Django 全文搜索优化 - Postgres

Question

我正在尝试利用 Django (v2.1) 和 Postgres (9.5) 创建地址自动完成功能的全文搜索，但性能不符合自动完成 table那一刻，我不明白我得到的性能结果背后的逻辑。有关信息，table 相当大，有 1400 万行。

我的模型：

from django.db import models
from postgres_copy import CopyManager
from django.contrib.postgres.indexes import GinIndex

class Addresses(models.Model):
date_update = models.DateTimeField(auto_now=True, null=True)
longitude = models.DecimalField(max_digits=9, decimal_places=6 , null=True)
latitude = models.DecimalField(max_digits=9, decimal_places=6 , null=True)
number = models.CharField(max_length=16, null=True, default='')
street = models.CharField(max_length=60, null=True, default='')
unit = models.CharField(max_length=50, null=True, default='')
city = models.CharField(max_length=50, null=True, default='')
district = models.CharField(max_length=10, null=True, default='')
region = models.CharField(max_length=5, null=True, default='')
postcode = models.CharField(max_length=5, null=True, default='')
addr_id = models.CharField(max_length=20, unique=True)
addr_hash = models.CharField(max_length=20, unique=True)
objects = CopyManager()

class Meta:
    indexes = [
        GinIndex(fields=['number', 'street', 'unit', 'city', 'region', 'postcode'], name='search_idx')
    ]

我创建了一个小测试来根据搜索中的单词数量检查性能：

    search_vector = SearchVector('number', 'street', 'unit', 'city', 'region', 'postcode')

    searchtext1 = "north"
    searchtext2 = "north bondi"
    searchtext3 = "north bondi blair"
    searchtext4 = "north bondi blair street 2026"

    print('Test1: 1 word')
    start_time = time.time()
    result = AddressesAustralia.objects.annotate(search=search_vector).filter(search=searchtext1)[:10]
    #print(len(result))
    time_exec = str(timedelta(seconds=time.time() - start_time))
    print(time_exec)
    print(' ')

    #print(AddressesAustralia.objects.annotate(search=search_vector).explain(verbose=True))

    print('Test2: 2 words')
    start_time = time.time()
    result = AddressesAustralia.objects.annotate(search=search_vector).filter(search=searchtext2)[:10]
    #print(len(result))
    time_exec = str(timedelta(seconds=time.time() - start_time))
    print(time_exec)
    print(' ')

    print('Test3: 3 words')
    start_time = time.time()
    result = AddressesAustralia.objects.annotate(search=search_vector).filter(search=searchtext3)[:10]
    #print(len(result))
    time_exec = str(timedelta(seconds=time.time() - start_time))
    print(time_exec)
    print(' ')

    print('Test4: 5 words')
    start_time = time.time()
    result = AddressesAustralia.objects.annotate(search=search_vector).filter(search=searchtext4)[:10]
    #print(len(result))
    time_exec = str(timedelta(seconds=time.time() - start_time))
    print(time_exec)
    print(' ')

我得到以下结果，这似乎很正确：

Test1: 1 word
0:00:00.001841

Test2: 2 words
0:00:00.001422

Test3: 3 words
0:00:00.001574

Test4: 5 words
0:00:00.001360

但是，如果我取消注释 print(len(results)) 行，我会得到以下结果：

Test1: 1 word
10
0:00:00.046392

Test2: 2 words
10
0:00:06.544732

Test3: 3 words
10
0:01:12.367157

Test4: 5 words
10
0:01:17.786596

这显然不适合table自动完成功能。

有人可以解释为什么在对查询集结果执行操作时执行时间更长吗？似乎数据库检索总是很快，但是通过结果需要时间，这对我来说没有意义，因为我将结果限制为 10，返回的查询集总是相同的大小。

另外，虽然我创建了一个GIN索引，但是这个索引好像没有用到。它似乎已正确创建：

=# \d public_data_au_addresses
                                   Table 
"public.public_data_au_addresses"
Column    |           Type           | Collation | Nullable |                            
Default                            
-------------+--------------------------+-----------+----------+------ 
---------------------------------------------------------
id          | integer                  |           | not null | 
nextval('public_data_au_addresses_id_seq'::regclass)
date_update | timestamp with time zone |           |          | 
longitude   | numeric(9,6)             |           |          | 
latitude    | numeric(9,6)             |           |          | 
number      | character varying(16)    |           |          | 
street      | character varying(60)    |           |          | 
unit        | character varying(50)    |           |          | 
city        | character varying(50)    |           |          | 
district    | character varying(10)    |           |          | 
region      | character varying(5)     |           |          | 
postcode    | character varying(5)     |           |          | 
addr_id     | character varying(20)    |           | not null | 
addr_hash   | character varying(20)    |           | not null | 
Indexes:
"public_data_au_addresses_pkey" PRIMARY KEY, btree (id)
"public_data_au_addresses_addr_hash_key" UNIQUE CONSTRAINT, btree (addr_hash)
"public_data_au_addresses_addr_id_key" UNIQUE CONSTRAINT, btree (addr_id)
"public_data_au_addresses_addr_hash_e8c67a89_like" btree (addr_hash varchar_pattern_ops)
"public_data_au_addresses_addr_id_9ee00c76_like" btree (addr_id varchar_pattern_ops)
"search_idx" gin (number, street, unit, city, region, postcode)

当我运行在我的查询中使用 explain() 方法时，我得到：

Test1: 1 word
Limit  (cost=0.00..1110.60 rows=10 width=140)
->  Seq Scan on public_data_au_addresses  (cost=0.00..8081472.41 rows=72767 width=140)
    Filter: (to_tsvector((((((((((((COALESCE(number, ''::character varying))::text || ' '::text) || (COALESCE(street, ''::character varying))::text) || ' '::text) || (COALESCE(unit, ''::character varying))::text) || ' '::text) || (COALESCE(city, ''::character varying))::text) || ' '::text) || (COALESCE(region, ''::character varying))::text) || ' '::text) || (COALESCE(postcode, ''::character varying))::text)) @@ plainto_tsquery('north'::text))

所以它仍然显示顺序扫描而不是使用索引扫描。有谁知道如何修复或调试它？

GIN 索引在有这么多字段要搜索的情况下仍然有效吗？

最后，有谁知道我可以如何改进代码以进一步提高性能？

谢谢！此致

更新

我尝试按照下面 Paolo 的建议创建一个搜索向量，但搜索似乎仍然是顺序的，没有利用 GIN 索引。

class AddressesQuerySet(CopyQuerySet):

    def update_search_vector(self):
        return self.update(search_vector=SearchVector('number', 'street', 'unit', 'city', 'region', 'postcode', config='english'))


class AddressesAustralia(models.Model):
    date_update = models.DateTimeField(auto_now=True, null=True)
    longitude = models.DecimalField(max_digits=9, decimal_places=6 , null=True)
    latitude = models.DecimalField(max_digits=9, decimal_places=6 , null=True)
    number = models.CharField(max_length=16, null=True, default='')
    street = models.CharField(max_length=60, null=True, default='')
    unit = models.CharField(max_length=50, null=True, default='')
    city = models.CharField(max_length=50, null=True, default='')
    district = models.CharField(max_length=10, null=True, default='')
    region = models.CharField(max_length=5, null=True, default='')
    postcode = models.CharField(max_length=5, null=True, default='')
    addr_id = models.CharField(max_length=20, unique=True)
    addr_hash = models.CharField(max_length=20, unique=True)
    search_vector = SearchVectorField(null=True, editable=False)

    objects = AddressesQuerySet.as_manager()

    class Meta:
        indexes = [
            GinIndex(fields=['search_vector'], name='search_vector_idx')
        ]

然后我使用更新命令更新了 search_vector 字段：

AddressesAustralia.objects.update_search_vector()

然后我运行使用相同的搜索向量进行测试的查询：

class Command(BaseCommand):

    def handle(self, *args, **options):

        search_vector = SearchVector('number', 'street', 'unit', 'city', 'region', 'postcode', config='english')

        searchtext1 = "north"

        print('Test1: 1 word')
        start_time = time.time()
        result = AddressesAustralia.objects.filter(search_vector=searchtext1)[:10].explain(verbose=True)
        print(len(result))
        print(result)
        time_exec = str(timedelta(seconds=time.time() - start_time))
        print(time_exec)

我得到以下结果，仍然显示顺序搜索：

Test1: 1 word
532
Limit  (cost=0.00..120.89 rows=10 width=235)
  Output: id, date_update, longitude, latitude, number, street, unit, city, district, region, postcode, addr_id, addr_hash, search_vector
  ->  Seq Scan on public.public_data_au_addressesaustralia  (cost=0.00..5061078.91 rows=418651 width=235)
        Output: id, date_update, longitude, latitude, number, street, unit, city, district, region, postcode, addr_id, addr_hash, search_vector
        Filter: (public_data_au_addressesaustralia.search_vector @@ plainto_tsquery('north'::text))
0:00:00.075262

我也试过：

在搜索向量中有和没有配置="english"（在更新和查询中）
删除GIN索引，然后重新创建然后重新运行 update_search_Vector

但还是一样的结果。知道我做错了什么或如何进一步排除故障吗？

Answer 1

您需要在搜索向量上创建一个 functional index。现在您在基础字段上有一个索引，但它仍然必须在过滤结果之前为每一行创建搜索向量。这就是它进行顺序扫描的原因。

Django 目前不支持 Meta.indexes 中的函数式索引，因此您需要手动创建它，例如使用 RunSQL operation.

RunSQL(
    """
    CREATE INDEX ON public_data_au_addresses USING GIN 
    (to_tsvector(...))
    """
)

to_tsvector() 表达式必须与查询中使用的表达式匹配。请务必通读 Postgres docs 以了解所有详细信息。

Answer 2

正如@knbk 已经建议的那样，为了提高性能，您必须阅读 Django 文档中的 Full-text search Performance 部分。

"If this approach becomes too slow, you can add a SearchVectorField to your model."

在您的代码中，您可以在模型中添加一个带有相关 GIN 索引的搜索向量字段，以及一个带有更新该字段的新方法的查询集：

from django.contrib.postgres.indexes import GinIndex
from django.contrib.postgres.search import SearchVector, SearchVectorField
from django.db import models
from postgres_copy import CopyQuerySet


class AddressesQuerySet(CopyQuerySet):

    def update_search_vector(self):
        return self.update(search_vector=SearchVector(
            'number', 'street', 'unit', 'city', 'region', 'postcode'
        ))


class Addresses(models.Model):
    date_update = models.DateTimeField(auto_now=True, null=True)
    longitude = models.DecimalField(max_digits=9, decimal_places=6, null=True)
    latitude = models.DecimalField(max_digits=9, decimal_places=6, null=True)
    number = models.CharField(max_length=16, null=True, default='')
    street = models.CharField(max_length=60, null=True, default='')
    unit = models.CharField(max_length=50, null=True, default='')
    city = models.CharField(max_length=50, null=True, default='')
    district = models.CharField(max_length=10, null=True, default='')
    region = models.CharField(max_length=5, null=True, default='')
    postcode = models.CharField(max_length=5, null=True, default='')
    addr_id = models.CharField(max_length=20, unique=True)
    addr_hash = models.CharField(max_length=20, unique=True)
    search_vector = SearchVectorField(null=True, editable=False)

    objects = AddressesQuerySet.as_manager()

    class Meta:
        indexes = [
            GinIndex(fields=['search_vector'], name='search_vector_idx')
        ]

您可以使用新的查询集方法更新新的搜索向量场：

>>> Addresses.objects.update_search_vector()
UPDATE "addresses_addresses"
SET "search_vector" = to_tsvector(
  COALESCE("addresses_addresses"."number", '') || ' ' ||
  COALESCE("addresses_addresses"."street", '') || ' ' ||
  COALESCE("addresses_addresses"."unit", '') || ' ' ||
  COALESCE("addresses_addresses"."city", '') || ' ' ||
  COALESCE("addresses_addresses"."region", '') || ' ' ||
  COALESCE("addresses_addresses"."postcode", '')
)

如果您执行查询并阅读说明，您可以看到使用的 GIN 索引：

>>> print(Addresses.objects.filter(search_vector='north').values('id').explain(verbose=True))
EXPLAIN (VERBOSE true)
SELECT "addresses_addresses"."id"
FROM "addresses_addresses"
WHERE "addresses_addresses"."search_vector" @@ (plainto_tsquery('north')) = true [0.80ms]
Bitmap Heap Scan on public.addresses_addresses  (cost=12.25..16.52 rows=1 width=4)
  Output: id
  Recheck Cond: (addresses_addresses.search_vector @@ plainto_tsquery('north'::text))
  ->  Bitmap Index Scan on search_vector_idx  (cost=0.00..12.25 rows=1 width=0)
        Index Cond: (addresses_addresses.search_vector @@ plainto_tsquery('north'::text))

如果您想进一步深入，可以阅读我写的一篇关于该主题的文章：

"Full-Text Search in Django with PostgreSQL"

更新

我尝试执行 Django ORM 生成的 SQL： http://sqlfiddle.com/#!17/f9aa9/1

Django 全文搜索优化 - Postgres

Django Full Text Search Optimization - Postgres

django

postgresql

full-text-search

autocomplete

django-models

更新

更新