请勿滥用unlogged table & hash index

Postgres2015全国用户大会将于11月20至21日在北京丽亭华苑酒店召开。本次大会嘉宾阵容强大，国内顶级PostgreSQL数据库专家将悉数到场，并特邀欧洲、俄罗斯、日本、美国等国家和地区的数据库方面专家助阵:
Postgres-XC项目的发起人铃木市一(SUZUKI Koichi)
Postgres-XL的项目发起人Mason Sharp
pgpool的作者石井达夫(Tatsuo Ishii)
PG-Strom的作者海外浩平(Kaigai Kohei)
Greenplum研发总监姚延栋
周正中(德哥), PostgreSQL中国用户会创始人之一
汪洋，平安科技数据库技术部经理
……

2015年度PG大象会报名地址：http://postgres2015.eventdove.com/PostgreSQL中国社区： http://postgres.cn/PostgreSQL专业1群： 3336901（已满）PostgreSQL专业2群： 100910388PostgreSQL专业3群： 150657323

unlogged table和hash index同样都不会写XLOG，所以如果你用流复制来搞HA，一定概要搞清楚一个问题，切换到备库的话unlogged table数据会被清掉，而hash index也没有，走hash index会失败。

unlogged table 的风险以及修复手段可以见 ：
http://blog.163.com/digoal@126/blog/static/163877040201582621345351/

hash index则风险略小，但是也必须重建，但是个人还是建议大家不要使用hash index，改用btree，因为性能确实相差无几。

批量将数据库集群的hash index修改为btree index的方法：

例子：

先并行创建一个btree索引，然后并行删除对应的hash 索引。

$ vi test.sh#!/bin/bash
for db in `psql -n -q -t -h 127.0.0.1 -p 1921 -U postgres postgres -c "copy (select datname from pg_database where datname <>'template0') to stdout;"`do 
psql -n -q -t -h 127.0.0.1 -p 1921 -U postgres $db -c "with t1(sql,nsp,idx) as (select regexp_replace(indexdef,'USING hash','USING btree'),schemaname,indexname from pg_indexes where indexdef ~ 'USING hash'), t2(sql_create,sql_drop) as (select regexp_replace(sql,'CREATE INDEX','CREATE INDEX CONCURRENTLY'), 'DROP INDEX CONCURRENTLY '||quote_ident(nsp)||'.'||quote_ident(idx) from t1) select regexp_replace(sql_create,'CONCURRENTLY (.*) ON','CONCURRENTLY \1_0926 ON')  ||'; '|| sql_drop ||'; ' from t2;"|psql -a -e -h 127.0.0.1 -p 1921 -U postgres $db -f -
done
$ . ./test.sh
这个with查询的结果如下举例：postgres=# with t1(sql,nsp,idx) as (select regexp_replace(indexdef,'USING hash','USING btree'),schemaname,indexname from pg_indexes where indexdef ~ 'USING hash'),t2(sql_create,sql_drop) as (select regexp_replace(sql,'CREATE INDEX','CREATE INDEX CONCURRENTLY'), 'DROP INDEX CONCURRENTLY '||quote_ident(nsp)||'.'||quote_ident(idx) from t1)select regexp_replace(sql_create,'CONCURRENTLY (.*) ON','CONCURRENTLY \1_0926 ON')  ||'; '|| sql_drop ||'; ' from t2;                                            ?column?                                             ------------------------------------------------------------------------------------------------- CREATE INDEX CONCURRENTLY hi1_0926 ON t USING btree (id); DROP INDEX CONCURRENTLY public.hi1;  CREATE INDEX CONCURRENTLY hi2_0926 ON s1.tbl USING btree (id); DROP INDEX CONCURRENTLY s1.hi2; (2 rows)

或者在每个数据库调用这个inline code：

do language plpgsql $$declare  v_sql text;   v_schema name;   v_idx name;   sql1 text;beginfor v_sql,v_schema,v_idx in select regexp_replace(indexdef,'USING hash','USING btree'),schemaname,indexname from pg_indexes where indexdef ~ 'USING hash'loop  sql1='DROP INDEX '||quote_ident(v_schema)||'.'||quote_ident(v_idx);   execute sql1;   execute v_sql; end loop; end; $$; 
postgres=# \d t       Table "public.t" Column |  Type   | Modifiers --------+---------+----------- id     | integer | not nullIndexes:    "t_pkey" PRIMARY KEY, btree (id)    "i1" btree (id)    "i2" btree (id)
postgres=# \d s1.tbl        Table "s1.tbl" Column |  Type   | Modifiers --------+---------+----------- id     | integer | Indexes:    "i1" btree (id)

关于hash和btree性能：
查询性能：

postgres=# create table tbl(id int, info text);CREATE TABLEpostgres=# insert into tbl select generate_series(1,1000000);INSERT 0 1000000postgres=# create index idx_tbl1 on tbl using hash (id);
$ vi test.sql\setrandom id 1 1000000select * from tbl where id=:id;
postgres@digoal-> pgbench -M prepared -n -r -P 1 -f ./test.sql -c 8 -j 8 -T 30progress: 1.0 s, 24219.3 tps, lat 0.258 ms stddev 0.436progress: 2.0 s, 29387.1 tps, lat 0.270 ms stddev 0.401progress: 3.0 s, 29281.0 tps, lat 0.271 ms stddev 0.442progress: 4.0 s, 29231.7 tps, lat 0.272 ms stddev 0.844......transaction type: Custom queryscaling factor: 1query mode: preparednumber of clients: 8number of threads: 8duration: 30 snumber of transactions actually processed: 876806latency average: 0.270 mslatency stddev: 0.592 mstps = 29202.503991 (including connections establishing)tps = 29379.956438 (excluding connections establishing)statement latencies in milliseconds:        0.003062        \setrandom id 1 1000000        0.266481        select * from tbl where id=:id;
postgres=# drop index idx_tbl1 ;DROP INDEXpostgres=# create index idx_tbl1 on tbl using btree (id);CREATE INDEX
postgres@digoal-> pgbench -M prepared -n -r -P 1 -f ./test.sql -c 8 -j 8 -T 30progress: 1.0 s, 28414.2 tps, lat 0.240 ms stddev 0.306progress: 2.0 s, 31192.2 tps, lat 0.255 ms stddev 0.605progress: 3.0 s, 31022.8 tps, lat 0.256 ms stddev 0.451progress: 4.0 s, 29587.1 tps, lat 0.268 ms stddev 0.671......transaction type: Custom queryscaling factor: 1query mode: preparednumber of clients: 8number of threads: 8duration: 30 snumber of transactions actually processed: 903467latency average: 0.263 mslatency stddev: 0.678 mstps = 30088.054150 (including connections establishing)tps = 30229.295069 (excluding connections establishing)statement latencies in milliseconds:        0.002900        \setrandom id 1 1000000        0.259402        select * from tbl where id=:id;

更新性能

$ vi test.sql\setrandom id 1 1000000update tbl set id=1+:id where id=:id;
postgres@digoal-> pgbench -M prepared -n -r -P 1 -f ./test.sql -c 8 -j 8 -T 30progress: 1.0 s, 12500.2 tps, lat 0.570 ms stddev 0.864progress: 2.0 s, 17456.9 tps, lat 0.456 ms stddev 0.641progress: 3.0 s, 18242.3 tps, lat 0.435 ms stddev 0.234progress: 4.0 s, 17693.0 tps, lat 0.450 ms stddev 0.909progress: 5.0 s, 17753.3 tps, lat 0.448 ms stddev 0.758......transaction type: Custom queryscaling factor: 1query mode: preparednumber of clients: 8number of threads: 8duration: 30 snumber of transactions actually processed: 521331latency average: 0.456 mslatency stddev: 0.702 mstps = 17372.386945 (including connections establishing)tps = 17430.542432 (excluding connections establishing)statement latencies in milliseconds:        0.003072        \setrandom id 1 1000000        0.452489        update tbl set id=1+:id where id=:id;
postgres=# drop index idx_tbl1 ;DROP INDEXpostgres=# create index idx_tbl1 on tbl using hash (id);CREATE INDEX
postgres@digoal-> pgbench -M prepared -n -r -P 1 -f ./test.sql -c 8 -j 8 -T 30progress: 1.0 s, 16321.8 tps, lat 0.411 ms stddev 0.521progress: 2.0 s, 17372.0 tps, lat 0.458 ms stddev 0.409progress: 3.0 s, 16731.5 tps, lat 0.475 ms stddev 1.094progress: 4.0 s, 16972.9 tps, lat 0.469 ms stddev 0.880progress: 5.0 s, 17392.7 tps, lat 0.457 ms stddev 0.607......transaction type: Custom queryscaling factor: 1query mode: preparednumber of clients: 8number of threads: 8duration: 30 snumber of transactions actually processed: 527778latency average: 0.450 mslatency stddev: 0.678 mstps = 17587.300360 (including connections establishing)tps = 17671.181609 (excluding connections establishing)statement latencies in milliseconds:        0.002955        \setrandom id 1 1000000        0.446435        update tbl set id=1+:id where id=:id;

SIZE

postgres=# create index idx_tbl1 on tbl using hash (id);CREATE INDEXpostgres=# create index idx_tbl2 on tbl using btree (id);CREATE INDEXpostgres=# \di+ idx_tbl*                         List of relations Schema |   Name   | Type  |  Owner   | Table | Size  | Description --------+----------+-------+----------+-------+-------+------------- public | idx_tbl1 | index | postgres | tbl   | 32 MB |  public | idx_tbl2 | index | postgres | tbl   | 21 MB | (2 rows)

小结:
查询和更新性能相差无几。
大小，btree是hash的2/3。