大量连接行的日期比较慢

Question

以下查询将Snippets 分组为ChannelId 并返回UnreadSnippetCount。

为了确定UnreadSnippetCount，将Channel 加入ChannelUsers 以获取User 上次读取Channel 的日期，并使用此LastReadDate 将计数限制为sn 所在的行-p 是在用户最后一次阅读频道之后创建的。

    SELECT c.Id, COUNT(s.Id) as [UnreadSnippetCount] 
    FROM Channels c
    INNER JOIN ChannelUsers cu
        ON cu.ChannelId = c.Id
    LEFT JOIN Snippets s
        ON cu.ChannelId = s.ChannelId
        AND s.CreatedByUserId <> @UserId
    WHERE cu.UserId = @UserId
    AND (cu.LastReadDate IS NULL OR s.CreatedDate > cu.LastReadDate)
    AND c.Id IN (select value from STRING_SPLIT(@ChannelIds, ','))

    GROUP BY c.Id

查询在逻辑上运行良好，但对于具有大量 Snippets (97691) 的 Channels，查询可能需要 10 分钟或更长时间才能返回。

创建以下索引：

CREATE NONCLUSTERED INDEX [IX_Snippets_CreatedDate] ON [dbo].[Snippets]
(
    [CreatedDate] ASC
)WITH (STATISTICS_NORECOMPUTE = OFF, DROP_EXISTING = OFF, ONLINE = OFF, OPTIMIZE_FOR_SEQUENTIAL_KEY = OFF) ON [PRIMARY]
GO

更新：

查询执行计划（原始查询）：

https://www.brentozar.com/pastetheplan/?id=B19sI105F

更新 2

按照建议将where 子句移入join：

    SELECT c.Id, COUNT(s.Id) as [UnreadSnippetCount] 
    FROM Channels c
    INNER JOIN ChannelUsers cu
        ON cu.ChannelId = c.Id
    LEFT JOIN Snippets s
        ON cu.ChannelId = s.ChannelId
        AND s.CreatedByUserId <> @UserId
        AND s.CreatedDate > cu.LastReadDate
    WHERE cu.UserId = @UserId
    AND c.Id IN (select value from STRING_SPLIT(@ChannelIds, ',')

产生这个执行计划：

https://www.brentozar.com/pastetheplan/?id=HkqwFk0ct

我可以使用更好的日期比较方法吗？

更新 3 - 解决方案

索引

CREATE NONCLUSTERED INDEX [IX_Snippet_Created] ON [dbo].[Snippets]
  (ChannelId ASC, CreatedDate ASC) INCLUDE (CreatedByUserId);

存储过程

ALTER PROCEDURE [dbo].[GetUnreadSnippetCounts2]
(
    @ChannelIds ChannelIdsType READONLY,
    @UserId nvarchar(36)
)
AS
    SET NOCOUNT ON

SELECT
    c.Id,
    COUNT(s.Id) as [UnreadSnippetCount] 
FROM Channels c
JOIN @ChannelIds cid
    ON cid.Id = c.Id
INNER JOIN ChannelUsers cu
    ON cu.ChannelId = c.Id
    AND cu.UserId = @UserId
JOIN Snippets s
    ON cu.ChannelId = s.ChannelId
    AND s.CreatedByUserId <> @UserId
    AND (cu.LastReadDate IS NULL OR s.CreatedDate > cu.LastReadDate)
GROUP BY c.Id;

这会在逻辑上给出正确的结果并快速返回。

产生的执行计划：

https://www.brentozar.com/pastetheplan/?id=S1GwRCCcK

Answer 1

我可以在查询计划中看到许多效率低下的地方。

1. 使用STRING_SPLIT 意味着编译器不知道返回了多少值，或者它们是唯一的，并且数据类型不匹配。理想情况下，你会传入一个表值参数，但是如果你不能这样做，那么另一种解决方案是将它们转储到表变量中

   DECLARE @tmp TABLE (Id int PRIMARY KEY);
   INSERT @tmp (Id)
   select value
   from STRING_SPLIT(@ChannelIds, ',')
   

2. 您需要对Snippets 进行更好的索引。我会建议以下

   CREATE NONCLUSTERED INDEX [IX_Snippet_Created] ON [dbo].[Snippets]
     (ChannelId ASC, CreatedDate ASC) INCLUDE (CreatedByUserId);
   

   将CreatedByUserId 放在键中是没有意义的，因为它是一个不等式。保存在INCLUDE

3. 正如您已经被告知的，最好将条件（对于左连接表）移至ON 子句。我不知道你是否还需要cu.LastReadDate IS NULL 支票，我已经把它留在里面了。

4. 我必须说，我不清楚你的架构，但INNER JOIN ChannelUsers cu 在这里感觉不对，也许应该是LEFT JOIN？如果没有看到您的完整设置和所需的输出，我不能再多说。

SELECT
  c.Id,
  COUNT(s.Id) as [UnreadSnippetCount] 
FROM Channels c
JOIN @tmp t
    ON t.Id = c.Id
INNER JOIN ChannelUsers cu
    ON cu.ChannelId = c.Id
    AND cu.UserId = @UserId
LEFT JOIN Snippets s
    ON cu.ChannelId = s.ChannelId
    AND s.CreatedByUserId <> @UserId
    AND (cu.LastReadDate IS NULL OR s.CreatedDate > cu.LastReadDate)
GROUP BY c.Id;