“它们会使删除记录变得更加麻烦——你不能删除“主”记录，而其他表中的记录中有外键会违反这种约束。”

重要的是要记住，SQL标准定义了删除或更新外键时所采取的操作。 我所知道的是:

ON DELETE RESTRICT - Prevents any rows in the other table that have keys in this column from being deleted. This is what Ken Ray described above. ON DELETE CASCADE - If a row in the other table is deleted, delete any rows in this table that reference it. ON DELETE SET DEFAULT - If a row in the other table is deleted, set any foreign keys referencing it to the column's default. ON DELETE SET NULL - If a row in the other table is deleted, set any foreign keys referencing it in this table to null. ON DELETE NO ACTION - This foreign key only marks that it is a foreign key; namely for use in OR mappers.

这些相同的操作也适用于ON UPDATE。

默认值似乎取决于您使用的sql server。

在DB2中，如果使用mqt(物化查询表)，优化器需要外键约束才能为任何给定的查询选择正确的计划。由于元数据包含基数信息，优化器会大量使用元数据来决定是否使用MQT。

在这里回答问题的许多人都过于关注通过引用约束实现的引用完整性的重要性。在具有引用完整性的大型数据库上工作性能不佳。Oracle似乎特别不擅长级联删除。我的经验法则是，应用程序永远不应该直接更新数据库，而应该通过存储过程更新。这将代码库保存在数据库中，并意味着数据库保持其完整性。

在许多应用程序可能正在访问数据库的地方，由于引用完整性约束确实会出现问题，但这取决于控件。

还有一个更广泛的问题，应用程序开发人员可能有非常不同的需求，而数据库开发人员可能并不那么熟悉。

From my experience its always better to avoid using FKs in Database Critical Applications. I would not disagree with guys here who say FKs is a good practice but its not practical where the database is huge and has huge CRUD operations/sec. I can share without naming ... one of the biggest investment bank of doesn't have a single FK in databases. These constrains are handled by programmers while creating applications involving DB. The basic reason is when ever a new CRUD is done it has to effect multiple tables and verify for each inserts/updates, though this won't be a big issue for queries affecting single rows but it does create a huge latency when you deal with batch processing which any big bank has to do as daily tasks.

最好避免fk，但它的风险必须由程序员来处理。

使用外键的原因:

you won't get Orphaned Rows you can get nice "on delete cascade" behavior, automatically cleaning up tables knowing about the relationships between tables in the database helps the Optimizer plan your queries for most efficient execution, since it is able to get better estimates on join cardinality. FKs give a pretty big hint on what statistics are most important to collect on the database, which in turn leads to better performance they enable all kinds of auto-generated support -- ORMs can generate themselves, visualization tools will be able to create nice schema layouts for you, etc. someone new to the project will get into the flow of things faster since otherwise implicit relationships are explicitly documented

不使用外键的原因:

you are making the DB work extra on every CRUD operation because it has to check FK consistency. This can be a big cost if you have a lot of churn by enforcing relationships, FKs specify an order in which you have to add/delete things, which can lead to refusal by the DB to do what you want. (Granted, in such cases, what you are trying to do is create an Orphaned Row, and that's not usually a good thing). This is especially painful when you are doing large batch updates, and you load up one table before another, with the second table creating consistent state (but should you be doing that sort of thing if there is a possibility that the second load fails and your database is now inconsistent?). sometimes you know beforehand your data is going to be dirty, you accept that, and you want the DB to accept it you are just being lazy :-)

我认为(我不确定!)大多数已建立的数据库都提供了一种指定外键的方法，这种方法不是强制的，只是一些元数据。由于不强制执行消除了不使用fk的所有理由，如果第二部分中的任何理由适用，您可能应该走那条路。

更大的问题是:你会戴着眼罩开车吗?如果你开发一个没有参考约束的系统就是这样。请记住，业务需求会改变，应用程序设计会改变，代码中相应的逻辑假设也会改变，逻辑本身也会被重构，等等。一般来说，数据库中的约束是在当代逻辑假设下放置的，对于特定的逻辑断言和假设集似乎是正确的。

在应用程序的整个生命周期中，引用检查和数据检查约束控制通过应用程序收集的数据，特别是当新需求驱动逻辑应用程序更改时。

从实时事务处理系统的角度来看，外键本身不会“提高性能”，也不会显著“降低性能”。然而，在大容量“批处理”系统中，约束检查存在一个聚合代价。这就是实时和批量事务处理的区别;批处理——通过约束检查，按顺序处理的批处理的总成本会对性能造成影响。

在一个设计良好的系统中，数据一致性检查将在处理批处理之前完成(然而，这里也有相关的成本);因此，加载时不需要外键约束检查。事实上，所有的约束，包括外键，都应该暂时禁用，直到批处理完成。

QUERY性能——如果表是在外键上连接的，那么要认识到外键列不是索引的事实(尽管根据定义，各自的主键是索引的)。通过索引一个外键，也就是索引任何键，在索引上连接表有助于提高性能，而不是通过连接一个带外键约束的非索引键。

Changing subjects, if a database is just supporting website display/rendering content/etc and recording clicks, then a database with full constraints on all tables is over kill for such purposes. Think about it. Most websites don’t even use a database for such. For similar requirements, where data is just being recorded and not referenced per say, use an in-memory database, which does not have constraints. This doesn’t mean that there is no data model, yes logical model, but no physical data model.