在使用少量数据进行分区时获得现实的查询计划

Question

在使用少量数据进行分区时获得现实的查询计划

Sim*_*ell 9 sql-server execution-plan partitioning sql-server-2012

我们正在使用分区来减少由于锁定而阻塞我们 OLTP 系统体验的数量，分区方案根据客户 ID 将工作表拆分为 100 个分区。然而，我们在测试过程中发现执行计划的选择方式并不完全符合我们的预期。

测试场景是具有 300,000 条联系人记录的单个客户（每个联系人的数据分布在两个表中），所有记录都位于单个分区中，并通过查询在客户分区中查找 500 条特定行。您会期望像哈希匹配这样的东西来消除不需要的 299,500 以在计划的早期启动，但看起来 SQL Server 选择获取整个表的记录数并在考虑如何之前对所有分区进行平均它将处理许多记录，这导致它选择嵌套循环并在该过程的后期消除不需要的记录。通常，对于非分区表进行相同查询所需的时间是其 9 倍。

奇怪的是，向选择添加一个选项（重新编译）给出了一个明智的计划，但我不知道为什么这会有所作为。这不是存储过程，在测试期间，我们在每次测试运行之前清除过程缓存。

当所涉及的表未分区时，不会看到此行为，即每次选择适当的计划，因为估计的行数与实际数相匹配

对这种行为的任何见解将不胜感激。

架构设置：

USE [Scratch]
GO
CREATE SCHEMA part
GO
CREATE PARTITION FUNCTION [ContactPartition](smallint) AS RANGE LEFT FOR VALUES (0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98)
GO
CREATE PARTITION SCHEME [ContactPartitionScheme] AS PARTITION [ContactPartition] ALL TO ([PRIMARY])
GO
CREATE TABLE [part].[Contact](
    [ContactId] [int] IDENTITY(1,1) NOT NULL,
    [CustomerPartitionKey] [smallint] NOT NULL,
    [CustomerId] [int] NOT NULL,
    [OptOut] [bit] NOT NULL,
 CONSTRAINT [cn_pk_cluContact_CustomerPartitionKey_ContactId] PRIMARY KEY CLUSTERED 
(
    [CustomerPartitionKey] ASC,
    [ContactId] ASC
) ON [ContactPartitionScheme]([CustomerPartitionKey])
) ON [ContactPartitionScheme]([CustomerPartitionKey])
GO
CREATE TABLE [part].[ContactIdentifier](
    [ContactIdentifierId] [int] IDENTITY(1,1) NOT NULL,
    [CustomerPartitionKey] [smallint] NOT NULL,
    [ContactId] [int] NOT NULL,
    [Identifier] [nvarchar](256) NOT NULL,
    [CustomerId] [int] NOT NULL,
 CONSTRAINT [cn_pk_cluContactIdentifier_CustomerPartitionKey_ContactId] PRIMARY KEY CLUSTERED 
(
    [CustomerPartitionKey] ASC,
    [ContactId] ASC
) ON [ContactPartitionScheme]([CustomerPartitionKey])
) ON [ContactPartitionScheme]([CustomerPartitionKey])
GO
CREATE NONCLUSTERED INDEX [idx_ncContactIdentifier_CustomerPartitionKey_ContactId] ON [part].[ContactIdentifier]
(
    [CustomerPartitionKey] ASC,
    [ContactId] ASC
)
 ON [ContactPartitionScheme]([CustomerPartitionKey])
GO
CREATE NONCLUSTERED INDEX [idx_ncContactIdentifier_CustomerPartitionKey_IdentifierType_Identifier] ON [part].[ContactIdentifier]
(
    [CustomerPartitionKey] ASC,
    [Identifier] ASC
) ON [ContactPartitionScheme]([CustomerPartitionKey])
GO
CREATE UNIQUE NONCLUSTERED INDEX [idx_ncuContactIdentifier_CustomerId_CustomerPartitionKey_Identifier] ON [part].[ContactIdentifier]
(
    [CustomerId] ASC,
    [CustomerPartitionKey] ASC,
    [Identifier] ASC
) ON [ContactPartitionScheme]([CustomerPartitionKey])
GO
ALTER TABLE [part].[ContactIdentifier]  WITH NOCHECK ADD  CONSTRAINT [cn_ContactIdentifier_CustomerPartitionKey_ContactId_fk_Contact_CustomerPartitionKey_ContactId] FOREIGN KEY([CustomerPartitionKey], [ContactId])
REFERENCES [part].[Contact] ([CustomerPartitionKey], [ContactId])
GO
ALTER TABLE [part].[ContactIdentifier] NOCHECK CONSTRAINT [cn_ContactIdentifier_CustomerPartitionKey_ContactId_fk_Contact_CustomerPartitionKey_ContactId]
GO
WITH TestData AS
(
    SELECT 1 As Ordinal
    UNION ALL
    SELECT td.Ordinal + 1
    FROM TestData td
    WHERE td.Ordinal < 30000
)
INSERT INTO part.Contact (CustomerPartitionKey, CustomerId, OptOut)
SELECT 3, 3, 0
FROM TestData OPTION (MAXRECURSION 30000);
GO 10
WITH TestData AS
(
    SELECT 1 AS Ordinal, ISNULL(MAX(ContactId) + 1, 1) AS ContactId FROM part.ContactIdentifier
    UNION ALL
    SELECT td.Ordinal + 1, td.ContactId + 1 AS ContactId
    FROM TestData td
    WHERE td.Ordinal < 30000
)
INSERT INTO part.ContactIdentifier (CustomerPartitionKey, CustomerId, ContactId, Identifier)
SELECT 3, 3, ContactId, CONCAT('User ', ContactId)
FROM TestData OPTION (MAXRECURSION 30000);
GO 10

Run Code Online (Sandbox Code Playgroud)

询问

USE Scratch
GO
DBCC FREEPROCCACHE
SET STATISTICS IO ON
SET STATISTICS TIME ON
GO
DECLARE @CustomerId int = 3, @CustomerPartitionKey smallint = 3

SET NOCOUNT ON

CREATE TABLE #identifiers (Ordinal int NOT NULL, Identifier nvarchar(256) NOT NULL)
INSERT INTO #identifiers
VALUES(0,N'User 0')
,(1,N'User 1')
,(2,N'User 2')
,(3,N'User 3')
,(4,N'User 4')
,(5,N'User 5')
,(6,N'User 6')
,(7,N'User 7')
,(8,N'User 8')
,(9,N'User 9')
,(10,N'User 10')
,(11,N'User 11')
,(12,N'User 12')
,(13,N'User 13')
,(14,N'User 14')
,(15,N'User 15')
,(16,N'User 16')
,(17,N'User 17')
,(18,N'User 18')
,(19,N'User 19')
,(20,N'User 20')
,(21,N'User 21')
,(22,N'User 22')
,(23,N'User 23')
,(24,N'User 24')
,(25,N'User 25')
,(26,N'User 26')
,(27,N'User 27')
,(28,N'User 28')
,(29,N'User 29')
,(30,N'User 30')
,(31,N'User 31')
,(32,N'User 32')
,(33,N'User 33')
,(34,N'User 34')
,(35,N'User 35')
,(36,N'User 36')
,(37,N'User 37')
,(38,N'User 38')
,(39,N'User 39')
,(40,N'User 40')
,(41,N'User 41')
,(42,N'User 42')
,(43,N'User 43')
,(44,N'User 44')
,(45,N'User 45')
,(46,N'User 46')
,(47,N'User 47')
,(48,N'User 48')
,(49,N'User 49')
,(50,N'User 50')
,(51,N'User 51')
,(52,N'User 52')
,(53,N'User 53')
,(54,N'User 54')
,(55,N'User 55')
,(56,N'User 56')
,(57,N'User 57')
,(58,N'User 58')
,(59,N'User 59')
,(60,N'User 60')
,(61,N'User 61')
,(62,N'User 62')
,(63,N'User 63')
,(64,N'User 64')
,(65,N'User 65')
,(66,N'User 66')
,(67,N'User 67')
,(68,N'User 68')
,(69,N'User 69')
,(70,N'User 70')
,(71,N'User 71')
,(72,N'User 72')
,(73,N'User 73')
,(74,N'User 74')
,(75,N'User 75')
,(76,N'User 76')
,(77,N'User 77')
,(78,N'User 78')
,(79,N'User 79')
,(80,N'User 80')
,(81,N'User 81')
,(82,N'User 82')
,(83,N'User 83')
,(84,N'User 84')
,(85,N'User 85')
,(86,N'User 86')
,(87,N'User 87')
,(88,N'User 88')
,(89,N'User 89')
,(90,N'User 90')
,(91,N'User 91')
,(92,N'User 92')
,(93,N'User 93')
,(94,N'User 94')
,(95,N'User 95')
,(96,N'User 96')
,(97,N'User 97')
,(98,N'User 98')
,(99,N'User 99')
,(100,N'User 100')
,(101,N'User 101')
,(102,N'User 102')
,(103,N'User 103')
,(104,N'User 104')
,(105,N'User 105')
,(106,N'User 106')
,(107,N'User 107')
,(108,N'User 108')
,(109,N'User 109')
,(110,N'User 110')
,(111,N'User 111')
,(112,N'User 112')
,(113,N'User 113')
,(114,N'User 114')
,(115,N'User 115')
,(116,N'User 116')
,(117,N'User 117')
,(118,N'User 118')
,(119,N'User 119')
,(120,N'User 120')
,(121,N'User 121')
,(122,N'User 122')
,(123,N'User 123')
,(124,N'User 124')
,(125,N'User 125')
,(126,N'User 126')
,(127,N'User 127')
,(128,N'User 128')
,(129,N'User 129')
,(130,N'User 130')
,(131,N'User 131')
,(132,N'User 132')
,(133,N'User 133')
,(134,N'User 134')
,(135,N'User 135')
,(136,N'User 136')
,(137,N'User 137')
,(138,N'User 138')
,(139,N'User 139')
,(140,N'User 140')
,(141,N'User 141')
,(142,N'User 142')
,(143,N'User 143')
,(144,N'User 144')
,(145,N'User 145')
,(146,N'User 146')
,(147,N'User 147')
,(148,N'User 148')
,(149,N'User 149')
,(150,N'User 150')
,(151,N'User 151')
,(152,N'User 152')
,(153,N'User 153')
,(154,N'User 154')
,(155,N'User 155')
,(156,N'User 156')
,(157,N'User 157')
,(158,N'User 158')
,(159,N'User 159')
,(160,N'User 160')
,(161,N'User 161')
,(162,N'User 162')
,(163,N'User 163')
,(164,N'User 164')
,(165,N'User 165')
,(166,N'User 166')
,(167,N'User 167')
,(168,N'User 168')
,(169,N'User 169')
,(170,N'User 170')
,(171,N'User 171')
,(172,N'User 172')
,(173,N'User 173')
,(174,N'User 174')
,(175,N'User 175')
,(176,N'User 176')
,(177,N'User 177')
,(178,N'User 178')
,(179,N'User 179')
,(180,N'User 180')
,(181,N'User 181')
,(182,N'User 182')
,(183,N'User 183')
,(184,N'User 184')
,(185,N'User 185')
,(186,N'User 186')
,(187,N'User 187')
,(188,N'User 188')
,(189,N'User 189')
,(190,N'User 190')
,(191,N'User 191')
,(192,N'User 192')
,(193,N'User 193')
,(194,N'User 194')
,(195,N'User 195')
,(196,N'User 196')
,(197,N'User 197')
,(198,N'User 198')
,(199,N'User 199')
,(200,N'User 200')
,(201,N'User 201')
,(202,N'User 202')
,(203,N'User 203')
,(204,N'User 204')
,(205,N'User 205')
,(206,N'User 206')
,(207,N'User 207')
,(208,N'User 208')
,(209,N'User 209')
,(210,N'User 210')
,(211,N'User 211')
,(212,N'User 212')
,(213,N'User 213')
,(214,N'User 214')
,(215,N'User 215')
,(216,N'User 216')
,(217,N'User 217')
,(218,N'User 218')
,(219,N'User 219')
,(220,N'User 220')
,(221,N'User 221')
,(222,N'User 222')
,(223,N'User 223')
,(224,N'User 224')
,(225,N'User 225')
,(226,N'User 226')
,(227,N'User 227')
,(228,N'User 228')
,(229,N'User 229')
,(230,N'User 230')
,(231,N'User 231')
,(232,N'User 232')
,(233,N'User 233')
,(234,N'User 234')
,(235,N'User 235')
,(236,N'User 236')
,(237,N'User 237')
,(238,N'User 238')
,(239,N'User 239')
,(240,N'User 240')
,(241,N'User 241')
,(242,N'User 242')
,(243,N'User 243')
,(244,N'User 244')
,(245,N'User 245')
,(246,N'User 246')
,(247,N'User 247')
,(248,N'User 248')
,(249,N'User 249')
,(250,N'User 250')
,(251,N'User 251')
,(252,N'User 252')
,(253,N'User 253')
,(254,N'User 254')
,(255,N'User 255')
,(256,N'User 256')
,(257,N'User 257')
,(258,N'User 258')
,(259,N'User 259')
,(260,N'User 260')
,(261,N'User 261')
,(262,N'User 262')
,(263,N'User 263')
,(264,N'User 264')
,(265,N'User 265')
,(266,N'User 266')
,(267,N'User 267')
,(268,N'User 268')
,(269,N'User 269')
,(270,N'User 270')
,(271,N'User 271')
,(272,N'User 272')
,(273,N'User 273')
,(274,N'User 274')
,(275,N'User 275')
,(276,N'User 276')
,(277,N'User 277')
,(278,N'User 278')
,(279,N'User 279')
,(280,N'User 280')
,(281,N'User 281')
,(282,N'User 282')
,(283,N'User 283')
,(284,N'User 284')
,(285,N'User 285')
,(286,N'User 286')
,(287,N'User 287')
,(288,N'User 288')
,(289,N'User 289')
,(290,N'User 290')
,(291,N'User 291')
,(292,N'User 292')
,(293,N'User 293')
,(294,N'User 294')
,(295,N'User 295')
,(296,N'User 296')
,(297,N'User 297')
,(298,N'User 298')
,(299,N'User 299')

SELECT 
    CI.ContactId,
    I.Ordinal,
    I.Identifier
FROM    #identifiers I
JOIN    part.ContactIdentifier AS CI ON CI.CustomerId = @CustomerId AND CI.CustomerPartitionKey = @CustomerPartitionKey AND 
                                        CI.Identifier = I.Identifier
JOIN    part.Contact AS C ON C.CustomerPartitionKey = @CustomerPartitionKey AND C.ContactId = CI.ContactId
WHERE   C.OptOut = 0

DROP TABLE #identifiers

Run Code Online (Sandbox Code Playgroud)

糟糕的执行计划：http : //pastebin.com/Us7HY4KF

Answer 1

Geo*_*son 4

看起来 SQL Server 正在生成一个参数化查询计划，该计划可以适用于 @CustomerPartitionKey 的任何值。为了做到这一点，它似乎将 @CustomerPartitionKey 视为您正在寻找的分区和列。如果我们看一下有错误估计的查询计划（估计有 3000 行，实际有 300000 行），我们会发现实际上有两个独立的查找谓词与part.Contact相关@CustomerPartitionKey：

Seek Keys[1]: Prefix: PtnId1004, [Test].[part].[Contact].CustomerPartitionKey = Scalar Operator([Expr1008]), Scalar Operator([@CustomerPartitionKey])

我认为后者 ([Test].[part].[Contact].CustomerPartitionKey = Scalar Operator([@CustomerPartitionKey])能够根据的值的参数嗅探获得正确的估计@CustomerPartitionKey。但是，前者 ( Prefix: PtnId1004 = Scalar Operator([Expr1008])) 可能无法做到这一点，也许Expr1008是因为处理分区消除的复杂表达式：[Expr1008]=RangePartitionNew([@CustomerPartitionKey],(0),(0),(1),(2),...,(97),(98))。

在这种情况下，有 100 个分区，行估计值正好低了 100 倍，因为 SQL Server 无法以与处理列上的实际查找并使用估计值相同的智能方式来处理分区消除。运行时参数值为 3。如果删除分区，估计的行数会发生变化，这一点支持了这一理论；如果您使用 90 个分区，则估计值为 3333.33 (300000 / 90)。

3在我们自己的查询中，每当我们编写要利用分区消除的查询时，我们通常使用文字（例如，在本例中）或使用 OPTION RECOMPILE。这种做法对我们来说效果相当好，因为系统上的查询数量并不多，而且针对大型分区表的查询的查询编译开销不是我们关心的问题。不一定是令人满意的答案，但它可能对您有用。

归档时间：	10 年，6 月前
查看次数：	507 次
最近记录：	10 年，6 月前