这是一个更新和改进的抽样形式。它基于与其他一些使用CHECKSUM / BINARY_CHECKSUM和modulus的答案相同的概念。

使用与此类似的实现的原因，而不是其他答案:

It is relatively fast over huge data sets and can be efficiently used in/with derived queries. Millions of pre-filtered rows can be sampled in seconds with no tempdb usage and, if aligned with the rest of the query, the overhead is often minimal. Does not suffer from CHECKSUM(*) / BINARY_CHECKSUM(*) issues with runs of data. When using the CHECKSUM(*) approach, the rows can be selected in "chunks" and not "random" at all! This is because CHECKSUM prefers speed over distribution. Results in a stable/repeatable row selection and can be trivially changed to produce different rows on subsequent query executions. Approaches that use NEWID(), such as CHECKSUM(NEWID()) % 100, can never be stable/repeatable. Allows for increased sample precision and reduces introduced statistical errors. The sampling precision can also be tweaked. CHECKSUM only returns an int value. Does not use ORDER BY NEWID(), as ordering can become a significant bottleneck with large input sets. Avoiding the sorting also reduces memory and tempdb usage. Does not use TABLESAMPLE and thus works with a WHERE pre-filter.

缺点/限制:

Slightly slower execution times and using CHECKSUM(*). Using hashbytes, as shown below, adds about 3/4 of a second of overhead per million lines. This is with my data, on my database instance: YMMV. This overhead can be eliminated if using a persisted computed column of the resulting 'well distributed' bigint value from HASHBYTES. Unlike the basic SELECT TOP n .. ORDER BY NEWID(), this is not guaranteed to return "exactly N" rows. Instead, it returns a percentage row rows where such a value is pre-determined. For very small sample sizes this could result in 0 rows selected. This limitation is shared with the CHECKSUM(*) approaches.

要点如下:

-- Allow a sampling precision [0, 100.0000].
declare @sample_percent decimal(7, 4) = 12.3456

select
    t.*
from t
where 1=1
    and t.Name = 'Mr. No Questionable Checksum Usages'
    and ( -- sample
        @sample_percent = 100
        or abs(
            -- Choose appropriate identity column(s) for hashbytes input.
            -- For demonstration it is assumed to be a UNIQUEIDENTIFIER rowguid column.
            convert(bigint, hashbytes('SHA1', convert(varbinary(32), t.rowguid)))
        ) % (1000 * 100) < (1000 * @sample_percent)
    )

注:

While SHA1 is technically deprecated since SQL Server 2016, it is both sufficient for the task and is slightly faster than either MD5 or SHA2_256. Use a different hashing function as relevant. If the table already contains a hashed column (with a good distribution), that could potentially be used as well. Conversion of bigint is critical as it allows 2^63 bits of 'random space' to which to apply the modulus operator; this is much more than the 2^31 range from the CHECKSUM result. This reduces the modulus error at the limit, especially as the precision is increased. The sampling precision can be changed as long as the modulus operand and sample percent are multiplied appropriately. In this case, that is 1000 * to account for the 4 digits of precision allowed in @sample_percent. Can multiply the bigint value by RAND() to return a different row sample each run. This effectively changes the permutation of the fixed hash values. If @sample_percent is 100 the query planner can eliminate the slower calculation code entirely. Remember 'parameter sniffing' rules. This allows the code to be left in the query regardless of enabling sampling.

计算@sample_percent，带下限/上限，并在查询中添加TOP“提示”，这在示例用于派生表上下文时可能有用。

-- Approximate max-sample and min-sample ranges.
-- The minimum sample percent should be non-zero within the precision.
declare @max_sample_size int = 3333333
declare @min_sample_percent decimal(7,4) = 0.3333
declare @sample_percent decimal(7,4) -- [0, 100.0000]
declare @sample_size int

-- Get initial count for determining sample percentages.
-- Remember to match the filter conditions with the usage site!
declare @rows int
select @rows = count(1)
    from t
    where 1=1
        and t.Name = 'Mr. No Questionable Checksum Usages'

-- Calculate sample percent and back-calculate actual sample size.
if @rows <= @max_sample_size begin
    set @sample_percent = 100
end else begin
    set @sample_percent = convert(float, 100) * @max_sample_size / @rows
    if @sample_percent < @min_sample_percent
        set @sample_percent = @min_sample_percent
end
set @sample_size = ceiling(@rows * @sample_percent / 100)

select *
from ..
join (
    -- Not a precise value: if limiting exactly at, can introduce more bias.
    -- Using 'option optimize for' avoids this while requiring dynamic SQL.
    select top (@sample_size + convert(int, @sample_percent + 5))
    from t
    where 1=1
        and t.Name = 'Mr. No Questionable Checksum Usages'
        and ( -- sample
            @sample_percent = 100
            or abs(
                convert(bigint, hashbytes('SHA1', convert(varbinary(32), t.rowguid)))
            ) % (1000 * 100) < (1000 * @sample_percent)
        )
) sampled
on ..

在MySQL中，你可以这样做:

SELECT `PRIMARY_KEY`, rand() FROM table ORDER BY rand() LIMIT 5000;

如果你(不像OP)需要特定数量的记录(这使得CHECKSUM方法很困难)，并且想要一个比TABLESAMPLE本身提供的更随机的样本，并且也想要比CHECKSUM更好的速度，你可以将TABLESAMPLE和NEWID()方法合并，如下所示:

DECLARE @sampleCount int = 50
SET STATISTICS TIME ON

SELECT TOP (@sampleCount) * 
FROM [yourtable] TABLESAMPLE(10 PERCENT)
ORDER BY NEWID()

SET STATISTICS TIME OFF

就我而言，这是随机性(我知道这并不是真的)和速度之间最直接的妥协。适当地改变TABLESAMPLE百分比(或行数)——百分比越高，样本的随机性越大，但速度会有线性下降。(注意，TABLESAMPLE不接受变量)

根据您的需要，TABLESAMPLE将为您提供几乎相同的随机和更好的性能。 这在MS SQL server 2005及更高版本上可用。

TABLESAMPLE将从随机页而不是随机行返回数据，因此它甚至不会检索它不会返回的数据。

我在一个非常大的桌子上测试

select top 1 percent * from [tablename] order by newid()

花了20多分钟。

select * from [tablename] tablesample(1 percent)

花了2分钟。

在TABLESAMPLE中，较小的样本也会提高性能，而使用newd()则不会。

请记住，这并不像newid()方法那样随机，但会给您一个不错的抽样。

请参见MSDN页面。

Newid ()/order by可以工作，但对于大型结果集来说代价非常高，因为它必须为每一行生成一个id，然后对它们进行排序。

从性能的角度来看，TABLESAMPLE()很好，但是您将得到结果的聚集(将返回页面上的所有行)。

为了获得更好的真实随机样本，最好的方法是随机过滤掉行。我在SQL Server Books Online文章使用TABLESAMPLE限制结果集中找到了以下代码示例:

If you really want a random sample of individual rows, modify your query to filter out rows randomly, instead of using TABLESAMPLE. For example, the following query uses the NEWID function to return approximately one percent of the rows of the Sales.SalesOrderDetail table: SELECT * FROM Sales.SalesOrderDetail WHERE 0.01 >= CAST(CHECKSUM(NEWID(),SalesOrderID) & 0x7fffffff AS float) / CAST (0x7fffffff AS int) The SalesOrderID column is included in the CHECKSUM expression so that NEWID() evaluates once per row to achieve sampling on a per-row basis. The expression CAST(CHECKSUM(NEWID(), SalesOrderID) & 0x7fffffff AS float / CAST (0x7fffffff AS int) evaluates to a random float value between 0 and 1.

当对一个有1,000,000行的表运行时，下面是我的结果:

SET STATISTICS TIME ON
SET STATISTICS IO ON

/* newid()
   rows returned: 10000
   logical reads: 3359
   CPU time: 3312 ms
   elapsed time = 3359 ms
*/
SELECT TOP 1 PERCENT Number
FROM Numbers
ORDER BY newid()

/* TABLESAMPLE
   rows returned: 9269 (varies)
   logical reads: 32
   CPU time: 0 ms
   elapsed time: 5 ms
*/
SELECT Number
FROM Numbers
TABLESAMPLE (1 PERCENT)

/* Filter
   rows returned: 9994 (varies)
   logical reads: 3359
   CPU time: 641 ms
   elapsed time: 627 ms
*/    
SELECT Number
FROM Numbers
WHERE 0.01 >= CAST(CHECKSUM(NEWID(), Number) & 0x7fffffff AS float) 
              / CAST (0x7fffffff AS int)

SET STATISTICS IO OFF
SET STATISTICS TIME OFF

如果您可以使用TABLESAMPLE，它将为您提供最佳性能。否则使用newwid ()/filter方法。如果结果集很大，Newid ()/order by应该是最后的选择。

我还没看出来答案有什么不同。我有一个额外的约束条件，给定一个初始种子，每次都要选择相同的行集。

对于MS SQL:

最小的例子:

select top 10 percent *
from table_name
order by rand(checksum(*))

规范化执行时间:1.00

NewId()例子:

select top 10 percent *
from table_name
order by newid()

规范化执行时间:1.02

NewId()比rand(checksum(*))慢不了多少，所以您可能不希望对大型记录集使用它。

初始种子选择:

declare @seed int
set @seed = Year(getdate()) * month(getdate()) /* any other initial seed here */

select top 10 percent *
from table_name
order by rand(checksum(*) % @seed) /* any other math function here */

如果给定一个种子，你需要选择相同的集合，这似乎是可行的。