PostgreSQL:计数查询花费太多时间

我的查询有一些问题-时间太长（2636124毫秒！）：

此查询由ORM（Django）生成。当我尝试（用ORM）执行它时，我的应用程序挂起，并且当我输入psql-psql挂起时

解释和分析：

Aggregate  (cost=329583550.40..329583550.41 rows=1 width=8) (actual 
time=2636109.932..2636109.933 rows=1 loops=1)
   ->  Seq Scan on dictionary_dictionary  (cost=0.00..329583390.76 
       rows=63856 width=0) (actual time=2636109.922..2636109.922 rows=0 loops=1)
           Filter: (NOT (SubPlan 1))
           Rows Removed by Filter: 127712
           SubPlan 1
             ->  Materialize  (cost=0.00..4821.74 rows=135828 width=4) (actual time=0.006..12.453 rows=63856 loops=127712)
                ->  Seq Scan on dictionary_frequencydata u1  (cost=0.00..3611.60 rows=135828 width=4) (actual time=0.299..95.915 rows=127712 loops=1)
                     Filter: (user_id = 1)
                     Rows Removed by Filter: 28054
 Planning time: 0.277 ms
 Execution time: 2636124.744 ms
 (11 wierszy)`

我来自Django的模特

class Dictionary(DateTimeModel):
    base_word = models.ForeignKey(BaseDictionary, related_name=_('dict_words'))
    word = models.CharField(max_length=64)
    version = models.ForeignKey(Version)

class FrequencyData(DateTimeModel):
    word = models.ForeignKey(Dictionary, related_name=_('frequency_data'))
    count = models.BigIntegerField(null=True, blank=True)
    source = models.ForeignKey(Source, related_name=_('frequency_data'), null=True, blank=True)
    user = models.ForeignKey(settings.AUTH_USER_MODEL, related_name=_('frequency_data'))
    user_ip_address = models.GenericIPAddressField(null=True, blank=True)
    date_of_checking = models.DateTimeField(null=True, blank=True)
    is_checked = models.BooleanField(default=False)

表格定义如下：

\d+字典\u字典
Tabela“public.dictionary\u dictionary”
Kolumna | Typ | Porównanie | Nullowalne | Domy"lnie | Przechowywanie | Cel statystyk | Opis
----------------------+--------------------------+------------+------------+--------------------------------------------------------------------+----------------+---------------+------
id | integer | not null | nextval（'dictionary_dictionary_id_seq'：：regclass）| plain |
创建日期|带时区的时间戳| |非空| |普通| |124;
日期|修改|带时区的时间戳| |非空| |普通| |
单词|字符变化（64）| |非空| |扩展| |
算法_版本| id |整数| |非空| |普通| |
基本单词id整数非空普通
索引：
“dictionary\u dictionary\u pkey”主键，btree（id）
“dictionary\u Phone\u algorithm\u version\u id\u 0f0af100”b树（algorithm\u version\u id）
“dictionary\u dictionary\u base\u word\u id\u 8db15cb4”b树（base\u word\u id）
Ograniczenia kluczy obcych：
“dictionary\u算法\u版本\u id\u 0f0af100\u fk\u拼音\u”外键（算法\u版本\u id）引用dictionary\u算法版本（id）可延迟初始延迟
“dictionary\u base\u word\u id\u 8db15cb4\u fk\u拼音”外键（base\u word\u id）引用dictionary\u语法dictionary（id）可延迟初始延迟
Wskazywany przez：
表“dictionary\u frequencydata”约束“dictionary\u word\u id\u c231110d\u fk\u拼音”外键（word\u id）引用dictionary\u dictionary（id）可延迟初始延迟
=========
\d+字典\u频率数据
Tabela“public.dictionary\u frequencydata”
Kolumna | Typ | Porównanie | Nullowalne | Domy"lnie | Przechowywanie | Cel statystyk | Opis
------------------+--------------------------+------------+------------+---------------------------------------------------------------+----------------+---------------+------
id | integer | not null | nextval（'dictionary_frequencydata_id_seq'：：regclass）| plain |
创建日期|带时区的时间戳| |非空| |普通| |124;
日期|修改|带时区的时间戳| |非空| |普通| |
计数| bigint | | | | |普通| |
用户| ip |地址| inet | | | | main |124;
日期|检查|带时区的时间戳| | | |普通| |124;
是否选中|布尔| |非空| |普通| |
源代码|整数| | | |普通| |124;
用户id |整数| |非空| |普通| |
单词| id |整数| |非空| |普通| |
索引：
“dictionary\u frequencydata\u pkey”主键，btree（id）
“字典频率数据源id”B树（源id）
“字典频率数据用户id c6dfedce”B树（用户id）
“dictionary\u frequencydata\u word\u id\u c231110d”B树（word\u id）
Ograniczenia kluczy obcych：
“dictionary\u source\u id\u 38bb205a\u fk\u peopologic\u”外键（source\u id）引用dictionary\u frequencysource（id）可延迟初始延迟
“dictionary\u user\u id\u c6dfedce\u fk\u auth\u user”外键（user\u id）引用最初延迟的auth\u user（id）
“dictionary\u word\u id\u c231110d\u fk\u拼音”外键（word\u id）引用dictionary\u dictionary（id）可延迟初始延迟

它是共享主机。

---

Dictionary db tabel-120k rows FrequencyData-160k rows

在这种情况下，如果您像下面那样重新编写查询，您的查询应该会快得多，因为两个子查询都很快。最终结果相当于django生成的查询

在

dictionary\u dictionary

上使用过滤器操作的seq扫描似乎相当昂贵，但是

class Dictionary(DateTimeModel):
    base_word = models.ForeignKey(BaseDictionary, related_name=_('dict_words'))
    word = models.CharField(max_length=64)
    version = models.ForeignKey(Version)

class FrequencyData(DateTimeModel):
    word = models.ForeignKey(Dictionary, related_name=_('frequency_data'))
    count = models.BigIntegerField(null=True, blank=True)
    source = models.ForeignKey(Source, related_name=_('frequency_data'), null=True, blank=True)
    user = models.ForeignKey(settings.AUTH_USER_MODEL, related_name=_('frequency_data'))
    user_ip_address = models.GenericIPAddressField(null=True, blank=True)
    date_of_checking = models.DateTimeField(null=True, blank=True)
    is_checked = models.BooleanField(default=False)

SELECT 
tot - excl
from (select count(*) tot
      from dictionary_dictionary) t1
, (select count(DISTINCT d.id) excl
   from dictionary_dictionary d 
   join dictionary_frequencydata f
     on d.id = f.word_id 
   where f.user_id = 1 ) t2

SELECT COUNT(*) AS "__count" 
FROM "dictionary_dictionary" 
WHERE NOT ("dictionary_dictionary"."id" IN (SELECT distinct U1."word_id" AS Col1
                                            FROM "dictionary_frequencydata" U1 
                                            WHERE U1."user_id" = 1));