mysql - Percona 5.7 在许多连接上速度慢

Question

我们最近将 percona 5.5 sql server 升级到了 percona 5.7。到目前为止效果很好。不幸的是，我们有一个巨大的查询，在 5.7 下速度极慢。低于 5.5。即使使用 sql_no_cache，也只需不到一秒钟的时间。使用 Percona 5.7。执行此查询最多需要 1 分钟。奇怪的是，随着我们使用的组合指标越多，它变得越慢。删除所有组合索引导致执行时间为 30 秒。强制 sql_straight_join 使查询在不到一秒内运行。

所以这里是查询:

SELECT t0_.tree_id AS tree_id0, t1_.treetype_name AS treetype_name1, c2_.contentelement_id AS contentelement_id2, t0_.tree_name AS tree_name3, (CASE WHEN t3_.treetype_name <> 'global' THEN t4_.tree_name ELSE t0_.tree_name END) AS sclr4, p5_.picture_id AS picture_id5, t6_.tree_misc_value_text AS tree_misc_value_text6, (CASE WHEN t3_.treetype_name <> 'global' THEN t7_.tree_misc_value_text ELSE t6_.tree_misc_value_text END) AS sclr7, w8_.widgetgeneral_slug AS widgetgeneral_slug8, (CASE WHEN t3_.treetype_name <> 'global' THEN w9_.widgetgeneral_slug ELSE w8_.widgetgeneral_slug END) AS sclr9, t10_.tree_misc_value_text AS tree_misc_value_text10, t11_.tree_misc_value_text AS tree_misc_value_text11
FROM tree_relation t12_
INNER JOIN tree t4_ ON t12_.tree_relation_parent = t4_.tree_id
INNER JOIN treetype t3_ ON t4_.tree_type_id = t3_.treetype_id AND (t3_.treetype_name IN ('global', 'country'))
INNER JOIN contentelement c13_ ON t4_.tree_id = c13_.contentelement_tree_id
INNER JOIN contentleaf c14_ ON c13_.contentelement_contentleaf_id = c14_.contentleaf_id AND (c14_.contentleaf_contentbranch_id = 1)
INNER JOIN widgetgeneral w9_
INNER JOIN widgetabstract w15_ ON w9_.widgetabstract_id = w15_.widgetabstract_id AND (w15_.widgetabstract_contentelement_id = c13_.contentelement_id AND w15_.widgetabstract_discriminator IN ('general') AND w15_.widgetabstract_state = 'preview')
INNER JOIN tree t0_ ON t12_.tree_relation_child = t0_.tree_id
INNER JOIN treetype t1_ ON t0_.tree_type_id = t1_.treetype_id AND (t1_.treetype_name IN ('city','region'))
INNER JOIN contentelement c2_ ON t0_.tree_id = c2_.contentelement_tree_id
INNER JOIN contentleaf c16_ ON c2_.contentelement_contentleaf_id = c16_.contentleaf_id AND (c16_.contentleaf_contentbranch_id = 1)
INNER JOIN widgetgeneral w8_
INNER JOIN widgetabstract w17_ ON w8_.widgetabstract_id = w17_.widgetabstract_id AND (w17_.widgetabstract_contentelement_id = c2_.contentelement_id AND w17_.widgetabstract_discriminator IN ('general') AND w17_.widgetabstract_state = 'preview')
INNER JOIN widgetgeneral w18_
INNER JOIN widgetabstract w19_ ON w18_.widgetabstract_id = w19_.widgetabstract_id AND (w19_.widgetabstract_contentleaf_id = c16_.contentleaf_id AND w19_.widgetabstract_discriminator IN ('general') AND w19_.widgetabstract_state = 'preview')
LEFT JOIN picture p5_ ON t0_.tree_picture_id = p5_.picture_id
LEFT JOIN tree_misc t6_ ON t0_.tree_id = t6_.tree_misc_tree_id AND (t6_.tree_misc_attributetype_key = 'flagId')
LEFT JOIN tree_misc t7_ ON t4_.tree_id = t7_.tree_misc_tree_id AND (t7_.tree_misc_attributetype_key = 'flagId')
LEFT JOIN tree_misc t10_ ON t0_.tree_id = t10_.tree_misc_tree_id AND (t10_.tree_misc_attributetype_key = 'latitude')
LEFT JOIN tree_misc t11_ ON t0_.tree_id = t11_.tree_misc_tree_id AND (t11_.tree_misc_attributetype_key = 'longitude')
WHERE w17_.widgetabstract_visibility = 'active' OR (w17_.widgetabstract_visibility = 'parent' AND w19_.widgetabstract_visibility = 'active')

和 5.7 的解释:

我们尝试了升级以及完全空白安装。打开和关闭所有 sql 模式和查询优化器选项。如果您需要更多信息或服务器变量，请告诉我。

操作系统:Debian GNU/Linux 8 (jessie)服务器版本为:5.7.14-7-log Percona Server (GPL)，版本“7”，修订版“083e298”

也许你能暗示我们遗漏了什么。

编辑:添加配置

[mysqld]
port                            = 3306
user                            = mysql
socket                          = /var/run/mysqld/mysqld.sock
pid-file                        = /var/run/mysqld/mysqld.pid
basedir                         = /usr
datadir                         = /var/lib/mysql
tmpdir                          = /tmp
lc-messages-dir                 = /usr/share/mysql
max_connect_errors              = 1000000
log-error                       = /var/log/mysql/error.log
skip-external-locking
myisam-recover-options          = BACKUP
character-set-server            = utf8
collation-server                = utf8_general_ci
interactive_timeout             = 28800
wait_timeout                    = 28800
skip-name-resolve
group_concat_max_len            = 268435456

innodb_file_per_table
innodb_buffer_pool_size               = 48G
innodb_buffer_pool_instances          = 1
innodb_flush_log_at_trx_commit        = 1
innodb_data_file_path                 = ibdata1:2G:autoextend
innodb_log_file_size                  = 256M
innodb_log_buffer_size                = 64M
innodb_file_format                    = barracuda
innodb_flush_method                   = O_DIRECT[mysqld_safe]
syslog
numa_interleave

# Per Thread
sort_buffer_size        = 4M
read_buffer_size        = 2M

# Cache/connection relevant
thread_cache_size       = 850
table_open_cache        = 4048
max_connections         = 1300

# MyISAM settings (also valid for queries with temporary tables)
key_buffer_size         = 128M
myisam_sort_buffer_size = 16M

# Misc
max_allowed_packet      = 256M
max_heap_table_size     = 16M
thread_stack            = 192K
tmp_table_size          = 16M

# Query cache
query_cache_limit       = 5M
query_cache_size        = 4024M

server-id              = 102
log_bin                = /var/log/mysql/mysql-bin.log
binlog_format          = mixed
expire_logs_days       = 10
max_binlog_size        = 100M
# enforce syncing of every transation to binlog (crash safe, with bbu this should be fast)
sync_binlog            = 1
sync_relay_log         = 1
sync_master_info       = 1
sync_relay_log_info    = 1
relay-log              = mysqld-relay-bin
skip-slave-start
log-slave-updates

slow_query_log                 = 1
slow_query_log_file            = /var/log/mysql/mysql-slow.log
long_query_time                = 1
log-queries-not-using-indexes

编辑 2:为 5.5 添加说明

Answer 1

新的连接排序可能是由于 MySQL 5.7 高估了可以基于 WHERE 和 ON 子句进行的过滤的效果。在 MySQL 5.6 中，没有考虑过滤，这通常会导致选择不必要的昂贵连接顺序。一般来说，MySQL 5.7 通常能够通过考虑过滤来找到更好的连接顺序。但是，对于未索引列的条件，过滤估计只是一个猜测，对于选择性不是很强的条件可能效果不佳。

您可以通过设置 optimizer_switch='condition_fanout_filter=off' 恢复到 5.6 行为，或者您可以使用 STRAIGHT_JOIN 强制执行特定的连接顺序。