Question

我有一个非常基本的图像上传服务实现，您可以上传图像并标记它们。这是我的架构：

CREATE TABLE Tag(
    orm_id INTEGER PRIMARY KEY AUTOINCREMENT, 
    pid_high UNSIGNED BIG INT NOT NULL, 
    pid_low UNSIGNED BIG INT NOT NULL, 
    name STRING NOT NULL, 
    CONSTRAINT KeyConstraint UNIQUE (pid_high, pid_low) ON CONFLICT FAIL);

CREATE TABLE TagBridge(
    orm_id INTEGER PRIMARY KEY AUTOINCREMENT, 
    pid_high UNSIGNED BIG INT NOT NULL, 
    pid_low UNSIGNED BIG INT NOT NULL, 
    image_id_high UNSIGNED BIG INT NOT NULL, 
    image_id_low UNSIGNED BIG INT NOT NULL, 
    tag_id_high UNSIGNED BIG INT NOT NULL, 
    tag_id_low UNSIGNED BIG INT NOT NULL, 
    CONSTRAINT KeyConstraint UNIQUE (pid_high, pid_low) ON CONFLICT FAIL);

CREATE TABLE Image(
    orm_id INTEGER PRIMARY KEY AUTOINCREMENT, 
    pid_high UNSIGNED BIG INT NOT NULL,
    pid_low UNSIGNED BIG INT NOT NULL, 
    filehash STRING NOT NULL, 
    mime STRING NOT NULL, 
    uploadedDate INTEGER NOT NULL, 
    ratingsAverage REAL, 
    CONSTRAINT KeyConstraint UNIQUE (pid_high, pid_low) ON CONFLICT FAIL);

和指数

CREATE INDEX ImageTest on Image(pid_high, pid_low, uploadedDate DESC);
CREATE INDEX ImagefilehashIndex ON Image (filehash);
CREATE INDEX ImageuploadedDateIndex ON Image (uploadedDate);
CREATE INDEX TagnameIndex ON Tag (name);

有pid_high / pid_low字段而不是标准主键的原因是因为此服务使用客户端权威的128位GUID，但这不会显着影响查询速度。

由于这是互联网，此服务上的绝大多数图像都是猫，并标有“猫”。事实上，50,000张图片中约有47,000张标有“猫”字样。获取所有标记为'cat'的图像的查询是

select i.* from Tag t, TagBridge b, Image i 
where 
    b.tag_id_high = t.pid_high AND b.tag_id_low = t.pid_low 
AND b.image_id_high = i.pid_high and b.image_id_low = i.pid_low 
AND t.name ='cat' 
order by uploadedDate DESC LIMIT 20;

此查询计划是

sele  order          from  deta
----  -------------  ----  ----
0     0              0     SEARCH TABLE Tag AS t USING INDEX TagnameIndex (name=?) (~1 rows)
0     1              1     SCAN TABLE TagBridge AS b (~472 rows)
0     2              2     SEARCH TABLE Image AS i USING INDEX ImageTest (pid_high=? AND pid_low=?) (~1 rows)
0     0              0     USE TEMP B-TREE FOR ORDER BY

这里的主要问题是最后一行，USE TEMP B-TREE FOR ORDER BY。这会显着减慢查询速度。如果没有'order by'子句，整个查询大约需要0.001秒才能运行。使用order by子句，查询需要0.483秒，这是400倍的性能损失。

我想在0.1秒内得到这个查询，但我不知道如何。我已经尝试了许多其他查询，添加和删除索引，但这是我能够运行的最快的。

Answer 1

这是在过滤和排序索引之间进行选择的一般问题：

的 Choosing index

你应该保留一份热门标签列表（订购索引更有利），如果标签很受欢迎，请以某种方式禁止过滤索引，例如：

SELECT i.* FROM Tag t, TagBridge b, Image i WHERE b.tag_id_high = t.pid_high AND b.tag_id_low = t.pid_low AND b.image_id_high = i.pid_high AND b.image_id_low = i.pid_low AND t.name || '' = 'cat' ORDER BY i.uploadedDate DESC LIMIT 20

或者，您可以对模式进行非规范化，并将uploadedDate添加到TagBridge，并使用触发器或其他任何内容填充它。然后在TagBridge (pid_high, pid_low, uploadedDate, image_id_high, image_id_low)上创建一个复合索引，并稍微重写一下查询：

SELECT i.* FROM TagBridge b, Image i WHERE b.tag_id_high = ( SELECT t.pid_high FROM Tag t WHERE t.name = 'cat' ) AND b.tag_id_low = ( SELECT t.pid_low FROM Tag t WHERE t.name = 'cat' ) AND i.pid_high = b.image_id_high AND i.pid_low = b.image_id_low ORDER BY b.uploadedDate DESC LIMIT 20;

双子查询是因为SQLite不理解元组语法。

从SQLite查询中删除临时B树排序

1 个答案: