PostgreSQL , 大屏指标 , qps , long query , locks , active , idle in transaction , long idle in transaction , 2PC
最关键的一些数据库健康指标,趋势监测。
主要看趋势,直接与业务量挂钩
如果连接数接近max_connection水位,需要注意。
同时连接数应与数据库主机可用内存挂钩,每个连接保守估计10MB内存开销(这里还未计算SYSCACHE,RELCACHE)。
select count(*) from pg_stat_activity ;
演示,打印每秒的总连接数。
psql
select count(*) from pg_stat_activity ;
\watch 1
主要看趋势,直接与业务量挂钩
如果突发大量连接,可能是新增了业务服务器,或者是性能抖动过导致业务大量新建连接满足并发的请求。
突然连接数下降,可能原因是业务服务器突然释放连接,或者业务服务器挂了。
select count(*) from pg_stat_activity where now()-backend_start > '? second';
演示,打印每秒的5秒内新建连接数。
psql
select count(*) from pg_stat_activity where now()-backend_start > '5 second';
\watch 1
1、需要加载pg_stat_statements,如果需要跟踪IO时间,需要开启track_io_timing。
同时需要注意,由于pg_stat_statements跟踪的SQL有限,最近未访问过的SQL的跟踪信息可能被抛弃。所以统计并不是非常的精准。
postgres=# \d pg_stat_statements
View "public.pg_stat_statements"
Column | Type | Collation | Nullable | Default
---------------------+------------------+-----------+----------+---------
userid | oid | | |
dbid | oid | | |
queryid | bigint | | |
query | text | | |
calls | bigint | | |
total_time | double precision | | |
min_time | double precision | | |
max_time | double precision | | |
mean_time | double precision | | |
stddev_time | double precision | | |
rows | bigint | | |
shared_blks_hit | bigint | | |
shared_blks_read | bigint | | |
shared_blks_dirtied | bigint | | |
shared_blks_written | bigint | | |
local_blks_hit | bigint | | |
local_blks_read | bigint | | |
local_blks_dirtied | bigint | | |
local_blks_written | bigint | | |
temp_blks_read | bigint | | |
temp_blks_written | bigint | | |
blk_read_time | double precision | | |
blk_write_time | double precision | | |
QPS指标来自pg_stat_statements,由于这个插件有一个STATEMENT采集上限,可配置,例如最多采集1000条SQL,如果有新的SQL被采集到时,并且1000已用完,则会踢掉最老的SQL。所以我们这里统计的QPS并不是完全精确,不过还好PG内部会自动合并SQL,把一些条件替换成变量,这样即使不使用绑定变量,也能追踪到很多SQL。
对于业务SQL非常繁多并且大多数都是活跃SQL的场景,可以适当调大pg_stat_statements的track数,提高精准度。
除此之外,可以改进pg_stat_statements的功能,直接统计精准的QPS。
主要看趋势,直接与业务量挂钩
with
a as (select sum(calls) s, sum(case when ltrim(query,' ') ~* '^select' then calls else 0 end) q from pg_stat_statements),
b as (select sum(calls) s, sum(case when ltrim(query,' ') ~* '^select' then calls else 0 end) q from pg_stat_statements , pg_sleep(1))
select
b.s-a.s, -- QPS
b.q-a.q, -- 读QPS
b.s-b.q-a.s+a.q -- 写QPS
from a,b;
如果只想看QPS,使用
with
a as (select sum(calls) s from pg_stat_statements),
b as (select sum(calls) s from pg_stat_statements , pg_sleep(1))
select
b.s-a.s -- QPS
from a,b;
演示,打印每秒的QPS。
psql
with
a as (select sum(calls) s from pg_stat_statements),
b as (select sum(calls) s from pg_stat_statements , pg_sleep(1))
select
b.s-a.s -- QPS
from a,b;
\watch 0.000001
每秒处理了多少行,包括写入,读取,更新,删除等操作。
两次快照相减除以时间间隔
sum(pg_stat_statements.rows)
shared_blks_hit | bigint | | |
shared_blks_read | bigint | | |
shared_blks_dirtied | bigint | | |
shared_blks_written | bigint | | |
local_blks_hit | bigint | | |
local_blks_read | bigint | | |
local_blks_dirtied | bigint | | |
local_blks_written | bigint | | |
temp_blks_read | bigint | | |
temp_blks_written | bigint | | |
blk_read_time | double precision | | |
blk_write_time | double precision | | |
主要看趋势,直接与业务量挂钩
如果活跃会话数长时间超过CPU核数时,说明数据库响应变慢了,需要深刻关注。
select count(*) from pg_stat_activity where state='active';
演示,打印每秒的活跃会话数。
psql
select count(*) from pg_stat_activity where state='active';
\watch 1
活跃会话/qps = RT(秒)
当前系统中执行时间超过N秒的SQL有多少条,LONG QUERY与活跃会话的比例说明当前LONG SQL的占比。占比越高,说明该系统可能偏向OLAP,占比越低,说明该系统偏向OLTP业务。
select count(*) from pg_stat_activity where state='active' and now()-query_start > interval '? second';
演示,打印每秒系统中执行时间超过5秒的SQL有多少条。
psql
select count(*) from pg_stat_activity where state='active' and now()-query_start > interval '5 second';
\watch 1
当前系统中N秒未结束的事务有多少条
select count(*) from pg_stat_activity where now()-xact_start > interval '? second';
演示,打印每秒系统中5秒未结束的事务有多少条
psql
select count(*) from pg_stat_activity where now()-xact_start > interval '5 second';
\watch 1
当前系统中在事务中并且处于空闲状态的会话有多少,很多,说明业务端的处理可能比较慢,如果结合锁等待发现有大量锁等待,并且活跃会话数有突增,可能需要关注并排查业务逻辑的问题。
select count(*) from pg_stat_activity where state='idle in transaction';
演示,打印每秒系统中在事务中并且处于空闲状态的会话有多少
psql
select count(*) from pg_stat_activity where state='idle in transaction';
\watch 1
当前系统中,有多少长期(超过N秒)处于空闲的事务。如果有较多这样的事务,说明业务端的处理时间超过N秒的情况非常普遍,应该尽快排查业务。
比如前端开启了游标,等待用户的翻页动作,用户可能开小差了。又比如业务上使用了一些交互模式,等用户的一些输入等。
这种情况应该尽量避免,否则长时间占用连接资源。
select count(*) from pg_stat_activity where state='idle in transaction' and now()-state_change > interval '? second';
演示,打印每秒系统中在事务中并且处于空闲状态(超过5秒)的会话有多少
psql
select count(*) from pg_stat_activity where state='idle in transaction' and now()-state_change > interval '5 second';
\watch 1
当前系统中,处于等待中的会话有多少。
如果很多,说明出现了大量的锁等待,使用末尾文章进行排查。
select count(*) from pg_stat_activity where wait_event_type is not null;
演示,打印每秒系统中处于等待中的会话有多少。
psql
select count(*) from pg_stat_activity where wait_event_type is not null;
\watch 1
当前系统中,等待超过N秒的会话有多少。
select count(*) from pg_stat_activity where wait_event_type is not null and now()-state_change > interval '? second';
演示,打印每秒系统中等待超过5秒的会话有多少。
psql
select count(*) from pg_stat_activity where wait_event_type is not null and now()-state_change > interval '5 second';
\watch 1
当前系统中,2PC的事务有多少。如果接近max_prepared_transactions,需要注意。建议调大max_prepared_transactions,或者排查业务是否未及时提交。
select count(*) from pg_prepared_xacts;
演示,打印每秒系统中未结束的2PC事务数。
psql
select count(*) from pg_prepared_xacts;
\watch 1
当前系统中,超过N秒未结束的2PC的事务有多少。如果很多,需要排查业务为什么未及时提交。
select count(*) from pg_prepared_xacts where now() - prepared > interval '? second';
演示,打印每秒系统中5秒仍未结束的2PC事务数。
psql
select count(*) from pg_prepared_xacts where now() - prepared > interval '5 second';
\watch 1
时间间隔越大,说明越容易导致膨胀。
排查这几个方向,长事务,长SQL,2PC,持有SNAPSHOT的QUERY。必要时把不合理的老的会话干掉。
with a as
(select min(xact_start) m from pg_stat_activity where backend_xid is not null or backend_xmin is not null),
b as (select min(prepared) m from pg_prepared_xacts)
select now()-least(a.m,b.m) from a,b;
演示,打印每秒系统中多久以前的垃圾可以被回收
psql
with a as
(select min(xact_start) m from pg_stat_activity where backend_xid is not null or backend_xmin is not null),
b as (select min(prepared) m from pg_prepared_xacts)
select now()-least(a.m,b.m) from a,b;
\watch 1
看当前占用情况,打快照,看时间维度空间变化情况。
按库划分
postgres=# \l+
List of databases
Name | Owner | Encoding | Collate | Ctype | Access privileges | Size | Tablespace | Description
-----------+----------+-----------+------------+------------+-----------------------+---------+------------+--------------------------------------------
postgres | postgres | SQL_ASCII | en_US.UTF8 | en_US.UTF8 | | 54 GB | pg_default | default administrative connection database
template0 | postgres | SQL_ASCII | en_US.UTF8 | en_US.UTF8 | =c/postgres +| 7489 kB | pg_default | unmodifiable empty database
| | | | | postgres=CTc/postgres | | |
template1 | postgres | SQL_ASCII | en_US.UTF8 | en_US.UTF8 | =c/postgres +| 578 MB | pg_default | default template for new databases
| | | | | postgres=CTc/postgres | | |
test | test | SQL_ASCII | en_US.UTF8 | en_US.UTF8 | | 7489 kB | pg_default |
(4 rows)
按表空间划分
postgres=# \db+
List of tablespaces
Name | Owner | Location | Access privileges | Options | Size | Description
--------------------+----------+--------------------------------------+-------------------+---------+---------+-------------
dbt2_index1 | postgres | /data02/pg/tbs_tpcc/index1/ts | | | 452 MB |
dbt2_index2 | postgres | /data02/pg/tbs_tpcc/index2/ts | | | 869 MB |
dbt2_pk_customer | postgres | /data02/pg/tbs_tpcc/pk_customer/ts | | | 451 MB |
dbt2_pk_district | postgres | /data02/pg/tbs_tpcc/pk_district/ts | | | 236 kB |
dbt2_pk_item | postgres | /data02/pg/tbs_tpcc/pk_item/ts | | | 2212 kB |
dbt2_pk_new_order | postgres | /data02/pg/tbs_tpcc/pk_new_order/ts | | | 149 MB |
dbt2_pk_order_line | postgres | /data02/pg/tbs_tpcc/pk_order_line/ts | | | 4701 MB |
dbt2_pk_orders | postgres | /data02/pg/tbs_tpcc/pk_orders/ts | | | 490 MB |
dbt2_pk_stock | postgres | /data02/pg/tbs_tpcc/pk_stock/ts | | | 1768 MB |
dbt2_pk_warehouse | postgres | /data02/pg/tbs_tpcc/pk_warehouse/ts | | | 44 kB |
pg_default | postgres | | | | 46 GB |
pg_global | postgres | | | | 573 kB |
(12 rows)
数据占用的空间。
WAL日志占用的空间。
select application_name,client_addr,client_hostname,client_port,state,sync_priority,sync_state,pg_size_pretty(pg_wal_lsn_diff(pg_current_wal_lsn(), sent_lsn)) from pg_stat_replication;
select application_name,client_addr,client_hostname,client_port,state,sync_priority,sync_state,pg_size_pretty(pg_wal_lsn_diff(pg_current_wal_lsn(), replay_lag)) from pg_stat_replication;
select slot_name, plugin, slot_type, temporary, active, active_pid, pg_size_pretty(pg_wal_lsn_diff(pg_current_wal_lsn(), restart_lsn)) from pg_replication_slots;
最后一次归档失败时间减去最后一次归档成功的时间,求时间差。
select last_failed_time - last_archived_time from pg_stat_archiver;
以下都可以针对单个数据库输出,也可以输出整个实例的统计。
postgres=# \d pg_stat_database
View "pg_catalog.pg_stat_database"
Column | Type | Collation | Nullable | Default
----------------+--------------------------+-----------+----------+---------
datid | oid | | |
datname | name | | |
numbackends | integer | | |
xact_commit | bigint | | |
xact_rollback | bigint | | |
blks_read | bigint | | |
blks_hit | bigint | | |
tup_returned | bigint | | |
tup_fetched | bigint | | |
tup_inserted | bigint | | |
tup_updated | bigint | | |
tup_deleted | bigint | | |
conflicts | bigint | | |
temp_files | bigint | | |
temp_bytes | bigint | | |
deadlocks | bigint | | |
blk_read_time | double precision | | |
blk_write_time | double precision | | |
stats_reset | timestamp with time zone | | |
多次查询计算
select sum(xact_commit) from pg_stat_database; -- pg_stat_get_db_xact_commit 为stable函数,一个事务中两次调用之间只执行一次,所以需要外部多次执行。
select sum(xact_rollback) from pg_stat_database;
select sum(tup_returned) from pg_stat_database;
select sum(tup_fetched) from pg_stat_database;
select sum(tup_inserted) from pg_stat_database;
select sum(tup_updated) from pg_stat_database;
select sum(tup_deleted) from pg_stat_database;
select sum(conflicts) from pg_stat_database;
select sum(deadlocks) from pg_stat_database;
《PostgreSQL 锁等待监控 珍藏级SQL - 谁堵塞了谁》
《PostgreSQL snapshot too old补丁, 防止数据库膨胀》
《PostgreSQL垃圾回收代码分析 - why postgresql cann't reclaim tuple is HEAPTUPLE_RECENTLY_DEAD》
《PostgreSQL relcache在长连接应用中的内存霸占"坑"》
《PostgreSQL pg_stat_ pg_statio_ 统计信息(scan,read,fetch,hit)源码解读》