Django:尽管进行了优化，但随机查询速度很慢

Question

我构建了一个API，它应该返回10个从大型查询集中随机选择的结果。

我有以下4种型号：

class ScrapingOperation(models.Model):
completed = models.BooleanField(default=False)
(...)
indexes = [
models.Index(fields=['completed'], name='completed_idx'),
models.Index(fields=['trusted'], name='trusted_idx'),
]
@property
def ads(self):
"""returns all ads linked to the searches of this operation"""
return Ad.objects.filter(searches__in=self.searches.all())

class Search(models.Model):
completed = models.BooleanField(default=False)
scraping_operation = models.ForeignKey(
ScrapingOperation,
on_delete=models.CASCADE,
related_name='searches'
)
(...)

class Ad(models.Model):
searches = models.ManyToManyField('scraper.Search', related_name='ads')
(...)

class Label(models.Model):
value = models.Integerfield()
linked_ad = models.OneToOneField(
Ad, on_delete=models.CASCADE, related_name='labels'
)

数据库中当前有400000+个Ad对象，但平均ScrapingOperation有14000个Ad对象链接到它。我希望API从这些+/-14000中返回10个随机结果，这些结果还没有链接的Label对象(每个操作最多只存在几百个(

因此，必须从包含14.000个对象的查询中返回10个随机结果。

早期版本只需要返回1个结果，但使用了速度慢得多的sort_by('?')方法。当我不得不将其放大以返回随机的10个Ad对象时，我使用了一种新的方法，部分基于这种堆叠式流回答

以下是选择(并返回(10个随机对象的代码：

# Get all ads linked to the last completed operation
last_op_ads = ScrapingOperation.objects.filter(completed=True).last().ads
# Get all ads that don't have an label yet
random_ads = last_op_ads.filter(labels__isnull=True)
# Get list ids of all potential ads
id_list = random_ads.values_list('id', flat=True)
id_list = list(id_list)
# Select a random sample of 10, get objects with PK matches
samples = rd.sample(id_list, min(len(id_list), 10))
selected_samples = random_ads.filter(id__in=samples)
return selected_samples

然而，尽管我进行了优化，但这个查询需要10多秒才能完成，从而创建了一个非常缓慢的API。

这种长延迟只是随机查询固有的吗？(如果是这样的话，其他程序员如何处理这种限制？(或者我的代码中是否存在我遗漏的错误/效率低下？

编辑：根据响应，我在下面包含了原始sql查询(注意：这些查询在我的本地环境中运行，它只包含我的生产环境中包含的5%的数据(

{'sql': 'SELECT "scraper_scrapingoperation"."id", 
"scraper_scrapingoperation"."date_started", 
"scraper_scrapingoperation"."date_completed",
"scraper_scrapingoperation"."completed", 
"scraper_scrapingoperation"."round", 
"scraper_scrapingoperation"."trusted" FROM "scraper_scrapingoperation" 
WHERE "scraper_scrapingoperation"."completed" = true ORDER BY 
"scraper_scrapingoperation"."id" DESC LIMIT 1', 'time': '0.001'}

{'sql': 'SELECT "database_ad"."id" FROM "database_ad" INNER JOIN 
"database_ad_searches" ON ("database_ad"."id" = 
"database_ad_searches"."ad_id") LEFT OUTER JOIN "classifier_label" ON
("database_ad"."id" = "classifier_label"."ad_id") WHERE 
("database_ad_searches"."search_id" IN (SELECT U0."id" FROM 
"scraper_search" U0 WHERE U0."scraping_operation_id" = 6) AND 
"classifier_label"."id" IS NULL)', 'time': '1.677'}

编辑2：我尝试了另一种方法，使用更深的select_related参数

random_ads = ScrapingOperation.objects.prefetch_related(
'searches__ads__labels',
).filter(completed=True).last().ads.exclude(
labels__isnull=True
)
id_list = random_ads.values_list('id', flat=True)
id_list = list(id_list)
samples = rd.sample(id_list, min(
len(id_list), 10))
selected_samples = random_ads.filter(
id__in=samples)
return selected_samples

它生成以下SQL查询：

{'time': '0.008', 'sql': 'SELECT "scraper_search"."id", 
"scraper_search"."item_id", "scraper_search"."date_started", 
"scraper_search"."date_completed", "scraper_search"."completed", 
"scraper_search"."round", "scraper_search"."scraping_operation_id", 
"scraper_search"."trusted" FROM "scraper_search" WHERE 
"scraper_search"."scraping_operation_id" IN (6)'}

{'time': '0.113', 'sql': 'SELECT ("database_ad_searches"."search_id")
AS "_prefetch_related_val_search_id", "database_ad"."id", 
"database_ad"."item_id", "database_ad"."item_state", 
"database_ad"."title", "database_ad"."seller_id", 
"database_ad"."url", "database_ad"."price", 
"database_ad"."transaction_type", "database_ad"."transaction_method",
"database_ad"."first_seen", "database_ad"."last_seen", 
"database_ad"."promoted" FROM "database_ad" INNER JOIN 
"database_ad_searches" ON ("database_ad"."id" = 
"database_ad_searches"."ad_id") WHERE 
"database_ad_searches"."search_id" IN (130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160)'}

{'time': '0.041', 'sql': 'SELECT "classifier_label"."id", 
"classifier_label"."set_by_id", "classifier_label"."ad_id", 
"classifier_label"."date", "classifier_label"."phone_type", 
"classifier_label"."seller_type", "classifier_label"."sale_type" FROM
"classifier_label" WHERE "classifier_label"."ad_id" IN (1, 3, 6, 10, 20, 29, 30, 35, 43, (and MANY more of these numbers) ....'}

{'time': '1.498', 'sql': 'SELECT "database_ad"."id" FROM "database_ad"
INNER JOIN "database_ad_searches" ON ("database_ad"."id" = "database_ad_searches"."ad_id") LEFT OUTER JOIN "classifier_label" ON 
("database_ad"."id" = "classifier_label"."ad_id") WHERE 
("database_ad_searches"."search_id" IN (SELECT U0."id" FROM
"scraper_search" U0 WHERE U0."scraping_operation_id" = 6) AND NOT 
("classifier_label"."id" IS NOT NULL))'}

每个ScrapingOperation"仅"有+/-14000个链接广告，但生产中的广告总数为400000个(而且还在增长(。上面的所有代码在我的本地环境(只包含5%的数据(上返回有效结果，但在生产中的API上返回502个错误。

Answer 1

我会尝试首先隔离链接的广告，然后通过生成的随机列使用顺序从中随机获得10个。我不确定这在生成的sql中如何有效。可以肯定的是，我更喜欢在任务上创建一个存储过程，因为这显然是一个最终在随机样本上进行的数据挖掘操作。