PostgreSQL DBA(139) - PG 12（B-tree index improvement 1#）

本節簡單介紹了PostgreSQL 12 B-tree的改進：索引出現很多重複值時提升效能，減少空間佔用。

PG 11
建立資料表，建立索引

[local]:5110 xdb@testdb=# drop table rel;
DROP TABLE
Time: 130.868 ms
[local]:5110 xdb@testdb=# CREATE TABLE rel (
xdb@testdb(#    aid bigint NOT NULL,
xdb@testdb(#    bid bigint NOT NULL
xdb@testdb(# );
CREATE TABLE
Time: 16.041 ms
[local]:5110 xdb@testdb=#  
[local]:5110 xdb@testdb=# ALTER TABLE rel
xdb@testdb-#    ADD CONSTRAINT rel_pkey PRIMARY KEY (aid, bid);
ALTER TABLE
Time: 5.236 ms
[local]:5110 xdb@testdb=#  
[local]:5110 xdb@testdb=# CREATE INDEX rel_bid_idx ON rel (bid);
CREATE INDEX
Time: 1.838 ms
[local]:5110 xdb@testdb=#  
[local]:5110 xdb@testdb=# INSERT INTO rel (aid, bid)
xdb@testdb-#    SELECT i, i / 10000
xdb@testdb-#    FROM generate_series(1, 20000000) AS i; 
INSERT 0 20000000
Time: 152699.275 ms (02:32.699)
[local]:5110 xdb@testdb=# 
[local]:5110 xdb@testdb=#

檢視索引資訊

[local]:5110 xdb@testdb=# 
[local]:5110 xdb@testdb=# \d rel
                Table "public.rel"
 Column |  Type  | Collation | Nullable | Default 
--------+--------+-----------+----------+---------
 aid    | bigint |           | not null | 
 bid    | bigint |           | not null | 
Indexes:
    "rel_pkey" PRIMARY KEY, btree (aid, bid)
    "rel_bid_idx" btree (bid)
[local]:5110 xdb@testdb=# \di+ rel_pkey
                        List of relations
 Schema |   Name   | Type  | Owner | Table |  Size  | Description 
--------+----------+-------+-------+-------+--------+-------------
 public | rel_pkey | index | xdb   | rel   | 602 MB | 
(1 row)
[local]:5110 xdb@testdb=# \di+ rel_bid_idx
                          List of relations
 Schema |    Name     | Type  | Owner | Table |  Size  | Description 
--------+-------------+-------+-------+-------+--------+-------------
 public | rel_bid_idx | index | xdb   | rel   | 545 MB | 
(1 row)

PG 12
建立資料表，建立索引

[local:/run/pg12]:5120 pg12@testdb=# \timing on
Timing is on.
[local:/run/pg12]:5120 pg12@testdb=# drop table rel;
DROP TABLE
Time: 279.144 ms
[local:/run/pg12]:5120 pg12@testdb=# CREATE TABLE rel (
pg12@testdb(#    aid bigint NOT NULL,
pg12@testdb(#    bid bigint NOT NULL
pg12@testdb(# );
CREATE TABLE
Time: 1.579 ms
[local:/run/pg12]:5120 pg12@testdb=#  
[local:/run/pg12]:5120 pg12@testdb=# ALTER TABLE rel
pg12@testdb-#    ADD CONSTRAINT rel_pkey PRIMARY KEY (aid, bid);
ALTER TABLE
Time: 3.450 ms
[local:/run/pg12]:5120 pg12@testdb=#  
[local:/run/pg12]:5120 pg12@testdb=# CREATE INDEX rel_bid_idx ON rel (bid);
CREATE INDEX
Time: 1.201 ms
[local:/run/pg12]:5120 pg12@testdb=#  
[local:/run/pg12]:5120 pg12@testdb=# INSERT INTO rel (aid, bid)
pg12@testdb-#    SELECT i, i / 10000
pg12@testdb-#    FROM generate_series(1, 20000000) AS i; 
INSERT 0 20000000
Time: 124503.212 ms (02:04.503)
[local:/run/pg12]:5120 pg12@testdb=#

檢視索引資訊

[local:/run/pg12]:5120 pg12@testdb=# \di+ rel_pkey
                        List of relations
 Schema |   Name   | Type  | Owner | Table |  Size  | Description 
--------+----------+-------+-------+-------+--------+-------------
 public | rel_pkey | index | pg12  | rel   | 601 MB | 
(1 row)
[local:/run/pg12]:5120 pg12@testdb=# \di+ rel_bid_idx
                          List of relations
 Schema |    Name     | Type  | Owner | Table |  Size  | Description 
--------+-------------+-------+-------+-------+--------+-------------
 public | rel_bid_idx | index | pg12  | rel   | 408 MB | 
(1 row)
[local:/run/pg12]:5120 pg12@testdb=#

可以看到PK沒有太大的變化，但有很多重複值的bid列索引則有明顯的變化，比PG 11少了25%的空間。

原理
PG 11 vs PG 12

從上面兩個圖可以看出，PG 11的索引leaf page發生在middle，而PG 12發生在rightmost，middle葉子節點的page稠密度明顯要比PG 11好很多。

參考資料

Make heap TID a tiebreaker nbtree index column.