数据库系统原理与实践——完成第四次作业
This commit is contained in:
parent
4a4d9e0e21
commit
2a10a2adc5
513
数据库系统原理与实践/平时作业/第四次作业.tex
Normal file
513
数据库系统原理与实践/平时作业/第四次作业.tex
Normal file
@ -0,0 +1,513 @@
|
|||||||
|
\documentclass[全部作业]{subfiles}
|
||||||
|
\input{mysubpreamble}
|
||||||
|
|
||||||
|
\begin{document}
|
||||||
|
\setcounter{chapter}{3}
|
||||||
|
\chapter{第四次作业}
|
||||||
|
\begin{enumerate}
|
||||||
|
\item \choice[1]{关于视图的操作,下列哪条语句是正确的?
|
||||||
|
A
|
||||||
|
可以使用 UPDATE 语句更新视图中的任意数据
|
||||||
|
|
||||||
|
B
|
||||||
|
如果视图包含聚合函数,则该视图是不可(插入)更新的
|
||||||
|
|
||||||
|
C
|
||||||
|
删除基表不会影响视图
|
||||||
|
|
||||||
|
D
|
||||||
|
视图是只读的,不能对其执行 DELETE 操作
|
||||||
|
|
||||||
|
}{2}
|
||||||
|
\item \choice[1]{如果在 SQL 查询中使用右外连接(RIGHT OUTER JOIN),并且右表中某些记录没有与左表中的记录匹配,那么:
|
||||||
|
A
|
||||||
|
右表中没有匹配的记录将被过滤掉
|
||||||
|
|
||||||
|
B
|
||||||
|
右表中的记录将显示 NULL
|
||||||
|
|
||||||
|
C
|
||||||
|
右表中的记录将完整保留,左表没有匹配的记录会显示 NULL
|
||||||
|
|
||||||
|
D
|
||||||
|
查询将无法执行
|
||||||
|
|
||||||
|
}{3}
|
||||||
|
\item \choice[1]{给定以下 SQL 语句:
|
||||||
|
|
||||||
|
CREATE VIEW employee_view AS
|
||||||
|
|
||||||
|
SELECT id, name, department
|
||||||
|
|
||||||
|
FROM employees
|
||||||
|
|
||||||
|
WHERE department = 'Sales';
|
||||||
|
|
||||||
|
该语句的作用是?
|
||||||
|
|
||||||
|
|
||||||
|
|
||||||
|
A
|
||||||
|
创建一个名为 employee_view 的物理表
|
||||||
|
|
||||||
|
B
|
||||||
|
创建一个包含所有员工信息的物理表
|
||||||
|
|
||||||
|
C
|
||||||
|
创建一个只包含销售部门员工的虚拟表
|
||||||
|
|
||||||
|
D
|
||||||
|
修改现有的员工表以只包含销售部门员工}{3}
|
||||||
|
\item \choice[1]{以下哪个 SQL 语句可以删除视图?
|
||||||
|
A
|
||||||
|
DELETE VIEW view_name;
|
||||||
|
|
||||||
|
B
|
||||||
|
DROP VIEW view_name;
|
||||||
|
|
||||||
|
C
|
||||||
|
REMOVE VIEW view_name;
|
||||||
|
|
||||||
|
D
|
||||||
|
ERASE VIEW view_name;
|
||||||
|
|
||||||
|
}{2}
|
||||||
|
\item \choice[1]{假设有两张表,销售记录表 sales 和员工表 employees,要求查询出所有员工的姓名以及他们的销售记录,如果某个员工没有销售记录,销售记录应显示为 NULL。以下哪个 SQL 语句是正确的?
|
||||||
|
|
||||||
|
A
|
||||||
|
\mint{SQL}|SELECT employees.employee_name, sales.sale_amount FROM employees LEFT OUTER JOIN sales ON employees.employee_id = sales.employee_id;|
|
||||||
|
|
||||||
|
B
|
||||||
|
\mint{SQL}|SELECT employees.employee_name, sales.sale_amount FROM employees RIGHT OUTER JOIN sales ON employees.employee_id = sales.employee_id;|
|
||||||
|
|
||||||
|
C
|
||||||
|
\mint{SQL}|SELECT employees.employee_name, sales.sale_amount FROM employees FULL OUTER JOIN sales ON employees.employee_id = sales.employee_id;|
|
||||||
|
|
||||||
|
D
|
||||||
|
\mint{SQL}|SELECT employees.employee_name, sales.sale_amount FROM employees INNER JOIN sales ON employees.employee_id = sales.employee_id;|
|
||||||
|
}{1}
|
||||||
|
\questionandanswer[]{
|
||||||
|
上机题1:查询个人客户(individual)中“吕东”所有办理的业务信息。(输出:身份证号、姓名全名、客户编号、账户编号、开户行名称、可用余额、产品名称、产品类型名称);
|
||||||
|
|
||||||
|
注意:客户全名可用CONCAT关键字,CONCAT(str1,str2,…)。
|
||||||
|
}{}
|
||||||
|
{
|
||||||
|
\begin{minted}{SQL}
|
||||||
|
select ID_NUMBER 身份证号,
|
||||||
|
concat(LAST_NAME, FIRST_NAME) 姓名全名,
|
||||||
|
CUST_ID 客户编号,
|
||||||
|
ACCOUNT_ID 账户编号,
|
||||||
|
branch.NAME 开户行名称,
|
||||||
|
account.AVAIL_BALANCE 可用余额,
|
||||||
|
product.NAME 产品名称,
|
||||||
|
product_type.NAME 产品类型名称
|
||||||
|
from individual
|
||||||
|
left join account using (CUST_ID)
|
||||||
|
left join branch on account.OPEN_BRANCH_ID = branch.BRANCH_ID
|
||||||
|
left join product using (PRODUCT_CD)
|
||||||
|
left join product_type using (PRODUCT_TYPE_CD)
|
||||||
|
where concat(LAST_NAME, FIRST_NAME) = "吕东";
|
||||||
|
\end{minted}
|
||||||
|
\scriptsize\begin{csv}
|
||||||
|
,身份证号,姓名全名,客户编号,账户编号,开户行名称,可用余额,产品名称,产品类型名称
|
||||||
|
1,450881196612220768,吕东,4,10,上海市总行,59934.1200,个人活期储蓄账户,存款
|
||||||
|
2,450881196612220768,吕东,4,11,上海市总行,650600.0000,个人定期存款账户,存款
|
||||||
|
3,450881196612220768,吕东,4,12,上海市总行,67800.0000,个人通知存款账户,存款
|
||||||
|
\end{csv}
|
||||||
|
}
|
||||||
|
\questionandanswer[]{
|
||||||
|
上机题2:创建视图,查询客户(customer)表,列出所有客户的姓名全名、及其在单位的职位名称(title)(提示:对私客户职位为NULL)。(输出:全名,职位名称)。
|
||||||
|
|
||||||
|
注意:可使用CASE关键字处理对公对私客户,CASE [column_name] WHEN [value1] THEN [result1]... ELSE [default] END。
|
||||||
|
}{}
|
||||||
|
{
|
||||||
|
\begin{minted}{SQL}
|
||||||
|
create view 视图1 as
|
||||||
|
select (case CUST_TYPE_CD
|
||||||
|
when 'B' then concat(officer.LAST_NAME, officer.FIRST_NAME)
|
||||||
|
else concat(individual.LAST_NAME, individual.FIRST_NAME) end) as 姓名全名,
|
||||||
|
(case CUST_TYPE_CD when 'B' then TITLE else NULL end) as 职位名称
|
||||||
|
from customer
|
||||||
|
left join individual using (CUST_ID)
|
||||||
|
left join officer using (CUST_ID);
|
||||||
|
\end{minted}
|
||||||
|
\begin{csv}
|
||||||
|
,姓名全名,职位名称
|
||||||
|
1,尤青,
|
||||||
|
2,许文强,
|
||||||
|
3,何婕,
|
||||||
|
4,吕东,
|
||||||
|
5,施珊珊,
|
||||||
|
6,张晓,
|
||||||
|
7,孔庆东,
|
||||||
|
8,曹方,
|
||||||
|
9,严匡,
|
||||||
|
10,华哨,校长
|
||||||
|
11,金歌,董事长
|
||||||
|
12,魏俊杰,董事长
|
||||||
|
13,陶海桥,董事长
|
||||||
|
\end{csv}
|
||||||
|
}
|
||||||
|
\questionandanswer[]{
|
||||||
|
上机题3:查询各个银行中余额最多的账户的客户信息。(输出:银行名称、姓名全名、账户余额)
|
||||||
|
|
||||||
|
在account表中插入一条测试元组:
|
||||||
|
|
||||||
|
insert into account values (30, 2000000, null, null, '2022-10-12', '正常', 5, 2, 15, 'CS');
|
||||||
|
|
||||||
|
提示:需要处理银行中余额最多的不止一个客户的情况。
|
||||||
|
}{}
|
||||||
|
{
|
||||||
|
% 怎么感觉这么复杂,是不是哪里弄错了
|
||||||
|
\begin{minted}{SQL}
|
||||||
|
insert into account
|
||||||
|
values (30, 2000000, null, null, '2022-10-12', '正常', 5, 2, 15, 'CS');
|
||||||
|
select branch.NAME 银行名称,
|
||||||
|
(case CUST_TYPE_CD
|
||||||
|
when 'B' then concat(officer.LAST_NAME, officer.FIRST_NAME)
|
||||||
|
else concat(individual.LAST_NAME, individual.FIRST_NAME) end) 姓名全名,
|
||||||
|
account.AVAIL_BALANCE 账户余额
|
||||||
|
from (select account.ACCOUNT_ID, OPEN_BRANCH_ID, AVAIL_BALANCE, max_avail_balance
|
||||||
|
from account
|
||||||
|
join
|
||||||
|
(select OPEN_BRANCH_ID, max(AVAIL_BALANCE) as max_avail_balance
|
||||||
|
from account
|
||||||
|
group by OPEN_BRANCH_ID) as b using (OPEN_BRANCH_ID)
|
||||||
|
where AVAIL_BALANCE = max_avail_balance) as abAI
|
||||||
|
left join branch on BRANCH_ID = OPEN_BRANCH_ID
|
||||||
|
left join account using (ACCOUNT_ID)
|
||||||
|
left join individual using (CUST_ID)
|
||||||
|
left join officer using (CUST_ID)
|
||||||
|
left join customer using (CUST_ID);
|
||||||
|
\end{minted}
|
||||||
|
\begin{csv}
|
||||||
|
,银行名称,姓名全名,账户余额
|
||||||
|
1,上海市总行,张晓,3330000.0000
|
||||||
|
2,建国支行,许文强,2000000.0000
|
||||||
|
3,建国支行,施珊珊,2000000.0000
|
||||||
|
4,南京分行,何婕,650000.0000
|
||||||
|
5,杭州分行,施珊珊,340023.0000
|
||||||
|
\end{csv}
|
||||||
|
}
|
||||||
|
\questionandanswer[]{
|
||||||
|
上机题4:查询所有同时办理了产品编号为“CS”和“RS”的个人客户(individual表)的身份证号(ID_NUMBER)和姓名全名。(输出:ID_NUMBER,全名)。
|
||||||
|
|
||||||
|
要求:使用NOT EXISTS进行包含关系查询。
|
||||||
|
}{}
|
||||||
|
{
|
||||||
|
\begin{minted}{SQL}
|
||||||
|
select individual.ID_NUMBER 身份证号, concat(individual.LAST_NAME, individual.FIRST_NAME) 姓名全名
|
||||||
|
from individual
|
||||||
|
where not exists(select *
|
||||||
|
from account
|
||||||
|
where PRODUCT_CD != 'CS' and PRODUCT_CD != 'RS' and account.CUST_ID = individual.CUST_ID);
|
||||||
|
\end{minted}
|
||||||
|
\begin{csv}
|
||||||
|
,身份证号,姓名全名
|
||||||
|
1,41032619780815564X,许文强
|
||||||
|
2,370900196802069281,何婕
|
||||||
|
3,320623197108259227,施珊珊
|
||||||
|
4,321111197703197600,孔庆东
|
||||||
|
\end{csv}
|
||||||
|
}
|
||||||
|
|
||||||
|
\questionandanswer[]{
|
||||||
|
上机题5:查询所有不在2013年(例如在2011、2019年等)有交易记录的账户编号(ACCOUNT_ID)。(输出:账户编号)
|
||||||
|
|
||||||
|
要求:需要用外连接来写该查询语句;
|
||||||
|
|
||||||
|
注意:某些账户可能在2011、2013年都有交易记录,但只要在2013年有一条交易记录的账户都会被排除在结果集之外。
|
||||||
|
}{}
|
||||||
|
{\kaishu
|
||||||
|
这题表述不清啊,是指“在2011、2019年等有交易记录”呢?还是“在2013年没有交易记录”呢?还是两个条件都要?
|
||||||
|
|
||||||
|
那么暂且认为是“在2013年没有交易记录”吧,也就是这样:
|
||||||
|
|
||||||
|
\begin{csv}
|
||||||
|
,2013年有交易记录,2013年无交易记录
|
||||||
|
其他年有交易记录,不要,要
|
||||||
|
其他年无交易记录,不要,要
|
||||||
|
\end{csv}
|
||||||
|
\begin{minted}{SQL}
|
||||||
|
select ACCOUNT_ID
|
||||||
|
from account
|
||||||
|
left join (select *
|
||||||
|
from acc_transaction
|
||||||
|
where year(TXN_DATE) != 2013) as not2013 using (ACCOUNT_ID)
|
||||||
|
where TXN_DATE is not null;
|
||||||
|
\end{minted}
|
||||||
|
\begin{csv}
|
||||||
|
,ACCOUNT_ID
|
||||||
|
1,4
|
||||||
|
2,5
|
||||||
|
3,18
|
||||||
|
4,19
|
||||||
|
5,21
|
||||||
|
6,28
|
||||||
|
7,10
|
||||||
|
8,17
|
||||||
|
9,13
|
||||||
|
10,29
|
||||||
|
11,27
|
||||||
|
12,3
|
||||||
|
13,23
|
||||||
|
14,12
|
||||||
|
15,22
|
||||||
|
16,15
|
||||||
|
17,28
|
||||||
|
\end{csv}
|
||||||
|
}
|
||||||
|
\questionandanswer[]{
|
||||||
|
教材3.11: 使用大学University数据库模式,请用SQL写出如下查询。
|
||||||
|
|
||||||
|
a.找出至少选修了一门Comp. Sci.课程的每名学生的ID和姓名,保证结果中没有重复的姓名。
|
||||||
|
|
||||||
|
b.找出没有选修2017年之前开设的任何课程的每名学生的ID和姓名。
|
||||||
|
|
||||||
|
c.找出每个系的教师的最高工资值。可以假设每个系至少有一位教师。
|
||||||
|
|
||||||
|
d.从前述查询所计算出的每个系的最高工资中选出所有系中的最低值。
|
||||||
|
}{}
|
||||||
|
{\kaishu
|
||||||
|
\begin{enumerate}
|
||||||
|
\item \begin{minted}{SQL}
|
||||||
|
select distinct ID, name
|
||||||
|
from course
|
||||||
|
left join takes using (course_id)
|
||||||
|
left join student using (ID)
|
||||||
|
left join department on course.dept_name = department.dept_name
|
||||||
|
where course.dept_name = "Comp. Sci.";
|
||||||
|
\end{minted}
|
||||||
|
\item \begin{minted}{SQL}
|
||||||
|
select ID, name
|
||||||
|
from student
|
||||||
|
where not exists(select *
|
||||||
|
from takes
|
||||||
|
where takes.ID = student.ID and year < 2017);
|
||||||
|
\end{minted}
|
||||||
|
\item \begin{minted}{SQL}
|
||||||
|
select dept_name, max(salary)
|
||||||
|
from instructor
|
||||||
|
group by dept_name;
|
||||||
|
\end{minted}
|
||||||
|
\item \begin{minted}{SQL}
|
||||||
|
with a as (select dept_name, max(salary) max_salary
|
||||||
|
from instructor
|
||||||
|
group by dept_name)
|
||||||
|
select *
|
||||||
|
from a
|
||||||
|
where max_salary = (select min(max_salary)
|
||||||
|
from a);
|
||||||
|
\end{minted}
|
||||||
|
\end{enumerate}
|
||||||
|
}
|
||||||
|
\begin{figure}
|
||||||
|
\centering
|
||||||
|
\includegraphics[width=1\linewidth]{imgs/2024-10-19-09-53-31.png}
|
||||||
|
\caption{大学数据库的模式图}
|
||||||
|
\end{figure}
|
||||||
|
\begin{verification}
|
||||||
|
教材的前言中介绍了官网:\url{https://db-book.com/},官网上可以下载示例数据:\url{https://db-book.com/university-lab-dir/sample_tables-dir/index.html},于是可以验证了。
|
||||||
|
\begin{minted}{SQL}
|
||||||
|
select distinct ID, name
|
||||||
|
from course
|
||||||
|
left join takes using (course_id)
|
||||||
|
left join student using (ID)
|
||||||
|
left join department on course.dept_name = department.dept_name
|
||||||
|
where course.dept_name = "Comp. Sci.";
|
||||||
|
|
||||||
|
select ID, name
|
||||||
|
from student
|
||||||
|
where not exists(select *
|
||||||
|
from takes
|
||||||
|
where takes.ID = student.ID
|
||||||
|
and year < 2017);
|
||||||
|
|
||||||
|
select dept_name, max(salary)
|
||||||
|
from instructor
|
||||||
|
group by dept_name;
|
||||||
|
|
||||||
|
with a as (select dept_name, max(salary) max_salary
|
||||||
|
from instructor
|
||||||
|
group by dept_name)
|
||||||
|
select *
|
||||||
|
from a
|
||||||
|
where max_salary = (select min(max_salary)
|
||||||
|
from a);
|
||||||
|
\end{minted}
|
||||||
|
|
||||||
|
\begin{csv}
|
||||||
|
,ID,name
|
||||||
|
1,00128,Zhang
|
||||||
|
2,12345,Shankar
|
||||||
|
3,45678,Levy
|
||||||
|
4,54321,Williams
|
||||||
|
5,76543,Brown
|
||||||
|
6,98765,Bourikas
|
||||||
|
\end{csv}
|
||||||
|
|
||||||
|
\begin{csv}
|
||||||
|
,ID,name
|
||||||
|
1,00128,Zhang
|
||||||
|
2,12345,Shankar
|
||||||
|
3,19991,Brandt
|
||||||
|
4,23121,Chavez
|
||||||
|
5,44553,Peltier
|
||||||
|
6,45678,Levy
|
||||||
|
7,54321,Williams
|
||||||
|
8,55739,Sanchez
|
||||||
|
9,70557,Snow
|
||||||
|
10,76543,Brown
|
||||||
|
11,76653,Aoi
|
||||||
|
12,98765,Bourikas
|
||||||
|
13,98988,Tanaka
|
||||||
|
\end{csv}
|
||||||
|
|
||||||
|
\begin{csv}
|
||||||
|
,dept_name,max(salary)
|
||||||
|
1,Biology,72000.00
|
||||||
|
2,Comp. Sci.,92000.00
|
||||||
|
3,Elec. Eng.,80000.00
|
||||||
|
4,Finance,90000.00
|
||||||
|
5,History,62000.00
|
||||||
|
6,Music,40000.00
|
||||||
|
7,Physics,95000.00
|
||||||
|
\end{csv}
|
||||||
|
|
||||||
|
\begin{csv}
|
||||||
|
,dept_name,max_salary
|
||||||
|
1,Music,40000.00
|
||||||
|
\end{csv}
|
||||||
|
\end{verification}
|
||||||
|
|
||||||
|
\questionandanswer[]{
|
||||||
|
教材3.16:考虑图 \ref{fig:3-19} 中的雇员数据库,其中主码被加了下划线。请给出下面每个查询的SQL表达式
|
||||||
|
|
||||||
|
a.找出每位这样的雇员的ID和姓名:该雇员所居住的城市与其工作的公司所在城市一样。
|
||||||
|
|
||||||
|
b.找出所居住的城市和街道与其经理相同的每位雇员的ID和姓名。
|
||||||
|
|
||||||
|
c.找出工资高于其所在公司所有雇员平均工资的每位雇员的ID和姓名。
|
||||||
|
|
||||||
|
d.找出工资总和最小的公司。
|
||||||
|
}{}
|
||||||
|
\begin{figure}
|
||||||
|
\centering
|
||||||
|
\includegraphics[width=0.5\linewidth]{imgs/2024-10-19-14-28-27.png}
|
||||||
|
\caption{雇员数据库}\label{fig:3-19}
|
||||||
|
\end{figure}
|
||||||
|
|
||||||
|
{\kaishu
|
||||||
|
\begin{enumerate}
|
||||||
|
\item \begin{minted}{SQL}
|
||||||
|
select ID, person_name
|
||||||
|
from employee
|
||||||
|
join works using (ID)
|
||||||
|
join company using (company_name)
|
||||||
|
where employee.city = company.city;
|
||||||
|
\end{minted}
|
||||||
|
|
||||||
|
\item \begin{minted}{SQL}
|
||||||
|
select subordinate.ID, subordinate.person_name
|
||||||
|
from manages
|
||||||
|
join employee as subordinate using (ID)
|
||||||
|
join employee as superior on manages.manager_id = superior.ID
|
||||||
|
where superior.street = subordinate.street
|
||||||
|
and superior.city = subordinate.city;
|
||||||
|
\end{minted}
|
||||||
|
|
||||||
|
\item \begin{minted}{SQL}
|
||||||
|
-- 这里假设employee.ID <--> works.ID 是一一对应的关系
|
||||||
|
select ID, person_name
|
||||||
|
from works
|
||||||
|
join (select company_name, avg(salary) as avg_salary
|
||||||
|
from works
|
||||||
|
group by company_name) as a using (company_name)
|
||||||
|
join employee using (ID)
|
||||||
|
where works.salary > a.avg_salary;
|
||||||
|
\end{minted}
|
||||||
|
|
||||||
|
\item \begin{minted}{SQL}
|
||||||
|
with a as (select company_name, sum(salary) as sum_salary
|
||||||
|
from company
|
||||||
|
join works using (company_name)
|
||||||
|
group by company_name)
|
||||||
|
select company_name
|
||||||
|
from a
|
||||||
|
where sum_salary = (select min(sum_salary) from a);
|
||||||
|
\end{minted}
|
||||||
|
\end{enumerate}
|
||||||
|
}
|
||||||
|
\questionandanswer[]{
|
||||||
|
教材3.17:考虑图 \ref{fig:3-19} 中的雇员数据库。请给出下面每个查询的SQL表达式。
|
||||||
|
|
||||||
|
a.为“First Bank Corporation"的所有雇员增长10\%的工资。
|
||||||
|
|
||||||
|
b.为“First Bank Corporation”的所有经理增长10\%的工资。
|
||||||
|
|
||||||
|
c.删除“Small Bank Corporation"的雇员在works关系中的所有元组。
|
||||||
|
}{}
|
||||||
|
{\kaishu
|
||||||
|
\begin{enumerate}
|
||||||
|
\item \begin{minted}{SQL}
|
||||||
|
update works
|
||||||
|
set salary = salary * 1.1
|
||||||
|
where company_name = "First Bank Corporation";
|
||||||
|
\end{minted}
|
||||||
|
\item \begin{minted}{SQL}
|
||||||
|
update works
|
||||||
|
join manages on works.ID = manages.manager_id
|
||||||
|
set salary = salary * 1.1
|
||||||
|
where company_name = "First Bank Corporation";
|
||||||
|
\end{minted}
|
||||||
|
\item \begin{minted}{SQL}
|
||||||
|
delete from works
|
||||||
|
where company_name = "Small Bank Corporation";
|
||||||
|
\end{minted}
|
||||||
|
\end{enumerate}
|
||||||
|
}
|
||||||
|
\questionandanswer[]{
|
||||||
|
教材3.18:请给出图 \ref{fig:3-19} 的雇员数据库的SQL模式定义。为每个属性选择合适的域,并为每个关系模式选择合适的主码。引人任何合理的外码约束。
|
||||||
|
}{}
|
||||||
|
{\kaishu
|
||||||
|
\begin{minted}{SQL}
|
||||||
|
create table employee
|
||||||
|
(
|
||||||
|
ID varchar(10) primary key,
|
||||||
|
person_name varchar(100),
|
||||||
|
street varchar(100),
|
||||||
|
city varchar(100)
|
||||||
|
);
|
||||||
|
|
||||||
|
create table works
|
||||||
|
(
|
||||||
|
ID varchar(10) primary key,
|
||||||
|
company_name varchar(100),
|
||||||
|
salary decimal(8, 2)
|
||||||
|
);
|
||||||
|
|
||||||
|
create table company
|
||||||
|
(
|
||||||
|
company_name varchar(100) primary key,
|
||||||
|
city varchar(100)
|
||||||
|
);
|
||||||
|
|
||||||
|
create table manages
|
||||||
|
(
|
||||||
|
ID varchar(10) primary key,
|
||||||
|
manager_id varchar(10)
|
||||||
|
);
|
||||||
|
|
||||||
|
|
||||||
|
alter table works
|
||||||
|
add (foreign key (company_name) references company (company_name),
|
||||||
|
foreign key (ID) references employee (ID));
|
||||||
|
|
||||||
|
alter table manages
|
||||||
|
add (foreign key (manager_id) references employee (ID),
|
||||||
|
foreign key (ID) references employee (ID));
|
||||||
|
\end{minted}
|
||||||
|
}
|
||||||
|
\end{enumerate}
|
||||||
|
|
||||||
|
\end{document}
|
Loading…
Reference in New Issue
Block a user