那些遇到过的问题

使用“NOT IN”时性能不佳

RuS*_*SSe 6 performance sql-server t-sql query-performance

在我的应用程序中，我有一个在“文件”表中执行搜索的查询。

该表files按f.created（参见表定义）进行分区，并为客户端 19 ( f.cid = 19)提供约 1 亿行。

我正在使用我上一个问题的答案中的这个查询， SQL Server 的慢顺序：

WITH PartitionNumbers AS
(
    -- Each partition of the table
    SELECT P.partition_number
    FROM sys.partitions AS P
    WHERE P.[object_id] = OBJECT_ID(N'dbo.files', N'U')
    AND P.index_id = 1
)
SELECT
    FF.id,
    FF.[name],
    FF.[year],
    FF.cid,
    FF.created,
    vnVE0.keywordValueCol0_numeric
FROM PartitionNumbers AS PN
CROSS APPLY
(
    SELECT
        F100.*
    FROM 
    (
        -- 50 rows in order for year 2013
        SELECT
            F.id,
            F.[name],
            F.[year],
            F.cid,
            F.created
        FROM dbo.files AS F
        WHERE
            F.grapado IS NULL
            AND F.masterversion IS NULL
            AND F.[year] = 2013
            AND F.cid = 19
            AND F.eid = 8
            AND $PARTITION.PF_files_partitioning(F.created) = PN.partition_number
        ORDER BY
            F.[name]
            OFFSET 0 ROWS
            FETCH FIRST 50 ROWS ONLY

        UNION ALL

        -- 50 rows in order for year 0
        SELECT
            F.id,
            F.[name],
            F.[year],
            F.cid,
            F.created
        FROM dbo.files AS F
        WHERE
            F.grapado IS NULL
            AND F.masterversion IS NULL
            AND F.[year] = 0
            AND F.cid = 19
            AND F.eid = 8
            AND $PARTITION.PF_files_partitioning(F.created) = PN.partition_number
        ORDER BY
            F.[name]
            OFFSET 0 ROWS
            FETCH FIRST 50 ROWS ONLY
    ) AS F100
) AS FF
OUTER APPLY
(
    -- Lookup distinct values
    SELECT
        keywordValueCol0_numeric = 
            CASE
                WHEN VN.[value] IS NOT NULL AND VN.[value] <> ''
                THEN CONVERT(decimal(28, 2), VN.[value])
                ELSE CONVERT(decimal(28, 2), 0)
            END
    FROM dbo.value_number AS VN
    WHERE
        VN.id_file = FF.id
        AND VN.id_field = 260
    GROUP BY
        VN.[value]
) AS vnVE0
ORDER BY
    FF.[name]
    OFFSET 0 ROWS
    FETCH FIRST 50 ROWS ONLY;
Run Code Online (Sandbox Code Playgroud)

这里的关键是这个查询表现良好，但如果我改变f.eid = 8在WHERE给f.eid NOT IN (10,12)，查询变得太慢（超过10分钟）。

执行计划使用f.eid = 8：https : //www.brentozar.com/pastetheplan/?id=HJ_Fbb2qM

简单条件

执行计划使用f.eid NOT IN (8,10)：https : //www.brentozar.com/pastetheplan/?id=B1-zmbnqz

不在

我试着做用一个JOINentidades表，其id被引用eid的files表。

执行计划与entidades：https : //www.brentozar.com/pastetheplan/? id =rJarHZh5M

加入 entidades

我怎样才能提高这里的性能？

附加信息

分区功能PF_files_partitioning：

CREATE PARTITION FUNCTION PF_files_partitioning (DATETIME2(7))
AS
RANGE LEFT FOR VALUES ( '2013-03-31 23:59:59', 
                        '2013-06-30 23:59:59',
                        '2013-09-30 23:59:59',
                        '2013-12-31 23:59:59',
                        '2014-03-31 23:59:59', 
                        '2014-06-30 23:59:59',
                        '2014-09-30 23:59:59',
                        '2014-12-31 23:59:59',
                        '2015-03-31 23:59:59', 
                        '2015-06-30 23:59:59',
                        '2015-09-30 23:59:59',
                        '2015-12-31 23:59:59',
                        '2016-03-31 23:59:59', 
                        '2016-06-30 23:59:59',
                        '2016-09-30 23:59:59',
                        '2016-12-31 23:59:59',
                        '2017-03-31 23:59:59', 
                        '2017-06-30 23:59:59',
                        '2017-09-30 23:59:59',
                        '2017-12-31 23:59:59',
                        '2018-03-31 23:59:59')
Run Code Online (Sandbox Code Playgroud)

分区方案PS_files_partitioning：

CREATE PARTITION SCHEME PS_files_partitioning AS PARTITION PF_files_partitioning ALL TO ([PRIMARY]);
Run Code Online (Sandbox Code Playgroud)

注意：我将在每个分区中有大约 1500 万行。

表files：

CREATE TABLE [dbo].[files](
    [id] [bigint] IDENTITY(1,1) NOT NULL,
    [cid] [tinyint] NOT NULL,
    [eid] [bigint] NOT NULL,
    [cat_id] [bigint] NOT NULL,
    [tip_id] [bigint] NULL,
    [sub_id] [bigint] NULL,
    [year] [smallint] NOT NULL,
    [caducidad] [smallint] NULL,
    [grapadopri] [int] NOT NULL,
    [grapado] [bigint] NULL,
    [name] [nvarchar](255) NOT NULL,
    [extension] [tinyint] NOT NULL,
    [size] [bigint] NOT NULL,
    [id_doc] [bit] NOT NULL,
    [observaciones] [nvarchar](255) NOT NULL,
    [indexed] [bit] NOT NULL,
    [signed] [bit] NOT NULL,
    [created] [datetime2](7) NOT NULL,
    [name_lower] [nvarchar](255) NOT NULL,
    [modified] [datetime2](7) NULL,
    [related] [bit] NOT NULL,
    [masterversion] [bigint] NULL,
    [versioned] [bit] NOT NULL,
    [hwsignature] [tinyint] NOT NULL,
    [blockedUserId] [smallint] NULL,
 CONSTRAINT [PK_files_id] PRIMARY KEY CLUSTERED 
(
    [id] ASC,
    [created] ASC
)WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, IGNORE_DUP_KEY = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON)
ON PS_files_partitioning([created]),
 CONSTRAINT [files$estructure_unique] UNIQUE NONCLUSTERED 
(
    [cat_id] ASC,
    [tip_id] ASC,
    [sub_id] ASC,
    [year] ASC,
    [name] ASC,
    [grapado] ASC,
    [created] ASC
)WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, IGNORE_DUP_KEY = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON)
)

ALTER TABLE [dbo].[files]  WITH NOCHECK ADD  CONSTRAINT [FK_files_entidad] FOREIGN KEY([eid])
REFERENCES [dbo].[entidades] ([id])
ON UPDATE CASCADE
ON DELETE CASCADE

ALTER TABLE [dbo].[files] CHECK CONSTRAINT [FK_files_entidad]
Run Code Online (Sandbox Code Playgroud)

表value_number：

CREATE TABLE [dbo].[value_number](
    [id] [bigint] IDENTITY(1,1) NOT NULL,
    [id_file] [bigint] NOT NULL DEFAULT ((0)),
    [id_field] [bigint] NOT NULL DEFAULT ((0)),
    [value] [nvarchar](255) NULL DEFAULT (NULL),
    [id_doc] [bigint] NULL DEFAULT (NULL)
 CONSTRAINT [PK_value_number_id] PRIMARY KEY CLUSTERED 
(
    [id] ASC
)WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, IGNORE_DUP_KEY = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON)
)
Run Code Online (Sandbox Code Playgroud)

此外，该表value_number由以下分区函数进行分区：

CREATE PARTITION FUNCTION PF_value_number (bigint) AS RANGE LEFT 
FOR VALUES (
29999999,
59999999,
89999999,
119999999,
149999999,
179999999,
209999999,
239999999
)
Run Code Online (Sandbox Code Playgroud)

和这个分区方案：

CREATE PARTITION SCHEME PS_value_number 
AS PARTITION PF_value_number
ALL TO ([PRIMARY]);

CREATE TABLE [dbo].[value_number](
    [id] [bigint] IDENTITY(1,1) NOT NULL,
    [id_file] [bigint] NOT NULL DEFAULT ((0)),
    [id_field] [bigint] NOT NULL DEFAULT ((0)),
    [value] [nvarchar](255) NULL DEFAULT (NULL),
    [id_doc] [bigint] NULL DEFAULT (NULL)
 CONSTRAINT [PK_value_number_id] PRIMARY KEY CLUSTERED 
(
    [id] ASC
)WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, IGNORE_DUP_KEY = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON)
ON PS_value_number([id_file])
)
Run Code Online (Sandbox Code Playgroud)

表entidades：

CREATE TABLE [dbo].[entidades](
    [id] [bigint] IDENTITY(1,1) NOT NULL,
    [cid] [bigint] NOT NULL,
    [mid] [bigint] NULL,
    [id_doc] [bigint] NULL,
    [name] [nvarchar](150) NOT NULL,
    [sincro] [tinyint] NOT NULL,
    [comprobar] [tinyint] NOT NULL,
    [op1] [tinyint] NOT NULL,
    [op2] [tinyint] NOT NULL,
    [op3] [tinyint] NOT NULL,
    [index] [tinyint] NOT NULL,
    [can_be_parent] [tinyint] NOT NULL,
    [accounting] [nchar](2) NULL,
    [nota] [nvarchar](60) NULL,
    [alias] [nvarchar](45) NULL,
    [element_type] [tinyint] NOT NULL,
    [size] [bigint] NOT NULL,
 CONSTRAINT [PK_entidades_id] PRIMARY KEY CLUSTERED 
(
    [id] ASC
)WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, IGNORE_DUP_KEY = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON),
 CONSTRAINT [codigo] UNIQUE NONCLUSTERED 
(
    [cid] ASC,
    [name] ASC
)WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, IGNORE_DUP_KEY = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON)
)

ALTER TABLE [dbo].[entidades]  WITH CHECK ADD  CONSTRAINT [FK_entidades_entidad] FOREIGN KEY([mid])
REFERENCES [dbo].[entidades] ([id])
GO
Run Code Online (Sandbox Code Playgroud)

files表的索引：

CREATE NONCLUSTERED INDEX [files_clientes] ON [dbo].[files]
(
    [cid] ASC
)
INCLUDE ([id]) WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, SORT_IN_TEMPDB = OFF, DROP_EXISTING = OFF, ONLINE = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON)
ON PS_files_partitioning([created])

CREATE NONCLUSTERED INDEX [files_grapado] ON [dbo].[files]
(
    [grapado] ASC
)
INCLUDE (   [id],
    [name]) WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, SORT_IN_TEMPDB = OFF, DROP_EXISTING = OFF, ONLINE = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON)
ON PS_files_partitioning([created])

CREATE NONCLUSTERED INDEX [files_mv] ON [dbo].[files]
(
    [masterversion] ASC,
    [year] ASC,
    [cat_id] ASC,
    [cid] ASC,
    [eid] ASC,
    [grapado] ASC,
    [sub_id] ASC,
    [tip_id] ASC
)
INCLUDE (   [id],
    [name]) WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, SORT_IN_TEMPDB = OFF, DROP_EXISTING = OFF, ONLINE = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON)
ON PS_files_partitioning([created])

CREATE NONCLUSTERED INDEX [files_ocr] ON [dbo].[files]
(
    [cid] ASC,
    [grapado] ASC,
    [indexed] ASC,
    [masterversion] ASC,
    [extension] ASC
)
INCLUDE (   [id],
    [eid],
    [cat_id],
    [tip_id],
    [sub_id],
    [year],
    [name]) WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, SORT_IN_TEMPDB = OFF, DROP_EXISTING = OFF, ONLINE = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON)
ON PS_files_partitioning([created])

CREATE NONCLUSTERED INDEX [files_ocr2] ON [dbo].[files]
(
    [cid] ASC,
    [eid] ASC,
    [grapado] ASC,
    [indexed] ASC,
    [masterversion] ASC,
    [extension] ASC
)
INCLUDE (   [id],
    [cat_id],
    [tip_id],
    [sub_id],
    [year],
    [name]) WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, SORT_IN_TEMPDB = OFF, DROP_EXISTING = OFF, ONLINE = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON)
ON PS_files_partitioning([created])

CREATE NONCLUSTERED INDEX [files_ocr3] ON [dbo].[files]
(
    [cid] ASC,
    [cat_id] ASC,
    [grapado] ASC,
    [indexed] ASC,
    [masterversion] ASC,
    [extension] ASC
)
INCLUDE (   [eid],
    [tip_id],
    [sub_id],
    [year],
    [name]) WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, SORT_IN_TEMPDB = OFF, DROP_EXISTING = OFF, ONLINE = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON)
ON PS_files_partitioning([created])

CREATE NONCLUSTERED INDEX [busqueda_name] ON [dbo].[files]
(
    [cid] ASC,
    [eid] ASC,
    [grapado] ASC,
    [year] ASC
)
INCLUDE (   [id],
    [cat_id],
    [tip_id],
    [sub_id],
    [grapadopri],
    [name],
    [size],
    [id_doc],
    [signed],
    [created],
    [modified],
    [related],
    [masterversion]) WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, SORT_IN_TEMPDB = OFF, DROP_EXISTING = OFF, ONLINE = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON)
ON PS_files_partitioning([created])

CREATE NONCLUSTERED INDEX [busqueda2] ON [dbo].[files]
(
    [cid] ASC,
    [eid] ASC,
    [cat_id] ASC,
    [grapado] ASC,
    [masterversion] ASC,
    [year] ASC
)WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, SORT_IN_TEMPDB = OFF, DROP_EXISTING = OFF, ONLINE = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON)
ON PS_files_partitioning([created])

CREATE NONCLUSTERED INDEX [cid] ON [dbo].[files]
(
    [cid] ASC
)WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, SORT_IN_TEMPDB = OFF, DROP_EXISTING = OFF, ONLINE = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON)
ON PS_files_partitioning([created])

CREATE NONCLUSTERED INDEX [eid] ON [dbo].[files]
(
    [eid] ASC
)WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, SORT_IN_TEMPDB = OFF, DROP_EXISTING = OFF, ONLINE = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON)
ON PS_files_partitioning([created])

CREATE NONCLUSTERED INDEX [extension] ON [dbo].[files]
(
    [extension] ASC
)WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, SORT_IN_TEMPDB = OFF, DROP_EXISTING = OFF, ONLINE = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON)
ON PS_files_partitioning([created])

CREATE NONCLUSTERED INDEX [FK_files_archivo] ON [dbo].[files]
(
    [grapado] ASC
)WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, SORT_IN_TEMPDB = OFF, DROP_EXISTING = OFF, ONLINE = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON)
ON PS_files_partitioning([created])

CREATE NONCLUSTERED INDEX [FK_files_tipo] ON [dbo].[files]
(
    [tip_id] ASC
)WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, SORT_IN_TEMPDB = OFF, DROP_EXISTING = OFF, ONLINE = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON)
ON PS_files_partitioning([created])

CREATE NONCLUSTERED INDEX [grapadopri] ON [dbo].[files]
(
    [grapadopri] ASC
)WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, SORT_IN_TEMPDB = OFF, DROP_EXISTING = OFF, ONLINE = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON)
ON PS_files_partitioning([created])

CREATE NONCLUSTERED INDEX [index_all] ON [dbo].[files]
(
    [cid] ASC,
    [eid] ASC,
    [grapado] ASC,
    [masterversion] ASC
)
INCLUDE (   [cat_id],
    [tip_id],
    [sub_id],
    [year],
    [grapadopri],
    [name],
    [size],
    [id_doc],
    [signed],
    [created],
    [modified],
    [related],
    [versioned]) WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, SORT_IN_TEMPDB = OFF, DROP_EXISTING = OFF, ONLINE = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON)
ON PS_files_partitioning([created])

CREATE NONCLUSTERED INDEX [missing_index_7_6] ON [dbo].[files]
(
    [cid] ASC,
    [eid] ASC,
    [grapado] ASC,
    [name] ASC,
    [year] ASC
)WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, SORT_IN_TEMPDB = OFF, DROP_EXISTING = OFF, ONLINE = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON) 
ON PS_files_partitioning([created])

CREATE NONCLUSTERED INDEX [ocrCloudClients] ON [dbo].[files]
(
    [grapado] ASC,
    [indexed] ASC,
    [extension] ASC
)
INCLUDE (   [cid],
    [eid],
    [cat_id],
    [tip_id],
    [sub_id]) WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, SORT_IN_TEMPDB = OFF, DROP_EXISTING = OFF, ONLINE = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON)
ON PS_files_partitioning([created])

CREATE NONCLUSTERED INDEX [searchEntity] ON [dbo].[files]
(
    [cid] ASC,
    [eid] ASC,
    [grapado] ASC,
    [masterversion] ASC
)WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, SORT_IN_TEMPDB = OFF, DROP_EXISTING = OFF, ONLINE = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON)
ON PS_files_partitioning([created])

CREATE NONCLUSTERED INDEX [sub_id] ON [dbo].[files]
(
    [sub_id] ASC
)WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, SORT_IN_TEMPDB = OFF, DROP_EXISTING = OFF, ONLINE = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON)
ON PS_files_partitioning([created])

CREATE NONCLUSTERED INDEX [IX dbo.files cid, year, eid : grapado IS NULL AND masterversion IS NULL] ON [dbo].[files]
(
    [cid] ASC,
    [year] ASC,
    [eid] ASC
)
INCLUDE (   [grapado],
    [masterversion]) 
WHERE ([grapado] IS NULL AND [masterversion] IS NULL)
WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, SORT_IN_TEMPDB = OFF, DROP_EXISTING = OFF, ONLINE = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON)
ON PS_files_partitioning([created])

CREATE NONCLUSTERED INDEX [IX dbo.files cid, year, name : grapado IS NULL AND masterversion IS NULL] ON [dbo].[files]
(
    [cid] ASC,
    [year] ASC,
    [name] ASC
)
INCLUDE (   [grapado],
    [masterversion]) 
WHERE ([grapado] IS NULL AND [masterversion] IS NULL)
WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, SORT_IN_TEMPDB = OFF, DROP_EXISTING = OFF, ONLINE = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON)
ON PS_files_partitioning([created])

CREATE NONCLUSTERED INDEX [IX dbo.files cid, year, eid, name : grapado IS NULL AND masterversion IS NULL] ON [dbo].[files]
    (
        [cid] ASC,
        [year] ASC,
        [eid] ASC,
        [name] ASC
    )
    INCLUDE (   [grapado],
        [masterversion]) 
    WHERE ([grapado] IS NULL AND [masterversion] IS NULL)
    WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, SORT_IN_TEMPDB = OFF, DROP_EXISTING = OFF, ONLINE = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON)
    ON PS_files_partitioning([created])
Run Code Online (Sandbox Code Playgroud)

value_number表的索引：

CREATE NONCLUSTERED INDEX [searchValues] ON [dbo].[value_number]
(
    [id_field] ASC
)
INCLUDE (   [id_file],
    [value]) WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, SORT_IN_TEMPDB = OFF, DROP_EXISTING = OFF, ONLINE = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON)
ON PS_value_number([id_file])

CREATE NONCLUSTERED INDEX [search] ON [dbo].[value_number]
(
    [id_file] ASC,
    [id_field] ASC
)
INCLUDE (   [value]) WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, SORT_IN_TEMPDB = OFF, DROP_EXISTING = OFF, ONLINE = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON)
ON PS_value_number([id_file])

CREATE NONCLUSTERED INDEX [id_field] ON [dbo].[value_number]
(
    [id_field] ASC
)WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, SORT_IN_TEMPDB = OFF, DROP_EXISTING = OFF, ONLINE = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON)
ON PS_value_number([id_file])

CREATE NONCLUSTERED INDEX [FK_valueesN_documento] ON [dbo].[value_number]
(
    [id_doc] ASC
)WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, SORT_IN_TEMPDB = OFF, DROP_EXISTING = OFF, ONLINE = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON)
ON PS_value_number([id_file])

CREATE NONCLUSTERED INDEX [FK_valueesN_archivo] ON [dbo].[value_number]
(
    [id_file] ASC
)WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, SORT_IN_TEMPDB = OFF, DROP_EXISTING = OFF, ONLINE = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON)
ON PS_value_number([id_file])
Run Code Online (Sandbox Code Playgroud)

Joe*_*ish 7

性能差异是因为您有一个适合第一个查询 ( "f"."eid" = 8) 但不适用于第二个 ( F.eid NOT IN (8,10))的索引。令我感到非常困惑的一件事是查询实际上使用了不同的索引：

[dd_produccion_test2].[dbo].[files].[IX dbo.files cid, year, name : grapado IS NULL AND masterversion IS NULL]
[dd_produccion_test2].[dbo].[files].[IX dbo.files cid, year, eid, name : grapado IS NULL AND masterversion IS NULL] [F]
Run Code Online (Sandbox Code Playgroud)

您没有在问题中定义这些索引。您对此有一个定义：

CREATE NONCLUSTERED INDEX [IX dbo.files cid, year, eid : grapado IS NULL AND masterversion IS NULL] ON [dbo].[archivos]
(
    [cid] ASC,
    [year] ASC,
    [eid] ASC
)
INCLUDE (   [grapado],
    [masterversion]) 
WHERE ([grapado] IS NULL AND [masterversion] IS NULL)
WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, SORT_IN_TEMPDB = OFF, DROP_EXISTING = OFF, ONLINE = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON)
ON PS_files_partitioning([created])
Run Code Online (Sandbox Code Playgroud)

我将假设索引定义基于它们的名称遵循相同的模式。

让我们开始使用该索引与键列的快速查询cid，year，id，和name。您的快速查询在前三个键列上具有相等谓词，并ORDER BY在第四个键列上定义。这使您可以直接查找所需的行。没有剩余的IO。您只需要从索引中读取 50 行，第四个键列意味着数据已经按照您想要的方式排序。总而言之，查询可能会执行几千次索引查找之类的操作，因此速度很快也就不足为奇了。

对于慢查询，过滤器eid不再是等式谓词。这是一个非常重要的区别。与键列的索引cid，year，eid，name仍然覆盖，但使用它看起来像这样：

  • 第一个键列：相等过滤器
  • 第二个关键列：相等过滤器
  • 第三个键列：不等于过滤器
  • 第四个关键列： ORDER BY

在第三个键列上没有相等过滤器时，引擎不支持利用索引的有序性质来避免对第四列 ( name)进行排序。如果查询优化器将此索引用于此查询，则需要从索引中读取所有匹配的行并对它们进行排序以找到前 50 行。这种类型的操作可能有很大的估计成本。

相反，查询优化选取具有关键列的索引cid，year和name。使用前两个键列上的相等筛选器，SQL Server 能够利用索引的有序特性来避免显式排序name。但是，该eid列不存在于该索引中。这会导致保留顺序的键查找以过滤eid，这几乎可以肯定是导致性能问题的原因。我会尝试将其eid作为包含列添加到[dd_produccion_test2].[dbo].[files].[IX dbo.files cid, year, name : grapado IS NULL AND masterversion IS NULL]：

CREATE NONCLUSTERED INDEX 
    [IX dbo.files cid, year, name : grapado IS NULL AND masterversion IS NULL]
ON [dbo].[files]
    ( cid, year, name )
    INCLUDE ( eid )
    WHERE ( grapado IS NULL AND masterversion IS NULL ) 
... ;
Run Code Online (Sandbox Code Playgroud)

这应该避免键查找并提高性能，但您可能会执行一些剩余 IO，因为您无法立即查找与所有过滤器匹配的前 50 行。

您可能想知道为什么 SQL Server 选择执行如此糟糕的计划，或者为什么一个估计成本仅为 0.777646 优化器单元的计划需要超过 10 分钟的时间来执行。答案与行目标有关。我们来看看慢速计划中的key查找：

在此处输入图片说明

查询优化器认为在找到 50 行满足F.eid NOT IN (8,10). 也许您的统计数据已经过时（您的计划中的某些列有警告）或者它做出了过于乐观的假设。我怀疑它需要执行比 151 多得多的搜索。您可以尝试禁用计划中的行目标作为测试，但我怀疑如果不更改索引定义，您将不会看到良好的性能。

Pau*_*ite 6

分析

回顾您在此站点上的问题历史，您似乎正在努力寻找一种设计，该设计允许您执行具有不同过滤和排序要求的各种查询。并非所有这些过滤要求都可以通过简单的 b 树索引（例如NOT IN）来支持。

如果您有幸提前确切知道需要哪些过滤条件和顺序，并且组合的数量相对较少，那么熟练的数据库索引调优师或许能够提出合理数量的非聚集索引来满足这些需求.

另一方面，所需的索引数量可能变得不切实际。您关于过滤/排序要求的任何问题都没有足够的细节来进行评估。我们对您的数据关系也不够了解。但是从您已经拥有的重叠索引的数量来看，您似乎有沿着这条路走下去的风险。

推荐

作为替代方案，我建议您删除迄今为止为支持查询而创建的现有非聚集索引，并将它们替换为单个非聚集列存储索引。这些最近在 Azure SQL 数据库的标准层上可用。保留聚簇主键和任何约束或唯一索引以执行键。

基本思想是涵盖您在非聚集列存储索引中筛选和排序的所有列，并可选择按始终（或非常普遍）应用的任何条件筛选索引。例如：

CREATE NONCLUSTERED COLUMNSTORE INDEX 
    [NC dbo.files id, name, year, cid, eid, created, grapado, masterversion]
ON dbo.files
    (id, [name], [year], cid, eid, created, grapado, masterversion)
WHERE
    grapado IS NULL
    AND masterversion IS NULL;
Run Code Online (Sandbox Code Playgroud)

然后，这将使您能够非常快速地找到符合条件的行，适用于非常广泛的查询，其中一些很难或不可能用 b 树索引来适应。

例子

首先编写一个查询以返回files您的查询将返回的表的主键和 order-by 列。例如：

SELECT
    -- files table primary key and order by column
    F.id,
    F.created,
    F.[name]
FROM dbo.files AS F
WHERE
    F.grapado IS NULL
    AND F.masterversion IS NULL
    AND F.[year] IN (0, 2013)
    AND F.cid = 19
    AND F.eid NOT IN (10, 12)
ORDER BY
    F.[name] ASC
OFFSET 0 ROWS
FETCH NEXT 50 ROWS ONLY;
Run Code Online (Sandbox Code Playgroud)

无论您使用何种过滤或排序条件，您都会发现该查询的性能非常好。在这个阶段，我们不会从文件表中返回我们需要的所有列，以最小化我们需要排序的数据的大小。

现在我们有了所需 50 行的主键，我们可以扩展查询以添加剩余的列，包括来自任何查找表的列：

WITH FoundKeys AS
(
    SELECT
        -- files table primary key and order by column
        F.id,
        F.created,
        F.[name]
    FROM dbo.files AS F
        WITH (INDEX([NC dbo.files id, name, year, cid, eid, created, grapado, masterversion]))
    WHERE
        F.grapado IS NULL
        AND F.masterversion IS NULL
        AND F.[year] IN (0, 2013)
        AND F.cid = 19
        AND F.eid NOT IN (10, 12)
    ORDER BY
        F.[name] ASC
    OFFSET 0 ROWS
    FETCH NEXT 50 ROWS ONLY
)
SELECT
    F.id,
    F.[name],
    F.[year],
    F.cid,
    F.eid,
    F.created,
    vnVE0.keywordValueCol0_numeric
FROM FoundKeys AS FK
JOIN dbo.files AS F
    -- join on primary key
    ON F.id = FK.id
    AND F.created = FK.created
OUTER APPLY
(
    -- Lookup distinct values
    SELECT
        keywordValueCol0_numeric = 
            CASE
                WHEN VN.[value] IS NOT NULL AND VN.[value] <> ''
                THEN CONVERT(decimal(28, 2), VN.[value])
                ELSE CONVERT(decimal(28, 2), 0)
            END
    FROM dbo.value_number AS VN
    WHERE
        VN.id_file = F.id
        AND VN.id_field = 260
    GROUP BY
        VN.[value]
) AS vnVE0
ORDER BY
    FK.[name];
Run Code Online (Sandbox Code Playgroud)

这应该产生一个执行计划，如：

预计计划

我还鼓励您再次考虑对这些表进行分区。对我来说似乎根本没有必要，使查询过于复杂，并损害了您的唯一性约束。应该可以批量插入行并快速应用它们，而不会阻塞并发读取器，如果这是问题的话。

如果要将上述解决方案的性能与更简单的版本（排序可能较慢）进行比较，请尝试：

SELECT
    F.id,
    F.[name],
    F.[year],
    F.cid,
    F.eid,
    F.created,
    keywordValueCol0_numeric = 
        CASE
            WHEN vnVE0.[value] IS NOT NULL AND vnVE0.[value] <> ''
            THEN CONVERT(decimal(28, 2), vnVE0.[value])
            ELSE CONVERT(decimal(28, 2), 0)
        END
FROM dbo.files AS F
    WITH (INDEX([NC dbo.files id, name, year, cid, eid, created, grapado, masterversion]))
OUTER APPLY
(
    SELECT DISTINCT VN.[value]
    FROM dbo.value_number AS VN
    WHERE VN.id_file = F.id
    AND VN.id_field = 260
) AS vnVE0
WHERE
    F.grapado IS NULL
    AND F.masterversion IS NULL
    AND F.[year] IN (0, 2013)
    AND F.cid = 19
    AND F.eid NOT IN (10, 12)
ORDER BY
    F.[name] ASC
OFFSET 0 ROWS
FETCH NEXT 50 ROWS ONLY;
Run Code Online (Sandbox Code Playgroud)

更简单的查询

归档时间：

查看次数：

535 次

最近记录：

8 年，2 月 前
SQL Server 的慢速排序 12
更多相关链接
相关归档
什么时候应该使用唯一约束而不是唯一索引？ 238
对聚合使用索引视图 - 好得令人难以置信？ 29
对大量重复值使用什么索引？ 15
列出索引和约束 12
TSQL 为什么 Top 使用变量更快？ 11
使未使用的数据库脱机的性能优势 10
如何解决错误 Msg 547, Level 16, State 0, Line 1 6
将事务表记录传输到数据仓库的最佳 ETL 设计 5
SQL Server 如何在删除表后重用保留页 3
应用程序执行时存储过程出现 ANSI_NULLS 和 ANSI_WARNINGS 错误 2
难疑归档
MySQL 能否合理地对数十亿行进行查询？ 289
强制删除数据库，而其他人可能已连接 175
MongoDB 上的分片和复制之间的区别 113
如何在 SQL Server 中将日期和时间与 datetime2 结合起来？ 55
使用 SSD 时，数据库设计中的聚集索引的概念是否有意义？ 44
SQL Server - 授予对整个架构与对象的权限？ 39
MySQL：为什么 mysql.db 中有“测试”条目？ 37
对于 DST 之前或之后的日期，如何获得 UTC 和本地时间之间的正确偏移量？ 33
带索引的 JSONB 与 hstore 32
pg_stat_user_tables中n_live_tup和n_dead_tup是什么意思 26