postgresql哈希聚合查询优化

我正在尝试优化下面的查询

select cellid2 as cellid, max(endeks) as turkcell 
from (select a.cellid2 as cellid2, b.endeks 
    from (select geom, cellid as cellid2 from grd_90098780_7c48_11e3_8876_f0bf97e0dd001000000000 ) a join (select endeks, st_transform(geom,    2320) as geom_tmp from  turkcell_data ) b on st_intersects(a.geom, b.geom_tmp) ) x
group by cellid2 limit 5

并解释和分析回报

"Limit  (cost=81808.31..81808.36 rows=5 width=12) (actual time=271376.201..271376.204 rows=5 loops=1)"
"  ->  HashAggregate  (cost=81808.31..81879.63 rows=7132 width=12) (actual time=271376.200..271376.203 rows=5 loops=1)"
"        ->  Nested Loop  (cost=0.00..81772.65 rows=7132 width=12) (actual time=5.128..269753.647 rows=1237707 loops=1)"
"              Join Filter: _st_intersects(grd_90098780_7c48_11e3_8876_f0bf97e0dd001000000000.geom, st_transform(turkcell_data.geom, 2320))"
"              ->  Seq Scan on turkcell_data  (cost=0.00..809.40 rows=3040 width=3045) (actual time=0.031..7.426 rows=3040 loops=1)"
"              ->  Index Scan using grd_90098780_7c48_11e3_8876_f0bf97e0dd001000000000_geom_gist on grd_90098780_7c48_11e3_8876_f0bf97e0dd001000000000  (cost=0.00..24.76 rows=7 width=124) (actual time=0.012..0.799 rows=647 loops=3040)"
"                    Index Cond: (geom && st_transform(turkcell_data.geom, 2320))"
"Total runtime: 271387.499 ms"

几何体列和单元ID列上存在索引。我读到，与使用max相比，orderbydesc和limit1更有效。然而，由于我有GROUPBY子句，我认为它不起作用。是否有任何方法可以做到这一点，或者有任何其他方法可以提高性能

表定义：

CREATE TABLE grd_90098780_7c48_11e3_8876_f0bf97e0dd001000000000
(
  regionid numeric,
  geom geometry(Geometry,2320),
  cellid integer,
  turkcell double precision
)
WITH (
  OIDS=FALSE
);
ALTER TABLE grd_90098780_7c48_11e3_8876_f0bf97e0dd001000000000
  OWNER TO postgres;

-- Index: grd_90098780_7c48_11e3_8876_f0bf97e0dd001000000000_cellid

-- DROP INDEX grd_90098780_7c48_11e3_8876_f0bf97e0dd001000000000_cellid;

CREATE INDEX grd_90098780_7c48_11e3_8876_f0bf97e0dd001000000000_cellid
  ON grd_90098780_7c48_11e3_8876_f0bf97e0dd001000000000
  USING btree
  (cellid );

-- Index: grd_90098780_7c48_11e3_8876_f0bf97e0dd001000000000_geom_gist

-- DROP INDEX grd_90098780_7c48_11e3_8876_f0bf97e0dd001000000000_geom_gist;

CREATE INDEX grd_90098780_7c48_11e3_8876_f0bf97e0dd001000000000_geom_gist
  ON grd_90098780_7c48_11e3_8876_f0bf97e0dd001000000000
  USING gist
  (geom );

CREATE TABLE turkcell_data
(
  gid serial NOT NULL,
  objectid_1 integer,
  objectid integer,
  neighbourh numeric,
  endeks numeric,
  coorx numeric,
  coory numeric,
  shape_leng numeric,
  shape_le_1 numeric,
  shape_area numeric,
  geom geometry(MultiPolygon,4326),
  CONSTRAINT turkcell_data_pkey PRIMARY KEY (gid )
)
WITH (
  OIDS=FALSE
);
ALTER TABLE turkcell_data
  OWNER TO postgres;

-- Index: turkcell_data_geom_gist

-- DROP INDEX turkcell_data_geom_gist;

CREATE INDEX turkcell_data_geom_gist
  ON turkcell_data
  USING gist
  (geom );

---

存储重新投影到2320的数据，为该列编制索引，并在联接中使用它，或者在

turkcell_data

中的几何体变换投影上创建索引。我通常喜欢后者：

CREATE INDEX turkcell_data_geom_gist2320
  ON turkcell_data
  USING gist
  (st_transform(geom, 2320) );

---

另一个问题可能是，如果你的几何体非常复杂，如果你的任何多边形有相对较多的点，你可能会被卡在交叉点上嘎吱嘎吱作响。不过，请先尝试索引。

如果希望我们帮助优化查询，您需要向我们显示表和索引定义，以及每个表的行数。也许您的表定义不好。可能索引没有正确创建。也许你在你认为你有的专栏上没有索引。如果看不到表和索引定义，我们无法判断。我们还需要行计数，因为这会极大地影响查询优化。我添加了必要的定义。嵌套循环有N=7和N=3040的子查询，结果是N=1237707行。这比carsesian的产品还要糟糕！我放置了索引，但它没有太大的变化。首先，您是否可以“设置enable_seqscan=false”，然后运行并发布解释分析（以验证索引是否正确创建）？然后“设置enable_seqscan=true”，再次运行并发布解释分析。你能描述一下你的图层吗？一层中大约有600个功能，另一层中大约有3000个。是否有任何几何图形的点数过高？是否第一层的几乎所有几何体都与第二层的几乎所有几何体相交？