精妙的SQL语句
人气:0精妙的SQL语句
- 说明:复制表(只复制结构,源表名:a 新表名:b)
select * into b from a where 1<>1
- 说明:拷贝表(拷贝数据,源表名:a 目标表名:b)
insert into b(a, b, c) select d,e,f from b;
- 说明:显示文章、提交人和最后回复时间
select a.title,a.username,b.adddate from table a,(select max(adddate) adddate from table where table.title=a.title) b
- 说明:外连接查询(表名1:a 表名2:b)
select a.a, a.b, a.c, b.c, b.d, b.f from a LEFT OUT JOIN b ON a.a = b.c
- 说明:日程安排提前五分钟提醒
select * from 日程安排 where datediff('minute',f开始时间,getdate())>5
- 说明:两张关联表,删除主表中已经在副表中没有的信息
delete from info where not exists ( select * from infobz where info.infid=infobz.infid )
- 说明:--
SQL:
SELECT A.NUM, A.NAME, B.UPD_DATE, B.PREV_UPD_DATE
FROM TABLE1,
(SELECT X.NUM, X.UPD_DATE, Y.UPD_DATE PREV_UPD_DATE
FROM (SELECT NUM, UPD_DATE, INBOUND_QTY, STOCK_ONHAND
FROM TABLE2
WHERE TO_CHAR(UPD_DATE,'YYYY/MM') = TO_CHAR(SYSDATE, 'YYYY/MM')) X,
(SELECT NUM, UPD_DATE, STOCK_ONHAND
FROM TABLE2
WHERE TO_CHAR(UPD_DATE,'YYYY/MM') =
TO_CHAR(TO_DATE(TO_CHAR(SYSDATE, 'YYYY/MM') || '/01','YYYY/MM/DD') - 1, 'YYYY/MM') ) Y,
WHERE X.NUM = Y.NUM (+)
AND X.INBOUND_QTY + NVL(Y.STOCK_ONHAND,0) <> X.STOCK_ONHAND ) B
WHERE A.NUM = B.NUM
- 说明:--
select * from studentinfo where not exists(select * from student where studentinfo.id=student.id) and 系名称='"&strdepartmentname&"' and 专业名称='"&strprofessionname&"' order by 性别,生源地,高考总成绩
- 从数据库中去一年的各单位电话费统计(电话费定额贺电化肥清单两个表来源)
SELECT a.userper, a.tel, a.standfee, TO_CHAR(a.telfeedate, 'yyyy') AS telyear,
SUM(decode(TO_CHAR(a.telfeedate, 'mm'), '01', a.factration)) AS JAN,
SUM(decode(TO_CHAR(a.telfeedate, 'mm'), '02', a.factration)) AS FRI,
SUM(decode(TO_CHAR(a.telfeedate, 'mm'), '03', a.factration)) AS MAR,
SUM(decode(TO_CHAR(a.telfeedate, 'mm'), '04', a.factration)) AS APR,
SUM(decode(TO_CHAR(a.telfeedate, 'mm'), '05', a.factration)) AS MAY,
SUM(decode(TO_CHAR(a.telfeedate, 'mm'), '06', a.factration)) AS JUE,
SUM(decode(TO_CHAR(a.telfeedate, 'mm'), '07', a.factration)) AS JUL,
SUM(decode(TO_CHAR(a.telfeedate, 'mm'), '08', a.factration)) AS AGU,
SUM(decode(TO_CHAR(a.telfeedate, 'mm'), '09', a.factration)) AS SEP,
SUM(decode(TO_CHAR(a.telfeedate, 'mm'), '10', a.factration)) AS OCT,
SUM(decode(TO_CHAR(a.telfeedate, 'mm'), '11', a.factration)) AS NOV,
SUM(decode(TO_CHAR(a.telfeedate, 'mm'), '12', a.factration)) AS DEC
FROM (SELECT a.userper, a.tel, a.standfee, b.telfeedate, b.factration
FROM TELFEESTAND a, TELFEE b
WHERE a.tel = b.telfax) a
GROUP BY a.userper, a.tel, a.standfee, TO_CHAR(a.telfeedate, 'yyyy')
- 说明:四表联查问题
select * from a left inner join b on a.a=b.b right inner join c on a.a=c.c inner join d on a.a=d.d where .....
- 说明:得到表中最小的未使用的ID号
- SELECT (CASE WHEN EXISTS(SELECT * FROM Handle b WHERE b.HandleID = 1) THEN MIN(HandleID) + 1 ELSE 1 END) as HandleID FROM Handle WHERE NOT HandleID IN (SELECT a.HandleID - 1 FROM Handle a)
- 一个SQL语句的问题:行列转换
select * from v_temp
上面的视图结果如下:
user_name role_name
-------------------------
系统管理员 管理员
feng 管理员
feng 一般用户
test 一般用户
想把结果变成这样:
user_name role_name
---------------------------
系统管理员 管理员
feng 管理员,一般用户
test 一般用户
===================
create table a_test(name varchar(20),role2 varchar(20))
insert into a_test values('李','管理员')
insert into a_test values('张','管理员')
insert into a_test values('张','一般用户')
insert into a_test values('常','一般用户')
create function join_str(@content varchar(100))
returns varchar(2000)
as
begin
declare @str varchar(2000)
set @str=''
select @str=@str+','+rtrim(role2) from a_test where [name]=@content
select @str=right(@str,len(@str)-1)
return @str
end
go
--调用:
select [name],dbo.join_str([name]) role2 from a_test group by [name]
--select distinct name,dbo.uf_test(name) from a_test
- 快速比较结构相同的两表
结构相同的两表,一表有记录3万条左右,一表有记录2万条左右,我怎样快速查找两表的不同记录?
============================
给你一个测试方法,从northwind中的orders表取数据。
select * into n1 from orders
select * into n2 from orders
select * from n1
select * from n2
--添加主键,然后修改n1中若干字段的若干条
alter table n1 add constraint pk_n1_id primary key (OrderID)
alter table n2 add constraint pk_n2_id primary key (OrderID)
select OrderID from (select * from n1 union select * from n2) a group by OrderID having count(*) > 1
应该可以,而且将不同的记录的ID显示出来。
下面的适用于双方记录一样的情况,
select * from n1 where orderid in (select OrderID from (select * from n1 union select * from n2) a group by OrderID having count(*) > 1)
至于双方互不存在的记录是比较好处理的
--删除n1,n2中若干条记录
delete from n1 where orderID in ('10728','10730')
delete from n2 where orderID in ('11000','11001')
--*************************************************************
-- 双方都有该记录却不完全相同
select * from n1 where orderid in(select OrderID from (select * from n1 union select * from n2) a group by OrderID having count(*) > 1)
union
--n2中存在但在n1中不存的在10728,10730
select * from n1 where OrderID not in (select OrderID from n2)
union
--n1中存在但在n2中不存的在11000,11001
select * from n2 where OrderID not in (select OrderID from n1)
- 四种方法取表里n到m条纪录:
1.
select top m * into 临时表(或表变量) from tablename order by columnname -- 将top m笔插入
set rowcount n
select * from 表变量 order by columnname desc
2.
select top n * from (select top m * from tablename order by columnname) a order by columnname desc
3.如果tablename里没有其他identity列,那么:
select identity(int) id0,* into #temp from tablename
取n到m条的语句为:
select * from #temp where id0 >=n and id0 <= m
如果你在执行select identity(int) id0,* into #temp from tablename这条语句的时候报错,那是因为你的DB中间的select into/bulkcopy属性没有打开要先执行:
exec sp_dboption 你的DB名字,'select into/bulkcopy',true
4.如果表里有identity属性,那么简单:
select * from tablename where identitycol between n and m
- 如何删除一个表中重复的记录?
create table a_dist(id int,name varchar(20))
insert into a_dist values(1,'abc')
insert into a_dist values(1,'abc')
insert into a_dist values(1,'abc')
insert into a_dist values(1,'abc')
exec up_distinct 'a_dist','id'
select * from a_dist
create procedure up_distinct(@t_name varchar(30),@f_key varchar(30))
--f_key表示是分组字段﹐即主键字段
as
begin
declare @max integer,@id varchar(30) ,@sql varchar(7999) ,@type integer
select @sql = 'declare cur_rows cursor for select '+@f_key+' ,count(*) from ' +@t_name +' group by ' +@f_key +' having count(*) > 1'
exec(@sql)
open cur_rows
fetch cur_rows into @id,@max
while @@fetch_status=0
begin
select @max = @max -1
set rowcount @max
select @type = xtype from syscolumns where id=object_id(@t_name) and name=@f_key
if @type=56
select @sql = 'delete from '+@t_name+' where ' + @f_key+' = '+ @id
if @type=167
select @sql = 'delete from '+@t_name+' where ' + @f_key+' = '+''''+ @id +''''
exec(@sql)
fetch cur_rows into @id,@max
end
close cur_rows
deallocate cur_rows
set rowcount 0
end
select * from systypes
select * from syscolumns where id = object_id('a_dist')
- 查询数据的最大排序问题(只能用一条语句写)
CREATE TABLE hard (qu char (11) ,co char (11) ,je numeric(3, 0))
insert into hard values ('A','1',3)
insert into hard values ('A','2',4)
insert into hard values ('A','4',2)
insert into hard values ('A','6',9)
insert into hard values ('B','1',4)
insert into hard values ('B','2',5)
insert into hard values ('B','3',6)
insert into hard values ('C','3',4)
insert into hard values ('C','6',7)
insert into hard values ('C','2',3)
要求查询出来的结果如下:
qu co je
----------- ----------- -----
A 6 9
A 2 4
B 3 6
B 2 5
C 6 7
C 3 4
就是要按qu分组,每组中取je最大的前2位!!
而且只能用一句sql语句!!!
select * from hard a where je in (select top 2 je from hard b where a.qu=b.qu order by je)
- 求删除重复记录的sql语句?
怎样把具有相同字段的纪录删除,只留下一条。
例如,表test里有id,name字段
如果有name相同的记录 只留下一条,其余的删除。
name的内容不定,相同的记录数不定。
有没有这样的sql语句?
==============================
A:一个完整的解决方案:
将重复的记录记入temp1表:
select [标志字段id],count(*) into temp1 from [表名]
group by [标志字段id]
having count(*)>1
2、将不重复的记录记入temp1表:
insert temp1 select [标志字段id],count(*) from [表名] group by [标志字段id] having count(*)=1
3、作一个包含所有不重复记录的表:
select * into temp2 from [表名] where 标志字段id in(select 标志字段id from temp1)
4、删除重复表:
delete [表名]
5、恢复表:
insert [表名] select * from temp2
6、删除临时表:
drop table temp1
drop table temp2
================================
B:
create table a_dist(id int,name varchar(20))
insert into a_dist values(1,'abc')
insert into a_dist values(1,'abc')
insert into a_dist values(1,'abc')
insert into a_dist values(1,'abc')
exec up_distinct 'a_dist','id'
select * from a_dist
create procedure up_distinct(@t_name varchar(30),@f_key varchar(30))
--f_key表示是分组字段﹐即主键字段
as
begin
declare @max integer,@id varchar(30) ,@sql varchar(7999) ,@type integer
select @sql = 'declare cur_rows cursor for select '+@f_key+' ,count(*) from ' +@t_name +' group by ' +@f_key +' having count(*) > 1'
exec(@sql)
open cur_rows
fetch cur_rows into @id,@max
while @@fetch_status=0
begin
select @max = @max -1
set rowcount @max
select @type = xtype from syscolumns where id=object_id(@t_name) and name=@f_key
if @type=56
select @sql = 'delete from '+@t_name+' where ' + @f_key+' = '+ @id
if @type=167
select @sql = 'delete from '+@t_name+' where ' + @f_key+' = '+''''+ @id +''''
exec(@sql)
fetch cur_rows into @id,@max
end
close cur_rows
deallocate cur_rows
set rowcount 0
end
select * from systypes
select * from syscolumns where id = object_id('a_dist')
- 行列转换--普通
假设有张学生成绩表(CJ)如下
Name Subject Result
张三 语文 80
张三 数学 90
张三 物理 85
李四 语文 85
李四 数学 92
李四 物理 82
想变成
姓名 语文 数学 物理
张三 80 90 85
李四 85 92 82
declare @sql varchar(4000)
set @sql = 'select Name'
select @sql = @sql + ',sum(case Subject when '''+Subject+''' then Result end) ['+Subject+']'
from (select distinct Subject from CJ) as a
select @sql = @sql+' from test group by name'
exec(@sql)
行列转换--合并
有表A,
id pid
1 1
1 2
1 3
2 1
2 2
3 1
如何化成表B:
id pid
1 1,2,3
2 1,2
3 1
创建一个合并的函数
create function fmerg(@id int)
returns varchar(8000)
as
begin
declare @str varchar(8000)
set @str=''
select @str=@str+','+cast(pid as varchar) from 表A where id=@id
set @str=right(@str,len(@str)-1)
return(@str)
End
go
--调用自定义函数得到结果
select distinct id,dbo.fmerg(id) from 表A
- 如何取得一个数据表的所有列名
方法如下:先从SYSTEMOBJECT系统表中取得数据表的SYSTEMID,然后再SYSCOLUMN表中取得该数据表的所有列名。
SQL语句如下:
declare @objid int,@objname char(40)
set @objname = 'tablename'
select @objid = id from sysobjects where id = object_id(@objname)
select 'Column_name' = name from syscolumns where id = @objid order by colid
或
SELECT * FROM INFORMATION_SCHEMA.COLUMNS WHERE TABLE_NAME ='users'
- 通过SQL语句来更改用户的密码
修改别人的,需要sysadmin role
EXEC sp_password NULL, 'newpassword', 'User'
如果帐号为SA执行EXEC sp_password NULL, 'newpassword', sa
- 怎么判断出一个表的哪些字段不允许为空?
select COLUMN_NAME from INFORMATION_SCHEMA.COLUMNS where IS_NULLABLE='NO' and TABLE_NAME=tablename
- 如何在数据库里找到含有相同字段的表?
a. 查已知列名的情况
SELECT b.name as TableName,a.name as columnname
From syscolumns a INNER JOIN sysobjects b
ON a.id=b.id
AND b.type='U'
AND a.name='你的字段名字'
- 未知列名查所有在不同表出现过的列名
Select o.name As tablename,s1.name As columnname
From syscolumns s1, sysobjects o
Where s1.id = o.id
And o.type = 'U'
And Exists (
Select 1 From syscolumns s2
Where s1.name = s2.name
And s1.id <> s2.id
)
- 查询第xxx行数据
假设id是主键:
select * from (select top xxx * from yourtable) aa where not exists(select 1 from (select top xxx-1 * from yourtable) bb where aa.id=bb.id)
如果使用游标也是可以的
fetch absolute [number] from [cursor_name]
行数为绝对行数
- SQL Server日期计算
a. 一个月的第一天
SELECT DATEADD(mm, DATEDIFF(mm,0,getdate()), 0)
b. 本周的星期一
SELECT DATEADD(wk, DATEDIFF(wk,0,getdate()), 0)
c. 一年的第一天
SELECT DATEADD(yy, DATEDIFF(yy,0,getdate()), 0)
d. 季度的第一天
SELECT DATEADD(qq, DATEDIFF(qq,0,getdate()), 0)
e. 上个月的最后一天
SELECT dateadd(ms,-3,DATEADD(mm, DATEDIFF(mm,0,getdate()), 0))
f. 去年的最后一天
SELECT dateadd(ms,-3,DATEADD(yy, DATEDIFF(yy,0,getdate()), 0))
g. 本月的最后一天
SELECT dateadd(ms,-3,DATEADD(mm, DATEDIFF(m,0,getdate())+1, 0))
h. 本月的第一个星期一
select DATEADD(wk, DATEDIFF(wk,0,
dateadd(dd,6-datepart(day,getdate()),getdate())
), 0)
i. 本年的最后一天
SELECT dateadd(ms,-3,DATEADD(yy, DATEDIFF(yy,0,getdate())+1, 0))。
- 获取表结构[把 'sysobjects' 替换 成 'tablename' 即可]
SELECT CASE IsNull(I.name, '')
When '' Then ''
Else '*'
End as IsPK,
Object_Name(A.id) as t_name,
A.name as c_name,
IsNull(SubString(M.text, 1, 254), '') as pbc_init,
T.name as F_DataType,
CASE IsNull(TYPEPROPERTY(T.name, 'Scale'), '')
WHEN '' Then Cast(A.prec as varchar)
ELSE Cast(A.prec as varchar) + ',' + Cast(A.scale as varchar)
END as F_Scale,
A.isnullable as F_isNullAble
FROM Syscolumns as A
JOIN Systypes as T
ON (A.xType = T.xUserType AND A.Id = Object_id('sysobjects') )
LEFT JOIN ( SysIndexes as I
JOIN Syscolumns as A1
ON ( I.id = A1.id and A1.id = object_id('sysobjects') and (I.status & 0x800) = 0x800 AND A1.colid <= I.keycnt) )
ON ( A.id = I.id AND A.name = index_col('sysobjects', I.indid, A1.colid) )
LEFT JOIN SysComments as M
ON ( M.id = A.cdefault and ObjectProperty(A.cdefault, 'IsConstraint') = 1 )
ORDER BY A.Colid ASC
- 提取数据库内所有表的字段详细说明的SQL语句
SELECT
(case when a.colorder=1 then d.name else '' end) N'表名',
a.colorder N'字段序号',
a.name N'字段名',
(case when COLUMNPROPERTY( a.id,a.name,'IsIdentity')=1 then '√'else ''
end) N'标识',
(case when (SELECT count(*)
FROM sysobjects
WHERE (name in
(SELECT name
FROM sysindexes
WHERE (id = a.id) AND (indid in
(SELECT indid
FROM sysindexkeys
WHERE (id = a.id) AND (colid in
(SELECT colid
FROM syscolumns
WHERE (id = a.id) AND (name = a.name))))))) AND
(xtype = 'PK'))>0 then '√' else '' end) N'主键',
b.name N'类型',
a.length N'占用字节数',
COLUMNPROPERTY(a.id,a.name,'PRECISION') as N'长度',
isnull(COLUMNPROPERTY(a.id,a.name,'Scale'),0) as N'小数位数',
(case when a.isnullable=1 then '√'else '' end) N'允许空',
isnull(e.text,'') N'默认值',
isnull(g.[value],'') AS N'字段说明'
FROM syscolumns a
left join systypes b
on a.xtype=b.xusertype
inner join sysobjects d
on a.id=d.id and d.xtype='U' and d.name<>'dtproperties'
left join syscomments e
on a.cdefault=e.id
left join sysproperties g
on a.id=g.id AND a.colid = g.smallid
order by object_name(a.id),a.colorder
- 快速获取表test的记录总数[对大容量表非常有效]
快速获取表test的记录总数:
select rows from sysindexes where id = object_id('test') and indid in (0,1)
update 2 set KHXH=(ID+1)\2 2行递增编号
update [23] set id1 = 'No.'+right('00000000'+id,6) where id not like 'No%' //递增
update [23] set id1= 'No.'+right('00000000'+replace(id1,'No.',''),6) //补位递增
delete from [1] where (id%2)=1
奇数
- 替换表名字段
update [1] set domurl = replace(domurl,'Upload/Imgswf/','Upload/Photo/') where domurl like '%Upload/Imgswf/%'
- 截位
SELECT LEFT(表名, 5)
加载全部内容