pg9.4 VS pg12大表join
pg9.4 VS pg12大表join
1、环境
1、从ECM生产财务库、TC库分别同步数据到pg9.4和pg12版本，保证表数据量一致。
2、数据库配置比较
| 数据库版本 | CPU核数 | 内存 | 最大IOPS | 存储类型 |
|---|---|---|---|---|
| 9.4 | 8核 | 16G | 8000 | 本地SSD盘 |
| 12 | 8核 | 16G | 26800 | ESSD云盘 |
3、由于ECM的导出大部分都是按照时间范围的大数据量导出，测试以发票为例，一个月的数据量。
2、前期准备
改原表名
alter table f_invoice
rename to f_invoice_2;
创建表分区
pg12有个特性，只要在主表上创建索引，分区表上自动创建对应的索引（实测btree和gin都行）。
create table ins_dw_prd12.f_invoice
(
invoice_id bigint not null,
source_type smallint not null,
cause_type smallint not null,
invoice_type smallint not null,
bill_type smallint not null,
outer_order_id varchar(64),
order_id bigint,
shop_id integer,
platform_id smallint,
pay_time timestamp,
invoice_title varchar(64) not null,
invoice_status smallint not null,
buyer_nick varchar(64),
seller_taxer_name varchar(64),
seller_taxer_code varchar(32),
invoice_amount numeric not null,
invoice_num varchar(128),
invoice_code varchar(128),
red_invoice_num varchar(128),
invoice_date timestamp,
pdf_url varchar(128),
pic_url varchar(128),
is_urgent smallint not null,
is_auto smallint not null,
sms_type smallint,
sms_notice_status smallint not null,
is_mix_split smallint not null,
invoice_flow varchar(32),
blue_invoice_id bigint,
old_blue_invoice_id bigint,
red_notice_num varchar(56),
down_time timestamp,
return_time timestamp,
invoice_remark varchar(256),
remark varchar(256),
reject_reason smallint,
reject_desc varchar(256),
is_lost smallint,
is_apply_writeoff smallint,
is_writeoff_result smallint,
err_msg varchar(256),
is_lock smallint,
is_pre_reopen smallint,
is_ad smallint,
create_time timestamp not null,
update_time timestamp with time zone not null,
gl_date timestamp,
ims_customer_code varchar(32),
check_code varchar(64),
red_reason_type smallint,
receiver_email varchar(128),
primary key (invoice_id,create_time)
) partition by range (create_time);
-- 创建表分区
create table f_invoice_p2018 partition of f_invoice for values from (minvalue )to ('2019-01-01');
create table f_invoice_p201901 partition of f_invoice for values from ('2019-01-01' )to ('2019-02-01');
create table f_invoice_p201902 partition of f_invoice for values from ('2019-02-01' )to ('2019-03-01');
create table f_invoice_p201903 partition of f_invoice for values from ('2019-03-01' )to ('2019-04-01');
create table f_invoice_p201904 partition of f_invoice for values from ('2019-04-01' )to ('2019-05-01');
create table f_invoice_p201905 partition of f_invoice for values from ('2019-05-01' )to ('2019-06-01');
create table f_invoice_p201906 partition of f_invoice for values from ('2019-06-01' )to ('2019-07-01');
create table f_invoice_p201907 partition of f_invoice for values from ('2019-07-01' )to ('2019-08-01');
create table f_invoice_p201908 partition of f_invoice for values from ('2019-08-01' )to ('2019-09-01');
create table f_invoice_p201909 partition of f_invoice for values from ('2019-09-01' )to ('2019-10-01');
create table f_invoice_p201910 partition of f_invoice for values from ('2019-10-01' )to ('2019-11-01');
create table f_invoice_p201911 partition of f_invoice for values from ('2019-11-01' )to ('2019-12-01');
create table f_invoice_p201912 partition of f_invoice for values from ('2019-12-01' )to ('2020-01-01');
create table f_invoice_p202001 partition of f_invoice for values from ('2020-01-01' )to ('2020-02-01');
create table f_invoice_p202002 partition of f_invoice for values from ('2020-02-01' )to ('2020-03-01');
create table f_invoice_p202003 partition of f_invoice for values from ('2020-03-01' )to ('2020-04-01');
create table f_invoice_p202004 partition of f_invoice for values from ('2020-04-01' )to ('2020-05-01');
create table f_invoice_p202005 partition of f_invoice for values from ('2020-05-01' )to ('2020-06-01');
create table f_invoice_p202006 partition of f_invoice for values from ('2020-06-01' )to ('2020-07-01');
CREATE TABLE f_invoice_default PARTITION OF f_invoice DEFAULT;
-- 创建索引
create index idx_f_invoice_gin2
on ins_dw_prd12.f_invoice using gin (source_type, invoice_type, invoice_status, invoice_title, invoice_date,
seller_taxer_code, shop_id, create_time) ;
create index idx_f_invoice_invoice_date2
on ins_dw_prd12.f_invoice (invoice_date);
create index idx_f_invoice_seller_taxer_code2
on ins_dw_prd12.f_invoice (seller_taxer_code);
create index idx_invoice_createtime_btree2
on ins_dw_prd12.f_invoice (create_time);
copy数据
-- 从原表拷贝数据
insert into f_invoice select * from f_invoice_2;
3、pg12 VS pg9.4
3.1、普通表对比
pg12
explain(analyse, timing)
SELECT count(*)
FROM (select *
from ins_dw_prd12.f_invoice_2 fi
where fi.seller_taxer_code in ('91320200704046760T', '91340100149067617J', '91320214MA1YGE8F94')
and fi.create_time >= '2019-08-01 00:00:00'
and fi.create_time <= '2019-09-01 00:00:00') m
INNER JOIN (select *
from ins_dw_prd12.f_invoice_item
where invoice_id in (SELECT fi.invoice_id
FROM ins_dw_prd12.f_invoice_2 fi
WHERE fi.seller_taxer_code in
('91320200704046760T', '91340100149067617J', '91320214MA1YGE8F94')
and fi.create_time >= '2019-08-01 00:00:00'
and fi.create_time <= '2019-09-01 00:00:00')) fit
ON fit.invoice_id = m.invoice_id;
执行计划：
Finalize Aggregate (cost=322631.88..322631.89 rows=1 width=8) (actual time=6265.586..6265.586 rows=1 loops=1)
-> Gather (cost=322631.46..322631.87 rows=4 width=8) (actual time=6264.473..6282.713 rows=5 loops=1)
Workers Planned: 4
Workers Launched: 4
-> Partial Aggregate (cost=321631.46..321631.47 rows=1 width=8) (actual time=6259.589..6259.589 rows=1 loops=5)
-> Nested Loop (cost=154762.41..321630.97 rows=194 width=0) (actual time=4100.919..6254.834 rows=38654 loops=5)
-> Parallel Hash Join (cost=154761.84..308914.48 rows=173 width=16) (actual time=4099.216..4355.187 rows=35911 loops=5)
Hash Cond: (fi.invoice_id = fi_1.invoice_id)
-> Parallel Index Scan using idx_invoice_createtime_btree on f_invoice_2 fi (cost=0.57..153991.56 rows=61577 width=8) (actual time=0.061..236.040 rows=35911 loops=5)
Index Cond: ((create_time >= '2019-08-01 00:00:00'::timestamp without time zone) AND (create_time <= '2019-09-01 00:00:00'::timestamp without time zone))
Filter: ((seller_taxer_code)::text = ANY ('{91320200704046760T,91340100149067617J,91320214MA1YGE8F94}'::text[]))
Rows Removed by Filter: 447251
-> Parallel Hash (cost=153991.56..153991.56 rows=61577 width=8) (actual time=4098.823..4098.823 rows=35911 loops=5)
Buckets: 262144 Batches: 1 Memory Usage: 9152kB
-> Parallel Index Scan using idx_invoice_createtime_btree on f_invoice_2 fi_1 (cost=0.57..153991.56 rows=61577 width=8) (actual time=1.356..4083.857 rows=35911 loops=5)
Index Cond: ((create_time >= '2019-08-01 00:00:00'::timestamp without time zone) AND (create_time <= '2019-09-01 00:00:00'::timestamp without time zone))
Filter: ((seller_taxer_code)::text = ANY ('{91320200704046760T,91340100149067617J,91320214MA1YGE8F94}'::text[]))
Rows Removed by Filter: 447251
-> Index Only Scan using f_invoice_item_invoice_id_idx on f_invoice_item (cost=0.57..71.17 rows=234 width=8) (actual time=0.052..0.052 rows=1 loops=179556)
Index Cond: (invoice_id = fi_1.invoice_id)
Heap Fetches: 193269
Planning Time: 0.596 ms
Execution Time: 6282.776 ms
pg9.4
explain(analyse, timing)
SELECT count(*)
FROM (select *
from ins_dw_prd12.f_invoice fi
where fi.seller_taxer_code in ('91320200704046760T', '91340100149067617J', '91320214MA1YGE8F94')
and fi.create_time >= '2019-08-01 00:00:00'
and fi.create_time <= '2019-09-01 00:00:00') m
INNER JOIN (select *
from ins_dw_prd12.f_invoice_item
where invoice_id in (SELECT fi.invoice_id
FROM ins_dw_prd12.f_invoice fi
WHERE fi.seller_taxer_code in
('91320200704046760T', '91340100149067617J', '91320214MA1YGE8F94')
and fi.create_time >= '2019-08-01 00:00:00'
and fi.create_time <= '2019-09-01 00:00:00')) fit
ON fit.invoice_id = m.invoice_id;
执行计划
Aggregate (cost=1583349.05..1583349.06 rows=1 width=0) (actual time=8706.357..8706.357 rows=1 loops=1)
-> Nested Loop (cost=801491.83..1582945.16 rows=161555 width=0) (actual time=6551.816..8689.740 rows=193269 loops=1)
-> Hash Join (cost=801491.27..1573243.48 rows=641 width=16) (actual time=6551.575..6953.476 rows=179556 loops=1)
Hash Cond: (fi.invoice_id = fi_1.invoice_id)
-> Bitmap Heap Scan on f_invoice fi (cost=15206.25..782402.63 rows=236591 width=8) (actual time=276.081..385.549 rows=179556 loops=1)
Recheck Cond: (((seller_taxer_code)::text = ANY ('{91320200704046760T,91340100149067617J,91320214MA1YGE8F94}'::text[])) AND (create_time >= '2019-08-01 00:00:00'::timestamp without time zone) AND (create_time <= '2019-09-01 00:00:00'::timestamp without time zone))
Heap Blocks: exact=49524
-> Bitmap Index Scan on f_invoice_seller_taxer_code_create_time_idx (cost=0.00..15147.10 rows=236591 width=0) (actual time=263.542..263.542 rows=179556 loops=1)
Index Cond: (((seller_taxer_code)::text = ANY ('{91320200704046760T,91340100149067617J,91320214MA1YGE8F94}'::text[])) AND (create_time >= '2019-08-01 00:00:00'::timestamp without time zone) AND (create_time <= '2019-09-01 00:00:00'::timestamp without time zone))
-> Hash (cost=782402.63..782402.63 rows=236591 width=8) (actual time=6274.711..6274.711 rows=179556 loops=1)
Buckets: 16384 Batches: 32 (originally 4) Memory Usage: 4097kB
-> Bitmap Heap Scan on f_invoice fi_1 (cost=15206.25..782402.63 rows=236591 width=8) (actual time=54.600..6113.136 rows=179556 loops=1)
Recheck Cond: (((seller_taxer_code)::text = ANY ('{91320200704046760T,91340100149067617J,91320214MA1YGE8F94}'::text[])) AND (create_time >= '2019-08-01 00:00:00'::timestamp without time zone) AND (create_time <= '2019-09-01 00:00:00'::timestamp without time zone))
Heap Blocks: exact=49524
-> Bitmap Index Scan on f_invoice_seller_taxer_code_create_time_idx (cost=0.00..15147.10 rows=236591 width=0) (actual time=43.760..43.760 rows=179556 loops=1)
Index Cond: (((seller_taxer_code)::text = ANY ('{91320200704046760T,91340100149067617J,91320214MA1YGE8F94}'::text[])) AND (create_time >= '2019-08-01 00:00:00'::timestamp without time zone) AND (create_time <= '2019-09-01 00:00:00'::timestamp without time zone))
-> Index Only Scan using f_invoice_item_invoice_id_idx on f_invoice_item (cost=0.57..12.62 rows=252 width=8) (actual time=0.009..0.009 rows=1 loops=179556)
Index Cond: (invoice_id = fi_1.invoice_id)
Heap Fetches: 162
Planning time: 0.871 ms
Execution time: 8706.466 ms
pg12版本耗时：6282.776 ms，pg9.4版本耗时：8706.466 ms，效率提升38%。
分析：
pg9.6以后的版本开始引入并行，这里测试的用例数据量和索引都是一样的，所以效率提升是由于pg12内部索引并行的优化。执行计划中：Workers Planned: 4，说明用了4个并行度，是由于pg12购买的机器只有8核，并且参数max_parallel_workers默认是4，还能通过核数来提升效率。
3.2、引入分区表
前期准备用已经把数据copy到f_invoice，并建好相关的表分区。
同样的sql
explain(analyse, timing)
SELECT count(*)
FROM (select *
from ins_dw_prd12.f_invoice fi
where fi.seller_taxer_code in ('91320200704046760T', '91340100149067617J', '91320214MA1YGE8F94')
and fi.create_time >= '2019-08-01 00:00:00'
and fi.create_time <= '2019-09-01 00:00:00') m
INNER JOIN (select *
from ins_dw_prd12.f_invoice_item
where invoice_id in (SELECT fi.invoice_id
FROM ins_dw_prd12.f_invoice fi
WHERE fi.seller_taxer_code in
('91320200704046760T', '91340100149067617J', '91320214MA1YGE8F94')
and fi.create_time >= '2019-08-01 00:00:00'
and fi.create_time <= '2019-09-01 00:00:00')) fit
ON fit.invoice_id = m.invoice_id;
执行计划
Finalize Aggregate (cost=2131432.05..2131432.06 rows=1 width=8) (actual time=7250.107..7250.107 rows=1 loops=1)
-> Gather (cost=2131431.83..2131432.04 rows=2 width=8) (actual time=7250.098..7262.130 rows=3 loops=1)
Workers Planned: 2
Workers Launched: 2
-> Partial Aggregate (cost=2130431.83..2130431.84 rows=1 width=8) (actual time=7241.645..7241.645 rows=1 loops=3)
-> Nested Loop (cost=99443.16..2111203.69 rows=7691256 width=0) (actual time=6911.989..7235.295 rows=64423 loops=3)
Join Filter: (fi_2.invoice_id = f_invoice_item.invoice_id)
-> Hash Join (cost=99442.60..197345.88 rows=37876 width=16) (actual time=6911.949..7017.101 rows=59852 loops=3)
Hash Cond: (fi.invoice_id = fi_2.invoice_id)
-> Parallel Append (cost=0.43..96862.13 rows=75752 width=8) (actual time=1.115..62.620 rows=59852 loops=3)
-> Parallel Index Scan using f_invoice_p201908_seller_taxer_code_idx on f_invoice_p201908 fi (cost=0.43..96480.72 rows=75751 width=8) (actual time=1.112..54.402 rows=59852 loops=3)
Index Cond: ((seller_taxer_code)::text = ANY ('{91320200704046760T,91340100149067617J,91320214MA1YGE8F94}'::text[]))
Filter: ((create_time >= '2019-08-01 00:00:00'::timestamp without time zone) AND (create_time <= '2019-09-01 00:00:00'::timestamp without time zone))
-> Parallel Index Scan using f_invoice_p201909_create_time_idx on f_invoice_p201909 fi_1 (cost=0.43..2.65 rows=1 width=8) (actual time=0.006..0.007 rows=0 loops=1)
Index Cond: ((create_time >= '2019-08-01 00:00:00'::timestamp without time zone) AND (create_time <= '2019-09-01 00:00:00'::timestamp without time zone))
Filter: ((seller_taxer_code)::text = ANY ('{91320200704046760T,91340100149067617J,91320214MA1YGE8F94}'::text[]))
-> Hash (cost=99439.67..99439.67 rows=200 width=8) (actual time=6910.770..6910.770 rows=179556 loops=3)
Buckets: 131072 (originally 1024) Batches: 4 (originally 1) Memory Usage: 3073kB
-> HashAggregate (cost=99437.67..99439.67 rows=200 width=8) (actual time=6835.596..6874.903 rows=179556 loops=3)
Group Key: fi_2.invoice_id
-> Append (cost=0.43..98983.16 rows=181803 width=8) (actual time=0.038..6746.552 rows=179556 loops=3)
-> Index Scan using f_invoice_p201908_seller_taxer_code_idx on f_invoice_p201908 fi_2 (cost=0.43..98071.49 rows=181802 width=8) (actual time=0.038..6718.573 rows=179556 loops=3)
Index Cond: ((seller_taxer_code)::text = ANY ('{91320200704046760T,91340100149067617J,91320214MA1YGE8F94}'::text[]))
Filter: ((create_time >= '2019-08-01 00:00:00'::timestamp without time zone) AND (create_time <= '2019-09-01 00:00:00'::timestamp without time zone))
-> Index Scan using f_invoice_p201909_create_time_idx on f_invoice_p201909 fi_3 (cost=0.43..2.65 rows=1 width=8) (actual time=0.016..0.017 rows=0 loops=3)
Index Cond: ((create_time >= '2019-08-01 00:00:00'::timestamp without time zone) AND (create_time <= '2019-09-01 00:00:00'::timestamp without time zone))
Filter: ((seller_taxer_code)::text = ANY ('{91320200704046760T,91340100149067617J,91320214MA1YGE8F94}'::text[]))
-> Index Only Scan using f_invoice_item_invoice_id_idx on f_invoice_item (cost=0.57..47.60 rows=234 width=8) (actual time=0.003..0.003 rows=1 loops=179556)
Index Cond: (invoice_id = fi.invoice_id)
Heap Fetches: 193269
Planning Time: 12.368 ms
Execution Time: 7264.656 ms
Workers Planned: 2，并行度只有2，可以明显看到只扫描了f_invoice_p201908和f_invoice_p201909上的索引，由于时间选的是08-01~09-01，所以f_invoice_p201909几乎不耗时，大部分时间在扫描f_invoice_p201908的索引。
Index Scan using f_invoice_p201908_seller_taxer_code_idx on f_invoice_p201908 fi_2 (cost=0.43..98071.49 rows=181802 width=8) (actual time=0.038..6718.573 rows=179556 loops=3)
理论上可以通过增加cpu核数和max_parallel_workers来提升分区表的效率。
后面用了不同月份尝试，并行度不提升（分区表在pg12上跟普通表性能差不多）
4、奇葩的发现
4.1、背景
1、数据量：
| 表名 | 数据量 |
|---|---|
| f_invoice | 87346130 |
| f_invoice_item | 97535867 |
2、索引：
表：f_invoice_item
CREATE INDEX f_invoice_item_order_item_id_idx ON ins_dw_prd12.f_invoice_item USING btree (order_item_id)
CREATE INDEX f_invoice_item_invoice_id_idx ON ins_dw_prd12.f_invoice_item USING btree (invoice_id) WITH (fillfactor='100')
表：f_invoice
CREATE INDEX idx_f_invoice_gin ON ins_dw_prd12.f_invoice USING gin (source_type, invoice_type, invoice_status, invoice_title, invoice_date, seller_taxer_code, shop_id, create_time)
CREATE INDEX idx_f_invoice_invoice_date ON ins_dw_prd12.f_invoice USING btree (invoice_date) WITH (fillfactor='100')
CREATE INDEX idx_f_invoice_seller_taxer_code ON ins_dw_prd12.f_invoice USING btree (seller_taxer_code) WITH (fillfactor='100')
CREATE INDEX idx_invoice_createtime_btree ON ins_dw_prd12.f_invoice USING btree (create_time) WITH (fillfactor='100')
4.2、奇葩
pg9.4和pg12的普通表
explain(analyse, timing)
SELECT count(*)
FROM (select *
from ins_dw_prd12.f_invoice_2 fi
where fi.seller_taxer_code in ('91320200704046760T', '91340100149067617J', '91320214MA1YGE8F94')
and fi.create_time >= '2019-09-01 00:00:00'
and fi.create_time <= '2019-10-01 00:00:00') m
INNER JOIN ins_dw_prd12.f_invoice_item fit
ON fit.invoice_id = m.invoice_id;
执行计划：
Finalize Aggregate (cost=2709869.86..2709869.87 rows=1 width=8) (actual time=83504.746..83504.746 rows=1 loops=1)
-> Gather (cost=2709869.03..2709869.84 rows=8 width=8) (actual time=83488.900..83509.668 rows=9 loops=1)
Workers Planned: 8
Workers Launched: 8
-> Partial Aggregate (cost=2708869.03..2708869.04 rows=1 width=8) (actual time=83478.769..83478.769 rows=1 loops=9)
-> Parallel Hash Join (cost=163813.42..2708776.96 rows=36830 width=0) (actual time=58876.039..83475.920 rows=27042 loops=9)
Hash Cond: (fit.invoice_id = fi.invoice_id)
-> Parallel Seq Scan on f_invoice_item fit (cost=0.00..2512888.52 rows=12219052 width=8) (actual time=0.122..80316.204 rows=10854295 loops=9)
-> Parallel Hash (cost=163154.14..163154.14 rows=52743 width=8) (actual time=252.113..252.113 rows=24461 loops=9)
Buckets: 524288 Batches: 1 Memory Usage: 12864kB
-> Parallel Index Scan using idx_invoice_createtime_btree on f_invoice_2 fi (cost=0.57..163154.14 rows=52743 width=8) (actual time=0.225..239.003 rows=24461 loops=9)
Index Cond: ((create_time >= '2019-09-01 00:00:00'::timestamp without time zone) AND (create_time <= '2019-10-01 00:00:00'::timestamp without time zone))
Filter: ((seller_taxer_code)::text = ANY ('{91320200704046760T,91340100149067617J,91320214MA1YGE8F94}'::text[]))
Rows Removed by Filter: 235679
Planning Time: 0.298 ms
Execution Time: 83509.737 ms
pg12分区表
explain(analyse, timing)
SELECT count(*)
FROM (select *
from ins_dw_prd12.f_invoice fi
where fi.seller_taxer_code in ('91320200704046760T', '91340100149067617J', '91320214MA1YGE8F94')
and fi.create_time >= '2019-09-01 00:00:00'
and fi.create_time <= '2019-10-01 00:00:00') m
INNER JOIN ins_dw_prd12.f_invoice_item fit
ON fit.invoice_id = m.invoice_id;
执行计划
Finalize Aggregate (cost=4539009.89..4539009.90 rows=1 width=8) (actual time=4089.118..4089.118 rows=1 loops=1)
-> Gather (cost=4539009.67..4539009.88 rows=2 width=8) (actual time=4088.801..4101.957 rows=3 loops=1)
Workers Planned: 2
Workers Launched: 2
-> Partial Aggregate (cost=4538009.67..4538009.68 rows=1 width=8) (actual time=4085.523..4085.524 rows=1 loops=3)
-> Nested Loop (cost=1.00..4483735.31 rows=21709746 width=0) (actual time=2.101..4076.649 rows=81125 loops=3)
-> Parallel Append (cost=0.43..106404.25 rows=92916 width=8) (actual time=1.383..2078.226 rows=73384 loops=3)
-> Parallel Index Scan using f_invoice_p201909_seller_taxer_code_idx on f_invoice_p201909 fi (cost=0.43..105937.02 rows=92915 width=8) (actual time=1.380..2066.570 rows=73384 loops=3)
Index Cond: ((seller_taxer_code)::text = ANY ('{91320200704046760T,91340100149067617J,91320214MA1YGE8F94}'::text[]))
Filter: ((create_time >= '2019-09-01 00:00:00'::timestamp without time zone) AND (create_time <= '2019-10-01 00:00:00'::timestamp without time zone))
-> Parallel Index Scan using f_invoice_p201910_create_time_idx on f_invoice_p201910 fi_1 (cost=0.43..2.65 rows=1 width=8) (actual time=0.007..0.007 rows=0 loops=1)
Index Cond: ((create_time >= '2019-09-01 00:00:00'::timestamp without time zone) AND (create_time <= '2019-10-01 00:00:00'::timestamp without time zone))
Filter: ((seller_taxer_code)::text = ANY ('{91320200704046760T,91340100149067617J,91320214MA1YGE8F94}'::text[]))
-> Index Only Scan using f_invoice_item_invoice_id_idx on f_invoice_item fit (cost=0.57..44.77 rows=234 width=8) (actual time=0.026..0.027 rows=1 loops=220151)
Index Cond: (invoice_id = fi.invoice_id)
Heap Fetches: 243376
Planning Time: 6.910 ms
Execution Time: 4102.010 ms
分析
理论上来讲：f_invoice的子查询join在关联字段上有索引（invoice_id）的大表f_invoice_item，会以f_invoice为驱动表，f_invoice_item会通过索引来扫描，但是事实是走了全表扫描。
但是在分区表中：并行扫描了f_invoice_p201909和f_invoice_p201910的索引后，再Parallel Append，最后跟f_invoice_item来Nested Loop。此时可以看到，f_invoice_item是走了Index Only Scan（只扫了索引，不用会表）
两者性能比较：pg12分区表：4102.010 ms，pg12普通表:83509.737 ms。
5、总体结论
可以通过升级pg12，购买8核的实例，利用pg12自带的并行算法来提升性能。预计可以提升30~50%，如果购买16核的实例，把并行度打开，引入分区表，估计还能提升更大。