执行计划路径选择
postgres查询规划过程中，查询请求的不同执行方案是通过建立不同的路径来表达的，在生成许多符合条件的路径之后，要从中选择出代价最小的路径，把它转化为一个计划，传递给执行器执行，规划器的核心工作就是生成多条路径，然后从中找出最优的那一条。
代价评估
评估路径优劣的依据是用系统表pg_statistic中的统计信息估算出来的不同路径的代价(cost)，PostgreSQL估计计划成本的方式：基于统计信息估计计划中各个节点的成本。PostgreSQL会分析各个表来获取一个统计信息样本（这个操作通常是由autovacuum这个守护进程周期性的执行analyze，来收集这些统计信息，然后保存到pg_statistic和pg_class里面）。
用于估算代价的参数postgresql.conf
# - Planner Cost Constants -
#seq_page_cost = 1.0 # measured on an arbitrary scale 顺序磁盘扫描时单个页面的开销
#random_page_cost = 4.0 # same scale as above 随机磁盘访问时单页面的读取开销
#cpu_tuple_cost = 0.01 # same scale as above cpu处理每一行的开销
#cpu_index_tuple_cost = 0.005 # same scale as above cpu处理每个索引行的开销
#cpu_operator_cost = 0.0025 # same scale as above cpu处理每个运算符或者函数调用的开销
#parallel_tuple_cost = 0.1 # same scale as above 计算并行处理的成本，如果成本高于非并行，则不会开启并行处理。
#parallel_setup_cost = 1000.0 # same scale as above
#min_parallel_relation_size = 8MB
#effective_cache_size = 4GB 再一次索引扫描中可用的文件系统内核缓冲区有效大小
也可以使用 show all的方式查看
路径的选择
--查看表信息
db_jcxxglpt=# \d t_jcxxgl_tjaj
Table "db_jcxx.t_jcxxgl_tjaj"
Column | Type | Modifiers
--------------+--------------------------------+-----------
c_bh | character(32) | not null
c_xzdm | character varying(300) |
c_jgid | character(32) |
c_ajbm | character(22) |
...
Indexes:
"t_jcxxgl_tjaj_pkey" PRIMARY KEY, btree (c_bh)
"idx_ttjaj_cah" btree (c_ah)
"idx_ttjaj_dslrq" btree (d_slrq)
首先更新统计信息vacuum analyze t_jcxxgl_tjaj，许多时候可能因为统计信息的不准确导致了不正常的执行计划
--执行计划，全表扫描
db_jcxxglpt=# explain (analyze,verbose,costs,buffers,timing)select c_bh,c_xzdm,c_jgid,c_ajbm from t_jcxxgl_tjaj where d_slrq >='2018-03-18';
QUERY PLAN
------------------------------------------------------------------------------------------------------------
Seq Scan on db_jcxx.t_jcxxgl_tjaj (cost=0.00..9.76 rows=3 width=96) (actual time=1.031..1.055 rows=3 loops
=1)
Output: c_bh, c_xzdm, c_jgid, c_ajbm
Filter: (t_jcxxgl_tjaj.d_slrq >= '2018-03-18'::date)
Rows Removed by Filter: 138
Buffers: shared hit=8
Planning time: 6.579 ms
Execution time: 1.163 ms
(7 rows)
--执行计划，关闭全表扫描
db_jcxxglpt=# set session enable_seqscan = off;
SET
db_jcxxglpt=# explain (analyze,verbose,costs,buffers,timing)select c_bh,c_xzdm,c_jgid,c_ajbm from t_jcxxgl_tjaj where d_slrq >='2018-03-18';
QUERY PLAN
------------------------------------------------------------------------------------------------------------
Index Scan using idx_ttjaj_dslrq on db_jcxx.t_jcxxgl_tjaj (cost=0.14..13.90 rows=3 width=96) (actual time=0.012..0.026 rows=3 loops=1)
Output: c_bh, c_xzdm, c_jgid, c_ajbm
Index Cond: (t_jcxxgl_tjaj.d_slrq >= '2018-03-18'::date)
Buffers: shared hit=4
Planning time: 0.309 ms
Execution time: 0.063 ms
(6 rows)
d_slrq上面有btree索引，但是查看执行计划并没有走索引，这是为什么呢？
代价计算：
一个路径的估算由三部分组成：启动代价(startup cost)，总代价(totalcost)，执行结果的排序方式(pathkeys)
代价估算公式：总代价=启动代价+I/O代价+CPU代价（cost=S+P+W*T）
P:执行时要访问的页面数，反应磁盘的I/O次数
T:表示在执行时所要访问的元组数，反映了cpu开销
W:表示磁盘I/O代价和CPU开销建的权重因子
统计信息：统计信息的其中一部分是每个表和索引中项的总数，以及每个表和索引占用的磁盘块数。这些信息保存在pg_class表的reltuples和relpages列中。我们可以这样查询相关信息：
--查看统计信息
db_jcxxglpt=# select relpages,reltuples from pg_class where relname ='t_jcxxgl_tjaj';
relpages | reltuples
----------+-----------
8 | 141
(1 row)
total_cost = 1(seq_page_cost)*8(磁盘总页数)+0.01(cpu_tuple_cost)*141(表的总记录数)+0.0025(cpu_operation_cost)*141(表的总记录数)=9.7625
可以看到走索引的cost=13.90比全表扫描cost=9.76要大。在表较小的情况下，全表扫描比索引扫描更有效， index scan 至少要发生两次I/O，一次是读取索引块，一次是读取数据块。
seq_scan源码
/*
* cost_seqscan
* Determines and returns the cost of scanning a relation sequentially.
*
* 'baserel' is the relation to be scanned
* 'param_info' is the ParamPathInfo if this is a parameterized path, else NULL
*/
void
cost_seqscan(Path *path, PlannerInfo *root,
RelOptInfo *baserel, ParamPathInfo *param_info)
{
Cost startup_cost = 0;
Cost cpu_run_cost;
Cost disk_run_cost;
double spc_seq_page_cost;
QualCost qpqual_cost;
Cost cpu_per_tuple;
/* Should only be applied to base relations */
Assert(baserel->relid > 0);
Assert(baserel->rtekind == RTE_RELATION);
/* Mark the path with the correct row estimate */
if (param_info)
path->rows = param_info->ppi_rows;
else
path->rows = baserel->rows;
if (!enable_seqscan)
startup_cost += disable_cost;
/* fetch estimated page cost for tablespace containing table */
get_tablespace_page_costs(baserel->reltablespace, NULL,&spc_seq_page_cost);
/*
* disk costs
*/
disk_run_cost = spc_seq_page_cost * baserel->pages;
/* CPU costs */
get_restriction_qual_cost(root, baserel, param_info, &qpqual_cost);
startup_cost += qpqual_cost.startup;
cpu_per_tuple = cpu_tuple_cost + qpqual_cost.per_tuple;
cpu_run_cost = cpu_per_tuple * baserel->tuples;
/* tlist eval costs are paid per output row, not per tuple scanned */
startup_cost += path->pathtarget->cost.startup;
cpu_run_cost += path->pathtarget->cost.per_tuple * path->rows;
/* Adjust costing for parallelism, if used. */
if (path->parallel_workers > 0)
{
double parallel_divisor = get_parallel_divisor(path);
/* The CPU cost is divided among all the workers. */
cpu_run_cost /= parallel_divisor;
/*
* It may be possible to amortize some of the I/O cost, but probably
* not very much, because most operating systems already do aggressive
* prefetching. For now, we assume that the disk run cost can't be
* amortized at all.
*/
/*
* In the case of a parallel plan, the row count needs to represent
* the number of tuples processed per worker.
*/
path->rows = clamp_row_est(path->rows / parallel_divisor);
}
path->startup_cost = startup_cost;
path->total_cost = startup_cost + cpu_run_cost + disk_run_cost;
}
一个SQL优化实例
慢SQL：
select c_ajbh, c_ah, c_cbfy, c_cbrxm, d_larq, d_jarq, n_dbjg, c_yqly from db_zxzhld.t_zhld_db dbxx join db_zxzhld.t_zhld_ajdbxx dbaj
on dbxx.c_bh = dbaj.c_dbbh where dbxx.n_valid=1 and dbxx.n_state in (1,2,3) and dbxx.c_dbztbh='1003'
and dbaj.c_zblx='1003' and dbaj.c_dbfy='0' and dbaj.c_gy = '2550'
and c_ajbh in (select distinct c_ajbh from db_zxzhld.t_zhld_zbajxx where n_dbzt = 1 and c_zblx = '1003' and c_gy = '2550' )
order by d_larq asc, c_ajbh asc limit 15 offset 0
慢sql耗时：7s
咋们先过下这个sql是干什么的、首先dbxx和dbaj的一个join连接然后dbaj.c_ajbh要包含在zbaj表里面，做了个排序，取了15条记录、大概就这样。
Sql有个缺点就是我不知道查询的字段是从那个表里面取的、建议加上表别名.字段。
查看该sql的表的数据量：
db_zxzhld.t_zhld_db ：1311
db_zxzhld.t_zhld_ajdbxx ：341296
db_zxzhld.t_zhld_zbajxx ：1027619
执行计划：
01 Limit (cost=36328.67..36328.68 rows=1 width=107) (actual time=88957.677..88957.729 rows=15 loops=1)
02 -> Sort (cost=36328.67..36328.68 rows=1 width=107) (actual time=88957.653..88957.672 rows=15 loops=1)
03 Sort Key: dbaj.d_larq, dbaj.c_ajbh
04 Sort Method: top-N heapsort Memory: 27kB
05 -> Nested Loop Semi Join (cost=17099.76..36328.66 rows=1 width=107) (actual time=277.794..88932.662 rows=8605 loops=1)
06 Join Filter: ((dbaj.c_ajbh)::text = (t_zhld_zbajxx.c_ajbh)::text)
07 Rows Removed by Join Filter: 37018710
08 -> Nested Loop (cost=0.00..19200.59 rows=1 width=107) (actual time=199.141..601.845 rows=8605 loops=1)
09 Join Filter: (dbxx.c_bh = dbaj.c_dbbh)
10 Rows Removed by Join Filter: 111865
11 -> Seq Scan on t_zhld_ajdbxx dbaj (cost=0.00..19117.70 rows=219 width=140) (actual time=198.871..266.182 rows=8605 loops=1)
12 Filter: ((n_valid = 1) AND ((c_zblx)::text = '1003'::text) AND ((c_dbfy)::text = '0'::text) AND ((c_gy)::text = '2550'::text))
13 Rows Removed by Filter: 332691
14 -> Materialize (cost=0.00..66.48 rows=5 width=33) (actual time=0.001..0.017 rows=14 loops=8605)
15 -> Seq Scan on t_zhld_db dbxx (cost=0.00..66.45 rows=5 width=33) (actual time=0.044..0.722 rows=14 loops=1)
16 Filter: ((n_valid = 1) AND ((c_dbztbh)::text = '1003'::text) AND (n_state = ANY ('{1,2,3}'::integer[])))
17 Rows Removed by Filter: 1297
18 -> Materialize (cost=17099.76..17117.46 rows=708 width=32) (actual time=0.006..4.890 rows=4303 loops=8605)
19 -> HashAggregate (cost=17099.76..17106.84 rows=708 width=32) (actual time=44.011..54.924 rows=8605 loops=1)
20 Group Key: t_zhld_zbajxx.c_ajbh
21 -> Bitmap Heap Scan on t_zhld_zbajxx (cost=163.36..17097.99 rows=708 width=32) (actual time=5.218..30.278 rows=8605 loops=1)
22 Recheck Cond: ((n_dbzt = 1) AND ((c_zblx)::text = '1003'::text))
23 Filter: ((c_gy)::text = '2550'::text)
24 Rows Removed by Filter: 21849
25 Heap Blocks: exact=960
26 -> Bitmap Index Scan on i_tzhldzbajxx_zblx_dbzt (cost=0.00..163.19 rows=5876 width=0) (actual time=5.011..5.011 rows=30458 loops=1)
27 Index Cond: ((n_dbzt = 1) AND ((c_zblx)::text = '1003'::text))
28 Planning time: 1.258 ms
29 Execution time: 88958.029 ms
执行计划解读：
1：第27->21行，通过索引i_tzhldzbajxx_zblx_dbzt过滤表t_zhld_zbajxx的数据，然后根据过滤条件(c_gy)::text = '2550'::text过滤最终返回8605条数据
2：第17->15行，根据条件过滤t_zhld_db表的数据，最终返回了14条数据
3：第20->19行，对表t_zhld_zbajxx做group by的操作
4：第13->11行，全表扫描t_zhld_ajdbxx 最终返回了8605条数据
5：第08行，根据t_zhld_ajdbxx返回的8605条结果集作为驱动表和t_zhld_db的结果集(14条)做嵌套循环，t_zhld_db的结果集被循环了8605次。然后过滤掉了其中的111865条记录，那么最终将得到（8605*14-111865） = 8605
6：第07->05行，根据第08和18行返回的结果集最终做了Nested Loop Semi Join，第18行的4303条结果集被循环了8605次,(4303*8605-37018710)=8605
7: 第04->02行，对最终的8605条记录进行排序
8：第01行，limit最终获取15条记录
整个执行计划中耗时最长的地方在05行Nested Loop Semi Join，actual time=277.794..88932.662，
表db_zxzhld.t_zhld_db dbxx和db_zxzhld.t_zhld_ajdbxx均是全表扫描
具体优化步骤
查看索引页并没有索引，创建c_ajbh，c_dbbh等逻辑外键的索引
drop index if exists I_T_ZHLD_AJDBXX_AJBH;
create index I_T_ZHLD_AJDBXX_AJBH on T_ZHLD_AJDBXX (c_ajbh);
commit;
drop index if exists I_T_ZHLD_AJDBXX_DBBH;
create index I_T_ZHLD_AJDBXX_DBBH on T_ZHLD_AJDBXX (c_dbbh);
commit;
创建d_larq,c_ajbh的排序索引：
drop index if exists I_T_ZHLD_AJDBXX_m6;
create index I_T_ZHLD_AJDBXX_m6 on T_ZHLD_AJDBXX (c_zblx,c_dbfy,c_gy,d_larq asc,c_ajbh asc);
commit;
drop index if exists I_T_ZHLD_ZBAJXX_h3 ;
create index I_T_ZHLD_ZBAJXX_h3 on db_zxzhld.t_zhld_zbajxx (n_dbzt,c_zblx,c_gy,c_gy);
commit;
创建索引后执行计划有了改变，原来的dbaj表和dbxx表先做nestedloop变成了zbaj和dbaj表先做了nestedloop join,总的cost也从36328.68降到了12802.87，
Limit (cost=12802.87..12802.87 rows=1 width=107) (actual time=4263.598..4263.648 rows=15 loops=1)
-> Sort (cost=12802.87..12802.87 rows=1 width=107) (actual time=4263.592..4263.609 rows=15 loops=1)
Sort Key: dbaj.d_larq, dbaj.c_ajbh
Sort Method: top-N heapsort Memory: 27kB
-> Nested Loop (cost=2516.05..12802.86 rows=1 width=107) (actual time=74.240..4239.723 rows=8605 loops=1)
Join Filter: (dbaj.c_dbbh = dbxx.c_bh)
Rows Removed by Join Filter: 111865
-> Nested Loop (cost=2516.05..12736.34 rows=1 width=140) (actual time=74.083..327.974 rows=8605 loops=1)
-> HashAggregate (cost=2515.62..2522.76 rows=714 width=32) (actual time=74.025..90.185 rows=8605 loops=1)
Group Key: ("ANY_subquery".c_ajbh)::text
-> Subquery Scan on "ANY_subquery" (cost=2499.56..2513.84 rows=714 width=32) (actual time=28.782..59.823 rows=8605 loops=1)
-> HashAggregate (cost=2499.56..2506.70 rows=714 width=32) (actual time=28.778..39.968 rows=8605 loops=1)
Group Key: zbaj.c_ajbh
-> Index Scan using i_t_zhld_zbajxx_h3 on t_zhld_zbajxx zbaj (cost=0.42..2497.77 rows=715 width=32) (actual time=0.062..15.104 rows=8605 loops=1)
Index Cond: ((n_dbzt = 1) AND ((c_zblx)::text = '1003'::text) AND ((c_gy)::text = '2550'::text))
-> Index Scan using i_t_zhld_ajdbxx_ajbh on t_zhld_ajdbxx dbaj (cost=0.42..14.29 rows=1 width=140) (actual time=0.015..0.021 rows=1 loops=8605)
Index Cond: ((c_ajbh)::text = ("ANY_subquery".c_ajbh)::text)
Filter: (((c_zblx)::text = '1003'::text) AND ((c_dbfy)::text = '0'::text) AND ((c_gy)::text = '2550'::text))
Rows Removed by Filter: 1
-> Seq Scan on t_zhld_db dbxx (cost=0.00..66.45 rows=5 width=33) (actual time=0.015..0.430 rows=14 loops=8605)
Filter: ((n_valid = 1) AND ((c_dbztbh)::text = '1003'::text) AND (n_state = ANY ('{1,2,3}'::integer[])))
Rows Removed by Filter: 1298
Planning time: 1.075 ms
Execution time: 4263.803 ms
执行的时间还是要4s左右仍然不满足需求，并且没有使用上I_T_ZHLD_AJDBXX_m6这个索引。
等价改写SQL(1)
等价改写：将排序条件加入db_zxzhld.t_zhld_ajdbxx让其先排序，再和t_zhld_db表连接。
修改后sql：
Select dbaj.c_ajbh, dbaj.c_ah, dbaj.c_cbfy, dbaj.c_cbrxm, dbaj.d_larq, dbaj.d_jarq, dbaj.n_dbjg, dbaj.c_yqly
from (select * from db_zxzhld.t_zhld_db where n_valid=1 and n_state in (1,2,3) and c_dbztbh='1003' )dbxx
join (select * from db_zxzhld.t_zhld_ajdbxx
where n_valid=1 and c_zblx='1003'
and c_dbfy='0' and c_gy = '2550' and
c_ajbh in (select distinct c_ajbh from db_zxzhld.t_zhld_zbajxx where n_dbzt = 1 and c_zblx = '1003' and c_gy = '2550' ) order by d_larq asc, c_ajbh asc)dbaj
on dbxx.c_bh = dbaj.c_dbbh
limit 15 offset 0
再次查看执行计划：
Limit (cost=3223.92..3231.97 rows=1 width=107) (actual time=127.291..127.536 rows=15 loops=1)
-> Nested Loop (cost=3223.92..3231.97 rows=1 width=107) (actual time=127.285..127.496 rows=15 loops=1)
-> Sort (cost=3223.64..3223.65 rows=1 width=140) (actual time=127.210..127.225 rows=15 loops=1)
Sort Key: t_zhld_ajdbxx.d_larq, t_zhld_ajdbxx.c_ajbh
Sort Method: quicksort Memory: 2618kB
-> Hash Semi Join (cost=2523.19..3223.63 rows=1 width=140) (actual time=55.913..107.265 rows=8605 loops=1)
Hash Cond: ((t_zhld_ajdbxx.c_ajbh)::text = (t_zhld_zbajxx.c_ajbh)::text)
-> Index Scan using i_t_zhld_ajdbxx_m6 on t_zhld_ajdbxx (cost=0.42..700.28 rows=219 width=140) (actual time=0.065..22.005 rows=8605 loops=1)
Index Cond: (((c_zblx)::text = '1003'::text) AND ((c_dbfy)::text = '0'::text) AND ((c_gy)::text = '2550'::text))
-> Hash (cost=2513.84..2513.84 rows=714 width=32) (actual time=55.802..55.802 rows=8605 loops=1)
Buckets: 16384 (originally 1024) Batches: 1 (originally 1) Memory Usage: 675kB
-> HashAggregate (cost=2499.56..2506.70 rows=714 width=32) (actual time=30.530..43.275 rows=8605 loops=1)
Group Key: t_zhld_zbajxx.c_ajbh
-> Index Scan using i_t_zhld_zbajxx_h3 on t_zhld_zbajxx (cost=0.42..2497.77 rows=715 width=32) (actual time=0.043..15.552 rows=8605 loops=1)
Index Cond: ((n_dbzt = 1) AND ((c_zblx)::text = '1003'::text) AND ((c_gy)::text = '2550'::text))
-> Index Scan using t_zhld_db_pkey on t_zhld_db (cost=0.28..8.30 rows=1 width=33) (actual time=0.009..0.011 rows=1 loops=15)
Index Cond: (c_bh = t_zhld_ajdbxx.c_dbbh)
Filter: (((c_dbztbh)::text = '1003'::text) AND (n_state = ANY ('{1,2,3}'::integer[])))
Planning time: 1.154 ms
Execution time: 127.734 ms
这一次可以看出，ajdbxx和zbajxx表做了hash semi join 消除了nestedloop,cost降到了3231.97。并且使用上了i_t_zhld_ajdbxx_m6
子查询中in的结果集有一万多条数据，尝试使用exists等价改写in，看能否有更好的结果
等价改写SQL(2)
等价改写：将in替换为exists：
select c_ajbh, c_ah, c_cbfy, c_cbrxm, d_larq, d_jarq, n_dbjg, c_yqly
from (select c_bh from db_zxzhld.t_zhld_db where n_state in (1,2,3) and c_dbztbh='1003' )dbxx
join (select c_ajbh, c_ah, c_cbfy, c_cbrxm, d_larq, d_jarq, n_dbjg, c_yqly,c_dbbh from db_zxzhld.t_zhld_ajdbxx ajdbxx
where c_zblx='1003'
and c_dbfy='0' and c_gy = '2550' and
exists (select distinct c_ajbh from db_zxzhld.t_zhld_zbajxx zbajxx where ajdbxx.c_ajbh = zbajxx.c_ajbh and n_dbzt = 1 and c_zblx = '1003' and c_gy = '2550' ) order by d_larq asc, c_ajbh asc)dbaj
on dbxx.c_bh = dbaj.c_dbbh
limit 15 offset 0
再次查看执行计划：
Limit (cost=1.12..2547.17 rows=1 width=107) (actual time=0.140..0.727 rows=15 loops=1)
-> Nested Loop (cost=1.12..2547.17 rows=1 width=107) (actual time=0.136..0.689 rows=15 loops=1)
-> Nested Loop Semi Join (cost=0.85..2538.84 rows=1 width=140) (actual time=0.115..0.493 rows=15 loops=1)
-> Index Scan using i_t_zhld_ajdbxx_m6 on t_zhld_ajdbxx t2 (cost=0.42..700.28 rows=219 width=140) (actual time=0.076..0.127 rows=15 loops=1)
Index Cond: (((c_zblx)::text = '1003'::text) AND ((c_dbfy)::text = '0'::text) AND ((c_gy)::text = '2550'::text))
-> Index Scan using i_t_zhld_zbajxx_c_ajbh on t_zhld_zbajxx t3 (cost=0.42..8.40 rows=1 width=32) (actual time=0.019..0.019 rows=1 loops=15)
Index Cond: ((c_ajbh)::text = (t2.c_ajbh)::text)
Filter: (((c_zblx)::text = '1003'::text) AND ((c_gy)::text = '2550'::text) AND (n_dbzt = 1))
-> Index Scan using t_zhld_db_pkey on t_zhld_db (cost=0.28..8.30 rows=1 width=33) (actual time=0.007..0.008 rows=1 loops=15)
Index Cond: (c_bh = t2.c_dbbh)
Filter: (((c_dbztbh)::text = '1003'::text) AND (n_state = ANY ('{1,2,3}'::integer[])))
Planning time: 1.268 ms
Execution time: 0.859 ms
可以看出使用exist效果更好，最终cost 2547.17
(1).少了t_zhld_zbajxx表的group by操作：Sort Key: t_zhld_ajdbxx.d_larq, t_zhld_ajdbxx.c_ajbh。(这一步是因为使用了索引中的排序)
(2).少了分组的操作：Group Key: t_zhld_zbajxx.c_ajbh。
第(2)为什么这个查询消除了t_zhld_zbajxx表的group by操作呢？
原因是exists替换了distinct的功能，一旦满足条件则立刻返回。所以使用exists的时候子查询可以直接去掉distinct。
优化无止境、、、
参考资料：
pg11添加了并行hash_join的功能，有兴趣可以了解：