Django中prefetch_related()函数优化实战指南

 

前言

对于多对多字段(ManyToManyField)和一对多字段, 可以使用prefetch_related()来进行优化

prefetch_related()和select_related()的设计目的很相似,都是为了减少SQL查询的数量,但是实现的方式不一样。后者是通过JOIN语句,在SQL查询内解决问题。但是对于多对多关系,使用SQL语句解决就显得有些不太明智,因为JOIN得到的表将会很长,会导致SQL语句运行时间的增加和内存占用的增加。若有n个对象,每个对象的多对多字段对应Mi条,就会生成Σ(n)Mi 行的结果表。prefetch_related()的解决方法是,分别查询每个表,然后用Python处理他们之间的关系。继续以上边的例子进行说明,如果我们要获得张三所有去过的城市,使用prefetch_related()应该是这么做:

zhangs = Person.objects.prefetch_related('visitation').get(firstname=u"张",lastname=u"三")
>>> for city in zhangs.visitation.all() :
...   print city

上述代码触发的SQL查询如下:

SELECT `QSOptimize_person`.`id`, `QSOptimize_person`.`firstname`,
`QSOptimize_person`.`lastname`, `QSOptimize_person`.`hometown_id`, `QSOptimize_person`.`living_id`
FROM `QSOptimize_person`
WHERE (`QSOptimize_person`.`lastname` = '三'  AND `QSOptimize_person`.`firstname` = '张');

SELECT (`QSOptimize_person_visitation`.`person_id`) AS `_prefetch_related_val`, `QSOptimize_city`.`id`,
`QSOptimize_city`.`name`, `QSOptimize_city`.`province_id`
FROM `QSOptimize_city`
INNER JOIN `QSOptimize_person_visitation` ON (`QSOptimize_city`.`id` = `QSOptimize_person_visitation`.`city_id`)
WHERE `QSOptimize_person_visitation`.`person_id` IN (1);

第一条SQL查询仅仅是获取张三的Person对象,第二条比较关键,它选取关系表`QSOptimize_person_visitation`中`person_id`为张三的行,然后和`city`表内联(INNER JOIN 也叫等值连接)得到结果表。

+----+-----------+----------+-------------+-----------+
| id | firstname | lastname | hometown_id | living_id |
+----+-----------+----------+-------------+-----------+
| 1 | 张 | 三 | 3 | 1 |
+----+-----------+----------+-------------+-----------+
1 row in set (0.00 sec)

+-----------------------+----+-----------+-------------+
| _prefetch_related_val | id | name | province_id |
+-----------------------+----+-----------+-------------+
| 1 | 1 | 武汉市 | 1 |
| 1 | 2 | 广州市 | 2 |
| 1 | 3 | 十堰市 | 1 |
+-----------------------+----+-----------+-------------+
3 rows in set (0.00 sec)

显然张三武汉、广州、十堰都去过。

又或者,我们要获得湖北的所有城市名,可以这样:

>>> hb = Province.objects.prefetch_related('city_set').get(name__iexact=u"湖北省")
>>> for city in hb.city_set.all():
... city.name
...

触发的SQL查询:

SELECT `QSOptimize_province`.`id`, `QSOptimize_province`.`name`
FROM `QSOptimize_province`
WHERE `QSOptimize_province`.`name` LIKE '湖北省' ;

SELECT `QSOptimize_city`.`id`, `QSOptimize_city`.`name`, `QSOptimize_city`.`province_id`
FROM `QSOptimize_city`
WHERE `QSOptimize_city`.`province_id` IN (1);

得到的表:

+----+-----------+
| id | name      |
+----+-----------+
|  1 | 湖北省    |
+----+-----------+
1 row in set (0.00 sec)

+----+-----------+-------------+
| id | name      | province_id |
+----+-----------+-------------+
|  1 | 武汉市    |           1 |
|  3 | 十堰市    |           1 |
+----+-----------+-------------+
2 rows in set (0.00 sec)

我们可以看见,prefetch使用的是 IN 语句实现的。这样,在QuerySet中的对象数量过多的时候,根据数据库特性的不同有可能造成性能问题。

 

使用方法

*lookups 参数

prefetch_related()在Django < 1.7 只有这一种用法。和select_related()一样,prefetch_related()也支持深度查询,例如要获得所有姓张的人去过的省:

>>> zhangs = Person.objects.prefetch_related('visitation__province').filter(firstname__iexact=u'张')
>>> for i in zhangs:
... for city in i.visitation.all():
... print city.province
...

触发的SQL:

SELECT `QSOptimize_person`.`id`, `QSOptimize_person`.`firstname`,
`QSOptimize_person`.`lastname`, `QSOptimize_person`.`hometown_id`, `QSOptimize_person`.`living_id`
FROM `QSOptimize_person`
WHERE `QSOptimize_person`.`firstname` LIKE '张' ;

SELECT (`QSOptimize_person_visitation`.`person_id`) AS `_prefetch_related_val`, `QSOptimize_city`.`id`,
`QSOptimize_city`.`name`, `QSOptimize_city`.`province_id` FROM `QSOptimize_city`
INNER JOIN `QSOptimize_person_visitation` ON (`QSOptimize_city`.`id` = `QSOptimize_person_visitation`.`city_id`)
WHERE `QSOptimize_person_visitation`.`person_id` IN (1, 4);

SELECT `QSOptimize_province`.`id`, `QSOptimize_province`.`name`
FROM `QSOptimize_province`
WHERE `QSOptimize_province`.`id` IN (1, 2);

获得的结果:

+----+-----------+----------+-------------+-----------+
| id | firstname | lastname | hometown_id | living_id |
+----+-----------+----------+-------------+-----------+
|  1 | 张        | 三       |           3 |         1 |
|  4 | 张        | 六       |           2 |         2 |
+----+-----------+----------+-------------+-----------+
2 rows in set (0.00 sec)

+-----------------------+----+-----------+-------------+
| _prefetch_related_val | id | name      | province_id |
+-----------------------+----+-----------+-------------+
|                     1 |  1 | 武汉市    |           1 |
|                     1 |  2 | 广州市    |           2 |
|                     4 |  2 | 广州市    |           2 |
|                     1 |  3 | 十堰市    |           1 |
+-----------------------+----+-----------+-------------+
4 rows in set (0.00 sec)

+----+-----------+
| id | name      |
+----+-----------+
|  1 | 湖北省    |
|  2 | 广东省    |
+----+-----------+
2 rows in set (0.00 sec)

值得一提的是,链式prefetch_related会将这些查询添加起来,就像1.7中的select_related那样。

要注意的是,在使用QuerySet的时候,一旦在链式操作中改变了数据库请求,之前用prefetch_related缓存的数据将会被忽略掉。这会导致Django重新请求数据库来获得相应的数据,从而造成性能问题。这里提到的改变数据库请求指各种filter()、exclude()等等最终会改变SQL代码的操作。而all()并不会改变最终的数据库请求,因此是不会导致重新请求数据库的。

举个例子,要获取所有人访问过的城市中带有“市”字的城市,这样做会导致大量的SQL查询:

plist =Person.objects.prefetch_related('visitation')
[p.visitation.filter(name__icontains=u"市")for p in plist]

因为数据库中有4人,导致了2+4次SQL查询:

SELECT `QSOptimize_person`.`id`, `QSOptimize_person`.`firstname`, `QSOptimize_person`.`lastname`,
`QSOptimize_person`.`hometown_id`, `QSOptimize_person`.`living_id`
FROM `QSOptimize_person`;

SELECT (`QSOptimize_person_visitation`.`person_id`) AS `_prefetch_related_val`, `QSOptimize_city`.`id`,
`QSOptimize_city`.`name`, `QSOptimize_city`.`province_id`
FROM `QSOptimize_city`
INNER JOIN `QSOptimize_person_visitation` ON (`QSOptimize_city`.`id` = `QSOptimize_person_visitation`.`city_id`)
WHERE `QSOptimize_person_visitation`.`person_id` IN (1, 2, 3, 4);

SELECT `QSOptimize_city`.`id`, `QSOptimize_city`.`name`, `QSOptimize_city`.`province_id`
FROM `QSOptimize_city`
INNER JOIN `QSOptimize_person_visitation` ON (`QSOptimize_city`.`id` = `QSOptimize_person_visitation`.`city_id`)
WHERE(`QSOptimize_person_visitation`.`person_id` = 1 AND `QSOptimize_city`.`name` LIKE '%市%' );

SELECT `QSOptimize_city`.`id`, `QSOptimize_city`.`name`, `QSOptimize_city`.`province_id`
FROM `QSOptimize_city`
INNER JOIN `QSOptimize_person_visitation` ON (`QSOptimize_city`.`id` = `QSOptimize_person_visitation`.`city_id`)
WHERE (`QSOptimize_person_visitation`.`person_id` = 2 AND `QSOptimize_city`.`name` LIKE '%市%' );

SELECT `QSOptimize_city`.`id`, `QSOptimize_city`.`name`, `QSOptimize_city`.`province_id`
FROM `QSOptimize_city`
INNER JOIN `QSOptimize_person_visitation` ON (`QSOptimize_city`.`id` = `QSOptimize_person_visitation`.`city_id`)
WHERE (`QSOptimize_person_visitation`.`person_id` = 3 AND `QSOptimize_city`.`name` LIKE '%市%' );

SELECT `QSOptimize_city`.`id`, `QSOptimize_city`.`name`, `QSOptimize_city`.`province_id`
FROM `QSOptimize_city`
INNER JOIN `QSOptimize_person_visitation` ON (`QSOptimize_city`.`id` = `QSOptimize_person_visitation`.`city_id`)
WHERE (`QSOptimize_person_visitation`.`person_id` = 4 AND `QSOptimize_city`.`name` LIKE '%市%' );

众所周知,QuerySet是lazy的,要用的时候才会去访问数据库。运行到第二行Python代码时,for循环将plist看做iterator,这会触发数据库查询。最初的两次SQL查询就是prefetch_related导致的。

虽然已经查询结果中包含所有所需的city的信息,但因为在循环体中对Person.visitation进行了filter操作,这显然改变了数据库请求。因此这些操作会忽略掉之前缓存到的数据,重新进行SQL查询。

但是如果有这样的需求了应该怎么办呢?在Django >= 1.7,可以通过下一节的Prefetch对象来实现,如果你的环境是Django < 1.7,可以在Python中完成这部分操作。

plist = Person.objects.prefetch_related('visitation')
[[city for city in p.visitation.all() if u"市" in city.name] for p in plist]

Prefetch对象

在Django >= 1.7,可以用Prefetch对象来控制prefetch_related函数的行为。

1.一个Prefetch对象只能指定一项prefetch操作。

2.Prefetch对象对字段指定的方式和prefetch_related中的参数相同,都是通过双下划线连接的字段名完成的。

3.可以通过 queryset 参数手动指定prefetch使用的QuerySet。

4.可以通过 to_attr 参数指定prefetch到的属性名。

5.Prefetch对象和字符串形式指定的lookups参数可以混用。

 

最佳实践

1.prefetch_related主要针一对多和多对多关系进行优化。

2.prefetch_related通过分别获取各个表的内容,然后用Python处理他们之间的关系来进行优化。

3.可以通过可变长参数指定需要select_related的字段名。指定方式和特征与select_related是相同的。

4.在Django >= 1.7可以通过Prefetch对象来实现复杂查询,但低版本的Django好像只能自己实现。

5.作为prefetch_related的参数,Prefetch对象和字符串可以混用。

6.prefetch_related的链式调用会将对应的prefetch添加进去,而非替换,似乎没有基于不同版本上区别。

7.可以通过传入None来清空之前的prefetch_related。

选择哪个函数

如果我们想要获得所有家乡是湖北的人,最无脑的做法是先获得湖北省,再获得湖北的所有城市,最后获得故乡是这个城市的人。就像这样:

>>> hb = Province.objects.get(name__iexact=u"湖北省")
>>> people = []
>>> for city in hb.city_set.all():
... people.extend(city.birth.all())
...

显然这不是一个明智的选择,因为这样做会导致1+(湖北省城市数)次SQL查询。反正是个反例,导致的查询和获得掉结果就不列出来了。prefetch_related() 或许是一个好的解决方法,让我们来看看。

>>> hb = Province.objects.prefetch_related("city_set__birth").objects.get(name__iexact=u"湖北省")
>>> people = []
>>> for city in hb.city_set.all():
... people.extend(city.birth.all())
...

因为是一个深度为2的prefetch,所以会导致3次SQL查询:

SELECT `QSOptimize_province`.`id`, `QSOptimize_province`.`name`
FROM `QSOptimize_province`
WHERE `QSOptimize_province`.`name` LIKE '湖北省' ;

SELECT `QSOptimize_city`.`id`, `QSOptimize_city`.`name`, `QSOptimize_city`.`province_id`
FROM `QSOptimize_city`
WHERE `QSOptimize_city`.`province_id` IN (1);

SELECT `QSOptimize_person`.`id`, `QSOptimize_person`.`firstname`, `QSOptimize_person`.`lastname`,
`QSOptimize_person`.`hometown_id`, `QSOptimize_person`.`living_id`
FROM `QSOptimize_person`
WHERE `QSOptimize_person`.`hometown_id` IN (1, 3);

嗯…看上去不错,但是3次查询么?倒过来查询可能会更简单?

>>> people = list(Person.objects.select_related("hometown__province").filter(hometown__province__name__iexact=u"湖北省"))
SELECT `QSOptimize_person`.`id`, `QSOptimize_person`.`firstname`, `QSOptimize_person`.`lastname`,
`QSOptimize_person`.`hometown_id`, `QSOptimize_person`.`living_id`, `QSOptimize_city`.`id`,
`QSOptimize_city`.`name`, `QSOptimize_city`.`province_id`, `QSOptimize_province`.`id`, `QSOptimize_province`.`name`
FROM `QSOptimize_person`
INNER JOIN `QSOptimize_city` ON (`QSOptimize_person`.`hometown_id` = `QSOptimize_city`.`id`)
INNER JOIN `QSOptimize_province` ON (`QSOptimize_city`.`province_id` = `QSOptimize_province`.`id`)
WHERE `QSOptimize_province`.`name` LIKE '湖北省';
+----+-----------+----------+-------------+-----------+----+--------+-------------+----+--------+
| id | firstname | lastname | hometown_id | living_id | id | name | province_id | id | name |
+----+-----------+----------+-------------+-----------+----+--------+-------------+----+--------+
| 1 | 张 | 三 | 3 | 1 | 3 | 十堰市 | 1 | 1 | 湖北省 |
| 2 | 李 | 四 | 1 | 3 | 1 | 武汉市 | 1 | 1 | 湖北省 |
| 3 | 王 | 麻子 | 3 | 2 | 3 | 十堰市 | 1 | 1 | 湖北省 |
+----+-----------+----------+-------------+-----------+----+--------+-------------+----+--------+
3 rows in set (0.00 sec)

完全没问题。不仅SQL查询的数量减少了,python程序上也精简了。select_related()的效率要高于prefetch_related()。因此,最好在能用select_related()的地方尽量使用它,也就是说,对于ForeignKey字段,避免使用prefetch_related()。对于同一个QuerySet,你可以同时使用这两个函数。在我们一直使用的例子上加一个model:Order (订单)

class Order(models.Model):
customer = models.ForeignKey(Person)
orderinfo = models.CharField(max_length=50)
time = models.DateTimeField(auto_now_add = True)
def __unicode__(self):
return self.orderinfo

如果我们拿到了一个订单的id 我们要知道这个订单的客户去过的省份。因为有ManyToManyField显然必须要用prefetch_related()。如果只用prefetch_related()会怎样呢?

>>> plist = Order.objects.prefetch_related('customer__visitation__province').get(id=1)
>>> for city in plist.customer.visitation.all():
... print city.province.name
...

显然,关系到了4个表:Order、Person、City、Province,根据prefetch_related()的特性就得有4次SQL查询

SELECT `QSOptimize_order`.`id`, `QSOptimize_order`.`customer_id`, `QSOptimize_order`.`orderinfo`, `QSOptimize_order`.`time`
FROM `QSOptimize_order`
WHERE `QSOptimize_order`.`id` = 1 ;

SELECT `QSOptimize_person`.`id`, `QSOptimize_person`.`firstname`, `QSOptimize_person`.`lastname`, `QSOptimize_person`.`hometown_id`, `QSOptimize_person`.`living_id`
FROM `QSOptimize_person`
WHERE `QSOptimize_person`.`id` IN (1);

SELECT (`QSOptimize_person_visitation`.`person_id`) AS `_prefetch_related_val`, `QSOptimize_city`.`id`,
`QSOptimize_city`.`name`, `QSOptimize_city`.`province_id`
FROM `QSOptimize_city`
INNER JOIN `QSOptimize_person_visitation` ON (`QSOptimize_city`.`id` = `QSOptimize_person_visitation`.`city_id`)
WHERE `QSOptimize_person_visitation`.`person_id` IN (1);

SELECT `QSOptimize_province`.`id`, `QSOptimize_province`.`name`
FROM `QSOptimize_province`
WHERE `QSOptimize_province`.`id` IN (1, 2);
+----+-------------+---------------+---------------------+
| id | customer_id | orderinfo     | time                |
+----+-------------+---------------+---------------------+
|  1 |           1 | Info of Order | 2014-08-10 17:05:48 |
+----+-------------+---------------+---------------------+
1 row in set (0.00 sec)

+----+-----------+----------+-------------+-----------+
| id | firstname | lastname | hometown_id | living_id |
+----+-----------+----------+-------------+-----------+
|  1 | 张        | 三       |           3 |         1 |
+----+-----------+----------+-------------+-----------+
1 row in set (0.00 sec)

+-----------------------+----+--------+-------------+
| _prefetch_related_val | id | name   | province_id |
+-----------------------+----+--------+-------------+
|                     1 |  1 | 武汉市 |           1 |
|                     1 |  2 | 广州市 |           2 |
|                     1 |  3 | 十堰市 |           1 |
+-----------------------+----+--------+-------------+
3 rows in set (0.00 sec)

+----+--------+
| id | name   |
+----+--------+
|  1 | 湖北省 |
|  2 | 广东省 |
+----+--------+
2 rows in set (0.00 sec)

更好的办法是先调用一次select_related()再调用prefetch_related(),最后再select_related()后面的表

>>> plist = Order.objects.select_related('customer').prefetch_related('customer__visitation__province').get(id=1)
>>> for city in plist.customer.visitation.all():
... print city.province.name
...

这样只会有3次SQL查询,Django会先做select_related,之后prefetch_related的时候会利用之前缓存的数据,从而避免了1次额外的SQL查询:

SELECT `QSOptimize_order`.`id`, `QSOptimize_order`.`customer_id`, `QSOptimize_order`.`orderinfo`,
`QSOptimize_order`.`time`, `QSOptimize_person`.`id`, `QSOptimize_person`.`firstname`,
`QSOptimize_person`.`lastname`, `QSOptimize_person`.`hometown_id`, `QSOptimize_person`.`living_id`
FROM `QSOptimize_order`
INNER JOIN `QSOptimize_person` ON (`QSOptimize_order`.`customer_id` = `QSOptimize_person`.`id`)
WHERE `QSOptimize_order`.`id` = 1 ;

SELECT (`QSOptimize_person_visitation`.`person_id`) AS `_prefetch_related_val`, `QSOptimize_city`.`id`,
`QSOptimize_city`.`name`, `QSOptimize_city`.`province_id`
FROM `QSOptimize_city`
INNER JOIN `QSOptimize_person_visitation` ON (`QSOptimize_city`.`id` = `QSOptimize_person_visitation`.`city_id`)
WHERE `QSOptimize_person_visitation`.`person_id` IN (1);

SELECT `QSOptimize_province`.`id`, `QSOptimize_province`.`name`
FROM `QSOptimize_province`
WHERE `QSOptimize_province`.`id` IN (1, 2);
+----+-------------+---------------+---------------------+----+-----------+----------+-------------+-----------+
| id | customer_id | orderinfo     | time                | id | firstname | lastname | hometown_id | living_id |
+----+-------------+---------------+---------------------+----+-----------+----------+-------------+-----------+
|  1 |           1 | Info of Order | 2014-08-10 17:05:48 |  1 | 张        | 三       |           3 |         1 |
+----+-------------+---------------+---------------------+----+-----------+----------+-------------+-----------+
1 row in set (0.00 sec)

+-----------------------+----+--------+-------------+
| _prefetch_related_val | id | name   | province_id |
+-----------------------+----+--------+-------------+
|                     1 |  1 | 武汉市 |           1 |
|                     1 |  2 | 广州市 |           2 |
|                     1 |  3 | 十堰市 |           1 |
+-----------------------+----+--------+-------------+
3 rows in set (0.00 sec)

+----+--------+
| id | name   |
+----+--------+
|  1 | 湖北省 |
|  2 | 广东省 |
+----+--------+
2 rows in set (0.00 sec)

值得注意的是,可以在调用prefetch_related之前调用select_related,并且Django会按照你想的去做:先select_related,然后利用缓存到的数据prefetch_related。然而一旦prefetch_related已经调用,select_related将不起作用。

 

小结

  • 因为select_related()总是在单次SQL查询中解决问题,而prefetch_related()会对每个相关表进行SQL查询,因此select_related()的效率通常比后者高。
  • 鉴于第一条,尽可能的用select_related()解决问题。只有在select_related()不能解决问题的时候再去想prefetch_related()。
  • 你可以在一个QuerySet中同时使用select_related()和prefetch_related(),从而减少SQL查询的次数。
  • 只有prefetch_related()之前的select_related()是有效的,之后的将会被无视掉。

 

总结

关于Django中prefetch_related()函数优化实战指南的文章就介绍至此,更多相关Djangoprefetch_related()函数优化内容请搜索编程宝库以前的文章,希望以后支持编程宝库

 tf.app.flags命令行参数解析模块说道命令行参数解析,就不得不提到 python 的 argparse 模块,详情可参考我之前的一篇文章:python ...