查看连接数变化
CPU利用率到达100%,首先怀疑,是不是业务高峰活跃连接陡增,而数据库预留的资源不足造成的结果。我们需要查看下,问题发生时,活跃的连接数是否比平时多很多。对于RDS for PG,数据库上的连接数变化,可以从控制台的监控信息中看到。而当前活跃的连接数>可以直接连接数据库,使用下列查询语句得到:
select count( * ) from pg_stat_activity where state not like '%idle';
追踪慢SQL
如果活跃连接数的变化处于正常范围,则很大概率可能是当时有性能很差的SQL被大量执行导致。由于RDS有慢SQL日志,我们可以通过这个日志,定位到当时比较耗时的SQL来进一步做分析。但通常问题发生时,整个系统都处于停滞状态,所有SQL都慢下来,当时记录的>慢SQL可能非常多,并不容易排查罪魁祸首。这里我们介绍几种在问题发生时,即介入追查慢SQL的方法。
1、第一种方法是使用pg_stat_statements插件定位慢SQL,步骤如下。
1.1 如果没有创建这个插件,需要手动创建。我们要利用插件和数据库系统里面的计数信息(如SQL执行时间累积等),而这些信息是不断累积的,包含了历史信息。为了更方便的排查当前的CPU满问题,我们要先重置计数器。
create extension pg_stat_statements; select pg_stat_reset(); select pg_stat_statements_reset();
1.2 等待一段时间(例如1分钟),使计数器积累足够的信息。
1.3 查询最耗时的SQL(一般就是导致问题的直接原因)。
select * from pg_stat_statements order by total_time desc limit 5;
1.4 查询读取Buffer次数最多的SQL,这些SQL可能由于所查询的数据没有索引,而导致了过多的Buffer读,也同时大量消耗了CPU。
select * from pg_stat_statements order by shared_blks_hit+shared_blks_read desc limit 5;
2、第二种方法是,直接通过pg_stat_activity视图,利用下面的查询,查看当前长时间执行,一直不结束的SQL。这些SQL对应造成CPU满,也有直接嫌疑。
select datname, usename, client_addr, application_name, state, backend_start, xact_start, xact_stay, query_start, query_stay, replace(query, chr(10), ' ') as query from (select pgsa.datname as datname, pgsa.usename as usename, pgsa.client_addr client_addr, pgsa.application_name as application_name, pgsa.state as state, pgsa.backend_start as backend_start, pgsa.xact_start as xact_start, extract(epoch from (now() - pgsa.xact_start)) as xact_stay, pgsa.query_start as query_start, extract(epoch from (now() - pgsa.query_start)) as query_stay , pgsa.query as query from pg_stat_activity as pgsa where pgsa.state != 'idle' and pgsa.state != 'idle in transaction' and pgsa.state != 'idle in transaction (aborted)') idleconnections order by query_stay desc limit 5;
datname | usename | client_addr | application_name | state | backend_start | xact_start | xact_stay | query_start | query_stay | query ---------+-------------+---------------+--------------------------+--------+-------------------------------+-------------------------------+---------------+-------------------------------+---------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- denali | denaliadmin | 10.222.16.45 | pgAdmin III - Query Tool | active | 2018-02-26 22:36:05.603781+00 | 2018-02-26 22:36:13.054396+00 | 187614.245395 | 2018-02-26 22:36:13.054396+00 | 187614.245395 | select * from gen3_search_eu_17q2_20171115_epl.place_name where place_id not in (select place_id from gen3_search_eu_17q1_20170308_epl.place_name ) and name not in (select name from gen3_search_eu_17q1_20170308_epl.place_name) and lang = 'ENG' limit 50 denali | denaliadmin | 10.222.16.45 | pgAdmin III - Query Tool | active | 2018-02-26 23:46:24.442846+00 | 2018-02-26 23:46:34.920261+00 | 183392.37953 | 2018-02-26 23:46:34.920261+00 | 183392.37953 | select * from gen3_search_eu_17q2_20171115_epl.place_name where place_id not in (select place_id from gen3_search_eu_17q1_20170308_epl.place_name ) and name not in (select name from gen3_search_eu_17q1_20170308_epl.place_name) and lang = 'ENG' limit 50 denali | denaliadmin | 10.222.16.45 | pgAdmin III - Query Tool | active | 2018-02-27 01:19:53.83589+00 | 2018-02-27 01:20:01.519778+00 | 177785.780013 | 2018-02-27 01:20:01.519778+00 | 177785.780013 | select * from gen3_search_eu_17q2_20171115_epl.place_name where place_id not in (select place_id from gen3_search_eu_17q1_20170308_epl.place_name ) and name not in (select name from gen3_search_eu_17q1_20170308_epl.place_name) limit 50 denali | denaliadmin | 10.222.16.45 | pgAdmin III - Query Tool | active | 2018-02-27 01:46:05.207888+00 | 2018-02-27 01:47:52.039779+00 | 176115.260012 | 2018-02-27 01:47:52.039779+00 | 176115.260012 | select a.place_id, a.metadata_dictionary_id,a.value, a.lang, b.place_id, b.metadata_dictionary_id, b.value, b.lang from gen3_search_eu_17q1_20170308_epl.place_address a inner join gen3_search_eu_17q2_20171115_epl.place_address b on a.place_id = b.place_id where a.metadata_dictionary_id = b.metadata_dictionary_id and a.lang = b.lang and a.value!=b.value and b.place_id not in (select poi_id from gen3_search_eu_17q2_20171115_epl.place_embeded_ids) limit 100 denali | denaliadmin | 10.224.14.148 | pgAdmin III - Query Tool | active | 2018-02-27 05:05:39.903885+00 | 2018-02-27 05:05:48.827779+00 | 164238.472012 | 2018-02-27 05:05:48.827779+00 | 164238.472012 | select a.place_id, a.metadata_dictionary_id,a.value, a.lang, b.place_id, b.metadata_dictionary_id, b.value, b.lang from gen3_search_eu_17q1_20170308_epl.place_address a inner join gen3_search_eu_17q2_20171115_epl.place_address b on a.place_id = b.place_id where a.metadata_dictionary_id = b.metadata_dictionary_id and a.lang = b.lang and a.value!=b.value and b.place_id not in (select poi_id from gen3_search_eu_17q2_20171115_epl.place_embeded_ids) limit 100 (5 rows)
3、第3种方法,是从数据表上表扫描(Table Scan)的信息开始查起,查找缺失索引的表。数据表如果缺失索引,大部分热数据又都在内存时(例如内存8G,热数据6G),此时数据库只能使用表扫描,并需要处理已在内存中的大量的无关记录,而耗费大量CPU。特别是对于表记录数超100的表,一次表扫描占用大量CPU(基本把一个CPU占满),多个连接并发(例如上百连接),把所有CPU占满。
3.1 通过下面的查询,查出使用表扫描最多的表:
select * from pg_stat_user_tables where n_live_tup > 100000 and seq_scan > 0 order by seq_tup_read desc limit 10;
3.2 查询当前正在运行的访问到上述表的慢查询:
select * from pg_stat_activity where query ilike '%<table name>%' and query_start - now() > interval '10 seconds';
3.3 也可以通过pg_stat_statements插件定位涉及到这些表的查询:
select * from pg_stat_statements where query ilike '%<table>%'order by shared_blks_hit+shared_blks_read desc limit 3;
处理慢SQL
对于上面的方法查出来的慢SQL,首先需要做的可能是Cancel或Kill掉他们,使业务先恢复:
select pg_cancel_backend(pid) from pg_stat_activity where query like '%<query text>%' and pid != pg_backend_pid(); select pg_terminate_backend(pid) from pg_stat_activity where query like '%<query text>%' and pid != pg_backend_pid();
如果这些SQL确实是业务上必需的,则需要对他们做优化。这方面有“三板斧”:
1、对查询涉及的表,执行ANALYZE <table>或VACUUM ANZLYZE <table>,更新表的统计信息,使查询计划更准确。注意,为避免对业务影响,最好在业务低峰执行。
2、执行explain (query text)或explain (buffers true, analyze true, verbose true) (query text)命令,查看SQL的执行计划(注意,前者不会实际执行SQL,后者会实际执行而且能得到详细的执行信息),对其中的Table Scan涉及的表,建立索引。
3、重新编写SQL,去除掉不必要的子查询、改写UNION ALL、使用JOIN CLAUSE固定连接顺序等到,都是进一步深度优化SQL的手段,这里不再深入说明。
https://www.centos.bz/2017/08/postgresql-cpu-100-slow-sql/