SQLite R*Tree 索引不与 DISTINCT 一起使用

Question

在 SQLite 3.20.1 中，我有一个 R*Tree 索引 (dog_bounds) 和一个临时 table (frisbees) 创建如下：

-- Changes infrequently and has ~100k entries
CREATE VIRTUAL TABLE dog_bounds USING rtree (
    dog_id,
    min_x, max_x,
    min_y, max_y
);

-- Changes frequently and has ~100 entries
CREATE TEMPORARY TABLE frisbees (
    frisbee_id,
    min_x, max_x,
    min_y, max_y
);

使用此索引查询速度很快，如下所示：

EXPLAIN QUERY PLAN
SELECT dog_id FROM dog_bounds AS db, frisbees AS f
    WHERE db.max_x >= f.min_x AND db.max_y >= f.min_y
    AND db.min_x < f.max_x AND db.min_y < f.max_y;

0|0|1|SCAN TABLE frisbees AS f
0|1|0|SCAN TABLE dog_bounds AS db VIRTUAL TABLE INDEX 2:D1D3C0C2

但是，如果我 select DISTINCT(dog_id)，索引不再使用，查询变慢，即使在 ANALYZE:

之后

EXPLAIN QUERY PLAN
SELECT DISTINCT(dog_id) FROM dog_bounds AS db, frisbees AS f
    WHERE db.max_x >= f.min_x AND db.max_y >= f.min_y
    AND db.min_x < f.max_x AND db.min_y < f.max_y;

0|0|0|SCAN TABLE dog_bounds AS db VIRTUAL TABLE INDEX 2:
0|1|1|SCAN TABLE frisbees AS f
0|0|0|USE TEMP B-TREE FOR DISTINCT

我怎样才能得到这里使用的 R*Tree 索引？复制狗太可惜了！

Answer 1

查询优化器认为不同的执行顺序更容易获得不同的 dog_id 值。

将 R 树查找移动到子查询中，以便查询优化器被迫分别执行这两项操作：

SELECT DISTINCT dog_id
FROM (SELECT dog_id
      FROM dog_bounds AS db, frisbees AS f
      WHERE db.max_x >= f.min_x AND db.max_y >= f.min_y
        AND db.min_x <  f.max_x AND db.min_y <  f.max_y);

QUERY PLAN
|--SCAN TABLE dog_bounds AS db VIRTUAL TABLE INDEX 2:
|--SCAN TABLE frisbees AS f
`--USE TEMP B-TREE FOR DISTINCT

哎呀，查询优化器太聪明了 flattened the subquery。但是有一些方法可以禁用它（规则 21）：

SELECT DISTINCT dog_id
FROM (SELECT dog_id
      FROM dog_bounds AS db, frisbees AS f
      WHERE db.max_x >= f.min_x AND db.max_y >= f.min_y
        AND db.min_x <  f.max_x AND db.min_y <  f.max_y
      LIMIT -1);

QUERY PLAN
|--CO-ROUTINE 0x892A90
|  |--SCAN TABLE frisbees AS f
|  `--SCAN TABLE dog_bounds AS db VIRTUAL TABLE INDEX 2:D1D3C0C2
|--SCAN SUBQUERY 0x892A90
`--USE TEMP B-TREE FOR DISTINCT

SQLite R*Tree 索引不与 DISTINCT 一起使用

SQLite R*Tree index not used with DISTINCT

sqlite

indexing

r-tree

query-planner