关系数据库设计 - 两个关系 1:1 还是一个 1:2?

发布于 2024-09-29 09:43:11 字数 2761 浏览 4 评论 0原文

这道题是关于如何设计SQL关系的。我在这件事上是个新手,我想知道更多专家的答案...

我目前正在使用 MeGrok 和 SqlAlchemy 将 ZopeDB(面向对象)数据库迁移到 MySQL(关系)(尽管我不这样做)认为这真的不太相关,因为我的问题更多是关于在关系数据库中设计关系)。

我有两个这样相关的类:

class Child(object):
    def __init__(self):
        self.field1 = "hello world"

class Parent(object):
    def __init__(self):
        self.child1 = Child()
        self.child2 = Child()

“Parent”类有两个不同的 Child() 类实例。我什至不确定如何对待这个(两个不同的 1:1 关系或 1:2 关系)。

目前,我有这个:

class Child(rdb.Model):
    rdb.metadata(metadata)
    rdb.tablename("children_table")
    id = Column("id", Integer, primary_key=True)
    field1 = Column("field1", String(64))   #Irrelevant
    def __init__(self):
        self.field1 = "hello world"

class Parent(rdb.Model):
    rdb.metadata(metadata)
    rdb.tablename("parent_table")

    id = Column("id", Integer, primary_key=True)
    child1_id = Column("child_1_id", Integer, ForeignKey("children_table.id"))
    child2_id = Column("child_2_id", Integer, ForeignKey("children_table.id"))

    child1 = relationship(Child,
        primaryjoin = ("parent_table.child1_id == children_table.id")
    )

    child2 = relationship(Child,
        primaryjoin = ("parent_table.child2_id == children_table.id")
    ) 

意思是...好吧,我将两个“子”id 作为外键存储在父级中,并使用该信息检索子级本身。

这工作正常,但我不知道这是否是最合适的解决方案。

或者我可以这样做:

class Child(rdb.Model):
    rdb.metadata(metadata)
    rdb.tablename("children_table")
    id = Column("id", Integer, primary_key=True)
    parent_id = Column("id", Integer, ForeignKey("parent_table.id"))  # New!
    type = Column("type", ShortInteger) # New!

    field1 = Column("field1", String(64))  #Irrelevant
    def __init__(self):
        self.field1 = "hello world"

class Parent(rdb.Model):
    rdb.metadata(metadata)
    rdb.tablename("parent_table")

    id = Column("id", Integer, primary_key=True)
    child1 = relationship(
        # Well... this I still don't know how to write it down,
        # but it would be something like:
        #   Give me all the children whose "parent_id" is my own "id"
        #   AND their type == 1
        # I'll deal with the joins and the actual implementation depending 
        # on your answer, guys
    )

    child2 = relationship(
        #  Would be same as above
        #  but selecting children whose type == 2
    )

这可能有利于向父类添加新的子级...如果我添加“Parent.child3”,我只需要创建一个与现有关系非常相似的新关系。

我现在的方式意味着创建一个新的关系并向父级添加一个新的外键。

另外,拥有一个带有一堆外键的“父”表可能不会使它成为世界上最好的“父”表,对吧?

我想知道对数据库了解更多的人是怎么想的:)

谢谢。

PS:相关帖子? 问题 3998545

This question is about how to design a SQL relationship. I am pretty newbie in this matter and I'd like to know the answers of (way) more experts guys...

I am currently migrating a ZopeDB (Object oriented) database to MySQL (relational) using MeGrok and SqlAlchemy (although I don't think that's really too relevant, since my question is more about designing a relationship in a relational database).

I have two classes related like this:

class Child(object):
    def __init__(self):
        self.field1 = "hello world"

class Parent(object):
    def __init__(self):
        self.child1 = Child()
        self.child2 = Child()

The "Parent" class has two different instances of a Child() class. I am not even sure about how to treat this (two different 1:1 relationships or a 1:2 relationship).

Currently, I have this:

class Child(rdb.Model):
    rdb.metadata(metadata)
    rdb.tablename("children_table")
    id = Column("id", Integer, primary_key=True)
    field1 = Column("field1", String(64))   #Irrelevant
    def __init__(self):
        self.field1 = "hello world"

class Parent(rdb.Model):
    rdb.metadata(metadata)
    rdb.tablename("parent_table")

    id = Column("id", Integer, primary_key=True)
    child1_id = Column("child_1_id", Integer, ForeignKey("children_table.id"))
    child2_id = Column("child_2_id", Integer, ForeignKey("children_table.id"))

    child1 = relationship(Child,
        primaryjoin = ("parent_table.child1_id == children_table.id")
    )

    child2 = relationship(Child,
        primaryjoin = ("parent_table.child2_id == children_table.id")
    ) 

Meaning... Ok, I store the two "children" ids as foreign keys in the Parent and retrieve the children itself using that information.

This is working fine, but I don't know if it's the most proper solution.

Or I could do something like:

class Child(rdb.Model):
    rdb.metadata(metadata)
    rdb.tablename("children_table")
    id = Column("id", Integer, primary_key=True)
    parent_id = Column("id", Integer, ForeignKey("parent_table.id"))  # New!
    type = Column("type", ShortInteger) # New!

    field1 = Column("field1", String(64))  #Irrelevant
    def __init__(self):
        self.field1 = "hello world"

class Parent(rdb.Model):
    rdb.metadata(metadata)
    rdb.tablename("parent_table")

    id = Column("id", Integer, primary_key=True)
    child1 = relationship(
        # Well... this I still don't know how to write it down,
        # but it would be something like:
        #   Give me all the children whose "parent_id" is my own "id"
        #   AND their type == 1
        # I'll deal with the joins and the actual implementation depending 
        # on your answer, guys
    )

    child2 = relationship(
        #  Would be same as above
        #  but selecting children whose type == 2
    )

This may be good for adding new children to the parent class... If I add a "Parent.child3", I just need to create a new relationship very similar to the already existing ones.

The way I have it now would imply creating a new relationship AND adding a new foreign key to the parent.

Also, having a "parent" table with a bunch of foreign keys may not make it the best "parent" table in the world, right?

I'd like to know what people that know much more about databases think :)

Thank you.

PS: Related post? Question 3998545

如果你对这篇内容有疑问,欢迎到本站社区发帖提问 参与讨论,获取更多帮助,或者扫码二维码加入 Web 技术交流群。

扫码二维码加入Web技术交流群

发布评论

需要 登录 才能够评论, 你可以免费 注册 一个本站的账号。

评论(3

凉城凉梦凉人心 2024-10-06 09:43:11

针对评论进行扩展

问题是,您正在用您知道(可以理解)的术语进行思考,并且您具有面向对象数据库的局限性......这不太好延续到关系数据库。因此,出于多种原因,最好简单地识别实体和关系,并对它们进行规范化。您用来调用的方法很容易改变,并且您不会仅限于现在所拥有的。

这里有一些很好的答案,但即使这些答案也是有限且不完整的。如果您标准化 Parent 和 Child(作为人,他们将有许多公共列),您将得到 Person,没有重复的列。

人们与其他人、他们的父母有“向上”的关系,但那是上下文,而不是父母首先作为一个人存在的事实(如果你愿意,你可以有两个以上)。人们与孩子也有“向下”的关系,也是有背景的。每个父母有两个孩子的限制是荒谬的(您可能必须检查您的方法/类:我怀疑一个是“向上”导航,另一个是“向下”)。并且您不想将关系存储为重复项(一旦 Fred 是 Sally 的父亲;两次 Sally 是 Fred 的孩子),该单个事实存在于单行中,可以解释为 Parent⇢Child 或父母⇠孩子。

这个要求在很多问题中都出现过,因此我使用一个通用但详细的说明。该模型定义了任何需要向上或向下遍历的树结构,通过简单的递归处理。它称为物料清单结构,最初是为库存控制系统创建的,可应用于任何树形结构要求。它是第五范式;没有重复的列;无更新异常。

物料清单

对于装配体和零部件,它们有很多公共列,它们被归一化为零件;它们是组件还是组件都是上下文相关的,并且这些上下文列位于关联(多对多)表中。

两个关系 1:1 还是一个 1:2?

实际上,它是 1::n 的两倍。

序号或排名在主键中是明确的(按时间顺序)。如果需要其他序数,只需向关联表添加一列即可。更好的是,它确实是一个派生列,因此可以在运行时根据当前值进行计算。

Expanded in Response to Comments

The issue is, you are thinking in the terms that you know (understandable), and you have the limitations of an OO database ... which would not be good to carry over into the Relational db. So for many reasons, it is best to simply identify the Entities and Relations, and to Normalise them. The method you use to call is easy to change and you will not be limited to only what you have now.

There are some good answers here, but even those are limited and incomplete. If you Normalise Parent and Child (being people, they will have many common columns), you get Person, with no duplicated columns.

People have "upward" relations to other people, their Parents, but that is context, not the fact that the Parent exists as a Person first (and you can have more than two if you like). People also have "downward" relations to their Children, also contextual. The limitation of two children per Parent is absurd (you may have to inspect your methods/classes: I suspect one is an "upward" navigation and the other is "downward"). And you do not want to have to store the relations as duplicates (once that Fred is a father of Sally; twice that Sally is a child of Fred), that single fact exists in a single row, which can be interpreted Parent⇢Child or Parent⇠Child.

This requirement has come up in many questions, therefore I am using a single generic, but detailed, illustration. The model defines any tree structure that needs to be walked up or down, handled by simple recursion. It is called a Bill of Materials structure, originally created for inventory control systems, and can be applied to any tree structure requirement. It is Fifth Normal Form; no duplicate columns; no Update Anomalies.

Bill of Materials

For Assemblies and Components, which would have many common columns, they are Normalised into Part; whether they are Assemblies or Components is contextual, and these contextual columns are located in the Associative (many-to-many) table.

Two Relations 1:1 or one 1:2 ?

Actually, it is two times 1::n.

Ordinals, or Ranking, is explicit in the Primary Key (chronological order). If some other ordinal is required, simply add a column to the Associative table. better yet, it is truly a derived column, so compute it at runtime from current values.

披肩女神 2024-10-06 09:43:11

我承认我对对象数据库不太熟悉,但从关系术语来看,这是一种简单的一对多(可选)关系。

create table parent (
  id           int PK,
  otherField   whatever
)

create table child (
  id           int PK,
  parent_id    int Fk,
  otherField   whatever
)

显然,这不是可用的代码......

我认为这与您的第二个示例类似。如果您需要跟踪子级在与父级的关系中的顺序位置,您可以向子表添加一列,例如:

create table child (
  id           int PK,
  parent_id    int Fk,
  birth_order  int,
  otherField   whatever
)

您必须负责在应用程序级别管理该字段,这不是什么您可以期待 DBMS 为您做的事情。

我将其称为可选关系,假设没有孩子的父母可以存在 - 如果这不是真的,那么它在逻辑上就成为必需的关系,尽管您仍然必须让 DBMS 创建新的父记录无子项,然后获取其 id 来创建子项,并再次在应用程序级别管理需求。

I'll admit that I'm not too familiar with object databases, but in relational terms this is a straightforward one-to-many (optional) relationship.

create table parent (
  id           int PK,
  otherField   whatever
)

create table child (
  id           int PK,
  parent_id    int Fk,
  otherField   whatever
)

Obviously, that's not usable code as it stands....

I think this is similar to your second example. If you need to track the ordinal postion of the children in their relationships to the parent, you'd add a column to the child table such as:

create table child (
  id           int PK,
  parent_id    int Fk,
  birth_order  int,
  otherField   whatever
)

You'd have to be responsible for managing that field at teh application level, it's not something you can expect the DBMS to do for you.

I called it an optional relationship on the assumption that childless parents can exist--if that's not true, it becomes a required relationship logically, though you'd still have to let the DBMS create a new parent record childlessly, then grab its id to create the child--and once again manage the requirement at the application level.

月光色 2024-10-06 09:43:11

这可能有点断章取义,因为我没有使用你提到的任何东西 - 但就总体设计而言,这里有一些想法:

  1. 保持基于常见类型的关系:has_one、has_many、belongs_to、has_and_belongs_to_many 。
  2. 对于孩子,最好不要显式指定 N 个孩子;要么没有,要么有一个,或者可能有很多。因此,您的 child1child2 模型声明将被替换为单个属性 - 包含子项的数组。

老实说,我不知道它与您正在使用的产品的契合程度如何。然而,这通常就是 ORM 意义上关系的运作方式。因此,基于此:

  1. 如果一个模型属于另一个模型(它具有另一个表的外键),则它将具有一个 parent [sic] 属性,并引用父对象
  2. 如果一个模型具有属于它的一个模型(另一个模型具有第一个模型表的外键),它将有一个 child [sic] 属性,其中包含对子对象的引用
  3. 如果一个模型有多个模型属于它的(许多其他模型具有第一个模型表的外键),它将具有一个 children [sic] 属性,该属性是对子对象的引用的数组
  4. 如果模型具有并属于许多其他模型...您可能需要考虑使用 parentschildren 属性,或类似的东西;命名法并不重要,重要的是您可以访问它所属的一组模型以及属于它的另一组模型。

抱歉,如果这完全没有帮助,但它可能会带来一些启发。 HTH。

This is probably a little out of context, since I use none of the things you've mentioned - but as far as the general design goes, here are a couple ideas:

  1. Keep relationships based on common types: has_one, has_many, belongs_to, has_and_belongs_to_many.
  2. With children, it's better to not specify N number of children explicitly; either there are none, one, or there could potentially be many. Thus your model declarations of child1 and child2 would be replaced by a single property - an array containing children.

To be totally honest, I don't know how well that fits in with what you're using. However, that's generally how relationships work in an ORM sense. So, based on this:

  1. If a model belongs to another (it has a foreign key for another table), it would have a parent [sic] property with a reference to the parent object
  2. If a model has one model that belongs to it (the other model has a foreign key to the first model's table), it would have a child [sic] property with a reference to the child object
  3. If a model has many models that belong to it (many other models have foreign keys to the first model's table), it would have a children [sic] property that is an array of references to child objects
  4. If a model has and belongs to many other models... you might want to consider using both parents and children properties, or something similar; nomenclature is less important than you having access to a group of models that it belongs to, and another group of models that belong to it.

Sorry if that's totally unhelpful, but it might shed some light. HTH.

~没有更多了~
我们使用 Cookies 和其他技术来定制您的体验包括您的登录状态等。通过阅读我们的 隐私政策 了解更多相关信息。 单击 接受 或继续使用网站,即表示您同意使用 Cookies 和您的相关数据。
原文