_cell_id >= cs.s2_cell_min) AND (tsc.s2_cell_id <= cs.s2_cell_max))
Rows Removed by Join Filter: 14430122
Buffers: shared hit=142735
-> Seq Scan on tiles.tileset_segment_counts tsc (cost=0.00..149546.77 rows=93135 width=1111) (actual time=0.119..214.902 rows=91951 loops=1)
Output: tsc.tileset_id, tsc.s2_cell_id, tsc.segments, tsc.num
Filter: ((tsc.num > 1000) AND (tsc.tileset_id = 1))
Rows Removed by Filter: 362013
Buffers: shared hit=142731
-> Materialize (cost=0.29..12.82 rows=157 width=16) (actual time=0.000..0.030 rows=157 loops=91951)
Output: cs.s2_cell_min, cs.s2_cell_max
Buffers: shared hit=4
-> Index Scan using covering_s2_cell_ids_covering_id_idx on cover.covering_s2_cell_ids cs (cost=0.29..12.04 rows=157 width=16) (actual time=0.015..0.052 rows=157 loops=1)
Output: cs.s2_cell_min, cs.s2_cell_max
Index Cond: (cs.covering_id = 2)
Buffers: shared hit=4
Planning time: 0.152 ms
Execution time: 6368.822 ms
我的想法是，在这种情况下，规划器应该更有可能想要在tileset_segment_counts上使用索引扫描，因为它返回更少的行。我确保在创建索引后对表进行真空分析。任何想法将不胜感激。我不明白为什么有这个限制性更强的谓词会推动规划者在索引条件下使用Join Filter +顺序扫描。
- 编辑 -
在enable_seqscan和enable_material设置为OFF的情况下，查询使用索引扫描或位图堆扫描并快速运行（下面为num> 1000）：
Nested Loop (cost=1010.05..744541.04 rows=1624688 width=1111) (actual time=0.048..8.272 rows=6185 loops=1)
Output: tsc.s2_cell_id, tsc.segments
Buffers: shared hit=2353
-> Index Scan using covering_s2_cell_ids_covering_id_idx on cover.covering_s2_cell_ids cs (cost=0.29..12.04 rows=157 width=16) (actual time=0.018..0.076 rows=157 loops=1)
Output: cs.covering_id, cs.s2_cell_id, cs.s2_cell_min, cs.s2_cell_max
Index Cond: (cs.covering_id = 2)
Buffers: shared hit=4
-> Bitmap Heap Scan on tiles.tileset_segment_counts tsc (cost=1009.77..4638.74 rows=10348 width=1111) (actual time=0.013..0.035 rows=39 loops=157)
Output: tsc.tileset_id, tsc.s2_cell_id, tsc.segments, tsc.num
Recheck Cond: ((tsc.tileset_id = 1) AND (tsc.s2_cell_id >= cs.s2_cell_min) AND (tsc.s2_cell_id <= cs.s2_cell_max) AND (tsc.num > 1000))
Heap Blocks: exact=1688
Buffers: shared hit=2349
-> Bitmap Index Scan on tileset_segment_counts_tileset_id_s2_cell_id_num_idx (cost=0.00..1007.18 rows=10348 width=0) (actual time=0.011..0.011 rows=39 loops=157)
Index Cond: ((tsc.tileset_id = 1) AND (tsc.s2_cell_id >= cs.s2_cell_min) AND (tsc.s2_cell_id <= cs.s2_cell_max) AND (tsc.num > 1000))
Buffers: shared hit=661
Planning time: 0.156 ms
Execution time: 9.492 ms
如果我只将enable_material设置为关闭并保留seq扫描，则情况也是如此。但是，仅设置enable_seqscan = OFF运行速度很慢，看起来与上面的慢速查询（第二个计划）完全相同，但不是顺序扫描，而是使用索引或位图堆扫描：
-> Bitmap Heap Scan on tiles.tileset_segment_counts tsc (cost=43731.56..178253.16 rows=93135 width=1111) (actual time=41.738..105.038 rows=91951 loops=1)
Output: tsc.tileset_id, tsc.s2_cell_id, tsc.segments, tsc.num
Recheck Cond: ((tsc.tileset_id = 1) AND (tsc.num > 1000))
Heap Blocks: exact=28833
Buffers: shared hit=38069
-> Bitmap Index Scan on tileset_segment_counts_tileset_id_s2_cell_id_num_idx (cost=0.00..43708.27 rows=93135 width=0) (actual time=36.765..36.765 rows=91951 loops=1)
Index Cond: ((tsc.tileset_id = 1) AND (tsc.num > 1000))
Buffers: shared hit=9236
我知道这是一个临时解决方案;我如何让计划者自己意识到这一点？这是成本问题吗？
- 编辑＃2 -
更新了第一次编辑中的第一个查询计划，以使num > 1000保持一致。
此外，我尝试将全局default_statistics_target设置为最大值（10000），对所有内容进行真空分析，然后重新运行。查询规划器仍然使用较慢的方法，使用大约6秒的连接过滤器进行顺序扫描。
更令人费解的是，如果我在联接中引入另一个表，它会让计划员使用更快的方法，无论统计目标或enable_material设置如何：
EXPLAIN (ANALYZE, COSTS, VERBOSE, BUFFERS)
SELECT tsc.s2_cell_id, tsc.segments
from public.markets d
join cover.covering_s2_cell_ids cs on d.default_covering_id = cs.covering_id
JOIN tiles.tileset_segment_counts tsc on tsc.s2_cell_id BETWEEN cs.s2_cell_min AND cs.s2_cell_max
WHERE d.id = 2
and tsc.tileset_id = 1
and num > 1000
Nested Loop (cost=0.71..99387.70 rows=1021667 width=1108) (actual time=0.033..7.860 rows=6185 loops=1)
Output: tsc.s2_cell_id, tsc.segments
Buffers: shared hit=2524
-> Nested Loop (cost=0.29..147.85 rows=100 width=16) (actual time=0.017..0.162 rows=157 loops=1)
Output: cs.s2_cell_min, cs.s2_cell_max
Buffers: shared hit=8
-> Seq Scan on public.markets d (cost=0.00..5.56 rows=1 width=4) (actual time=0.008..0.029 rows=1 loops=1)
Output: d.id, d.name, d.neilsen_id, d.market_key, d.default_covering_id, d.enabled
Filter: (d.id = 2)
Rows Removed by Filter: 204
Buffers: shared hit=3
-> Index Scan using covering_s2_cell_ids_pkey on cover.covering_s2_cell_ids cs (cost=0.29..141.29 rows=100 width=20) (actual time=0.006..0.072 rows=157 loops=1)
Output: cs.covering_id, cs.s2_cell_id, cs.s2_cell_min, cs.s2_cell_max
Index Cond: (cs.covering_id = d.default_covering_id)
Buffers: shared hit=5
-> Index Scan using tileset_segment_counts_tileset_id_s2_cell_id_num_idx on tiles.tileset_segment_counts tsc (cost=0.42..890.23 rows=10217 width=1108) (actual time=0.004..0.031 rows=39 loops=157)
Output: tsc.tileset_id, tsc.s2_cell_id, tsc.segments, tsc.num
Index Cond: ((tsc.tileset_id = 1) AND (tsc.s2_cell_id >= cs.s2_cell_min) AND (tsc.s2_cell_id <= cs.s2_cell_max) AND (tsc.num > 1000))
Buffers: shared hit=2516
Planning time: 0.416 ms
Execution time: 9.010 ms
真是亏本
- 编辑＃3 -
在cover.covering_s2_cell_ids(covering_id, s2_cell_min, s2_cell_max)上添加索引会使查询运行得更快，但仍然比预期慢得多：
Nested Loop (cost=0.29..267283.27 rows=1602272 width=1107) (actual time=68.918..411.856 rows=6185 loops=1)
Output: tsc.s2_cell_id, tsc.segments
Buffers: shared hit=340825 read=3810
-> Seq Scan on tiles.tileset_segment_counts tsc (cost=0.00..149528.89 rows=91850 width=1107) (actual time=0.010..213.931 rows=91951 loops=1)
Output: tsc.tileset_id, tsc.s2_cell_id, tsc.segments, tsc.num
Filter: ((tsc.num > 1000) AND (tsc.tileset_id = 1))
Rows Removed by Filter: 362013
Buffers: shared hit=138924 read=3807
-> Index Only Scan using covering_s2_cell_ids_covering_id_s2_cell_min_s2_cell_max_idx on cover.covering_s2_cell_ids cs (cost=0.29..1.11 rows=17 width=16) (actual time=0.002..0.002 rows=0 loops=91951)
Output: cs.covering_id, cs.s2_cell_min, cs.s2_cell_max
Index Cond: ((cs.covering_id = 2) AND (cs.s2_cell_min <= tsc.s2_cell_id) AND (cs.s2_cell_max >= tsc.s2_cell_id))
Heap Fetches: 0
Buffers: shared hit=201901 read=3
Planning time: 0.372 ms
Execution time: 413.054 ms
为tsc（tileset_id，num）添加索引会稍微改善它，但仍然很慢。我应该注意，覆盖表上的子查询只匹配大约157行。似乎快速（~10ms）和慢/中查询之间的差异是循环发生的顺序。请注意，在此最新查询中，我们必须在覆盖表上执行索引扫描的91951循环。在快速查询中，我们首先找到与覆盖匹配的行（157行），然后只需要在tileset表上执行157次索引扫描。我将尝试调整成本估算，因为规划明显高估了快速查询的成本。
答案 0 :(得分：0)
实现@ joop关于为minmax使用（int8）范围的想法（可以通过索引搜索）
create table covering_s2_cell_ids (
covering_id int -- references coverings(id) on delete cascade
, s2_cell_id bigint
-- , s2_cell_min bigint
-- , s2_cell_max bigint
, cellminmax int8range NOT NULL -- << HERE
, primary key (covering_id, s2_cell_id)
);
CREATE INDEX ON covering_s2_cell_ids USING GIST(cellminmax); -- << HERE
create table tileset_segment_counts (
tileset_id int -- references tilesets(id) on delete cascade
, s2_cell_id bigint not null
, segments jsonb not null
, num int not null
, primary key (tileset_id, s2_cell_id)
);
CREATE index on tileset_segment_counts(tileset_id, s2_cell_id, num);
VACUUM ANALYZE covering_s2_cell_ids;
VACUUM ANALYZE tileset_segment_counts;
EXPLAIN
SELECT tsc.s2_cell_id, tsc.segments
from covering_s2_cell_ids cs
JOIN tileset_segment_counts tsc
-- on tsc.s2_cell_id BETWEEN cs.s2_cell_min AND cs.s2_cell_max
on tsc.s2_cell_id <@ cs.cellminmax -- << HERE
WHERE cs.covering_id = 2
and tsc.tileset_id = 1
and num > 100
;
查询计划（无数据）：
-------------------------------------------------------------------------------------------------------------------------
Nested Loop (cost=3.24..9.61 rows=1 width=40)
-> Bitmap Heap Scan on tileset_segment_counts tsc (cost=1.22..3.34 rows=2 width=40)
Recheck Cond: ((tileset_id = 1) AND (num > 100))
-> Bitmap Index Scan on tileset_segment_counts_tileset_id_s2_cell_id_num_idx (cost=0.00..1.22 rows=2 width=0)
Index Cond: ((tileset_id = 1) AND (num > 100))
-> Bitmap Heap Scan on covering_s2_cell_ids cs (cost=2.02..3.13 rows=1 width=32)
Recheck Cond: ((tsc.s2_cell_id <@ cellminmax) AND (covering_id = 2))
-> BitmapAnd (cost=2.02..2.02 rows=1 width=0)
-> Bitmap Index Scan on covering_s2_cell_ids_cellminmax_idx (cost=0.00..0.66 rows=6 width=0)
Index Cond: (tsc.s2_cell_id <@ cellminmax)
-> Bitmap Index Scan on covering_s2_cell_ids_pkey (cost=0.00..1.21 rows=6 width=0)
Index Cond: (covering_id = 2)
(12 rows)