PostgreSQL的"天气预报" - 如何预测Freeze IO风暴-CSDN博客

还记得我写的这篇文档吗？《PostgreSQL 大表自动 freeze 优化思路》
https://yq.aliyun.com/articles/50411
文章主要针对如何优化大表的freeze调度来减少IO风暴的问题，请注意只是减少，不是避免。

作为一名有追求的PGer，要时刻保持警惕，生于忧患、死于安乐；
本文要给大家讲的是预测风暴，掌握了预测能力，才能未雨绸缪，淡定的面对暴风雨。

预测 IO 风暴

如何预测此类(prevent wrapped vacuum freeze) IO 风暴的来临呢？
首先需要测量几个维度的值。
.1. 表的大小以及距离它需要被强制vacuum freeze prevent wrap的年龄

.2. 每隔一段时间的XID值的采样（例如每分钟一次），采样越多越好，因为需要用于预测下一个时间窗口的XID。（其实就是每分钟消耗多少个事务号的数据）

.3. 通过第二步得到的结果，预测下一个时间窗口的每分钟的pXID（可以使用线性回归来进行预测）
预测方法这里不在细说，也可以参考我以前写的一些预测类的文章。

预测的结论包括未来一段时间的总Freeze IO量，以及分时的Freeze IO量。
预测结果范例
Freeze IO 时段总量
screenshot
Freeze IO 分时走势

详细的预测过程

.1.
每隔一段时间的XID值的采样（例如每分钟一次），采样越多越好，因为需要用于预测下一个时间窗口的XID。（其实就是每分钟消耗多少个事务号的数据）

vi xids.sh
#!/bin/bash
export PATH=/home/digoal/pgsql9.5/bin:$PATH
export PGHOST=127.0.0.1
export PGPORT=1921
export PGDATABASE=postgres
export PGUSER=postgres
export PGPASSWORD=postgres

psql -c "create table xids(crt_time timestamp, xids int8)"
for ((i=1;i>0;))
do
# 保留1个月的数据
psql -c "with a as (select ctid from xids order by crt_time desc limit 100 offset 43200) delete from xids where ctid in (select ctid from a);"
psql -c "insert into xids values (now(), txid_current());"
sleep 60
done

chmod 500 xids.sh

nohup ./xids.sh >/dev/null 2>&1 &

采集1天的数据可能是这样的

postgres=# select * from xids ;
          crt_time          | xids 
----------------------------+------
 2016-06-12 12:36:13.201315 | 2020
 2016-06-12 12:37:13.216002 | 9021
 2016-06-12 12:38:13.240739 | 21022
 2016-06-12 12:39:13.259203 | 32023
 2016-06-12 12:40:13.300604 | 42024
 2016-06-12 12:41:13.325874 | 52025
 2016-06-12 12:42:13.361152 | 62026
 2016-06-12 12:43:15.481609 | 72027
...

.2.
表的大小以及距离它需要被强制vacuum freeze prevent wrap的年龄

 (因为freeze是全集群的，所以需要把所有库得到的数据汇总到一起)

vi pred_io.sh

#!/bin/bash
export PATH=/home/digoal/pgsql9.5/bin:$PATH
export PGHOST=127.0.0.1
export PGPORT=1921
export PGDATABASE=postgres
export PGUSER=postgres
export PGPASSWORD=postgres

psql -c "drop table pred_io; create table pred_io(crt_time timestamp, bytes int8, left_live int8);"
for db in `psql -A -t -q -c "select datname from pg_database where datname <> 'template0'"`
do
psql -d $db -c " copy (
select now(), bytes, case when max_age>age then max_age-age else 0 end as xids from 
(select block_size*relpages bytes, 
case when d_max_age is not null and d_max_age<max_age then d_max_age else max_age end as max_age,
age from
(select 
(select setting from pg_settings where name='block_size')::int8 as block_size, 
(select setting from pg_settings where name='autovacuum_freeze_max_age')::int8 as max_age, 
relpages, 
substring(reloptions::text,'autovacuum_freeze_max_age=(\d+)')::int8 as d_max_age,
age(relfrozenxid) age
from pg_class where relkind in ('r', 't')) t) t
) to stdout;" | psql -d $PGDATABASE -c "copy pred_io from stdin"
done

. ./pred_io.sh

得到的数据可能是这样的

postgres=# select * from pred_io limit 10;
          crt_time          | bytes  | left_live 
----------------------------+--------+-----------
 2016-06-12 13:24:08.666995 | 131072 | 199999672
 2016-06-12 13:24:08.666995 |  65536 | 199999672
 2016-06-12 13:24:08.666995 |      0 | 199999672
 2016-06-12 13:24:08.666995 |      0 | 199999672
 2016-06-12 13:24:08.666995 |      0 | 199999672
 2016-06-12 13:24:08.666995 |      0 | 199999672
...

.3.
预测XIDs走势(略)，本文直接取昨天的同一时间点开始后的数据。

create view v_pred_xids as 
with b as (select min(crt_time) tbase from pred_io), 
       a as (select crt_time + interval '1 day' as crt_time, xids from xids,b where crt_time >= b.tbase - interval '1 day')  
select crt_time, xids - (select min(xids) from a) as xids from a ;

数据可能是这样的，预测未来分时的相对XIDs消耗量

          crt_time          | xids 
----------------------------+------
 2016-06-13 12:36:13.201315 |    0
 2016-06-13 12:37:13.216002 |    100
 2016-06-13 12:38:13.240739 |    200
 2016-06-13 12:39:13.259203 |    300
 2016-06-13 12:40:13.300604 |    400

.4.
结合pred_io与v_pred_xids 进行 io风暴预测
基准视图，后面的数据通过这个基准视图得到

create view pred_tbased_io as
with a as (select crt_time, xids as s_xids, lead(xids) over(order by crt_time) as e_xids from v_pred_xids)
select a.crt_time, sum(b.bytes) bytes from a, pred_io b where b.left_live >=a.s_xids and b.left_live < a.e_xids group by a.crt_time order by a.crt_time;

未来一天的总freeze io bytes预测

postgres=# select min(crt_time),max(crt_time),sum(bytes) from pred_tbased_io ;
            min             |            max             |   sum    
----------------------------+----------------------------+----------
 2016-06-13 12:36:13.201315 | 2016-06-14 12:35:26.104025 | 19685376
(1 row)

未来一天的freeze io bytes分时走势
得到的结果可能是这样的

postgres=# select * from pred_tbased_io ;
          crt_time          |  bytes   
----------------------------+----------
 2016-06-13 12:36:13.201315 |    65536
 2016-06-13 12:37:13.216002 |   581632
 2016-06-13 12:38:13.240739 |        0
 2016-06-13 12:39:13.259203 |        0
 2016-06-13 12:40:13.300604 |        0
 2016-06-13 12:41:13.325874 |        0
 2016-06-13 12:43:15.481609 |   106496
 2016-06-13 12:43:24.133055 |     8192
 2016-06-13 12:45:24.193318 |        0
 2016-06-13 12:46:24.225559 |    16384
 2016-06-13 12:48:24.296223 | 13434880
 2016-06-13 12:49:24.325652 |    24576
 2016-06-13 12:50:24.367232 |   401408
 2016-06-13 12:51:24.426199 |        0
 2016-06-13 12:52:24.457375 |   393216
......