> If we drop the "unique identifier" aspect of a key, and we can distinguish between different objects that share the same value according to one key by using a different key, then anything can be a key.
And the point the article is making, and that I was trying to emphasize further in my top-level comment, is that there are no perfectly unique identifiers. There are no UIDs in nature. UIDs do exist in human-designed systems, but they are only as good as the actual implementation of the system is, including its evolution over time.
Your own surrogate keys are the most reliable UIDs, because while you can still screw those up, it's in your power to avoid the mistake, and it's also in your power to fix any problem. Once you start using as keys "UIDs" that aren't owned by you, you lose both the ability to prevent bugs and to fix them.
For example, if you're making a library system and use a surrogate ID as a key, duplicates may happen - but they're always indicative of someone screwing up data entry, and can be trivially corrected by assigning a new ID for the entry. If you're using ISBN, or EAN-13, or some unholy hash of author/title/publication year, you're going to get a lot of corner cases, and you won't be able to fix them (you can't exactly change the barcode on everyone's copy of a printed book, just to undo the problems caused by the publisher playing funny games with SKUs).
> People say "natural key" when they're talking about a unique identifier (that's the "key" part) that is part of the domain model of an object (the "natural" part.)
Yes; under this phrasing, the point of TFA/my top comment is, your "natural key" is going to be only as good as your domain model of an object is - which, in practice, means it's going to be bad, because if your model defines some property correlated with real world to be unique, it's already broken.
Ultimately, the point is: there are no truly unique keys in the real world. Best you can do is to chose a key that's least likely to be duplicated in data. Perhaps so unlikely that, in non-critical systems, you can get away pretending it's unique.
And the point the article is making, and that I was trying to emphasize further in my top-level comment, is that there are no perfectly unique identifiers. There are no UIDs in nature. UIDs do exist in human-designed systems, but they are only as good as the actual implementation of the system is, including its evolution over time.
Your own surrogate keys are the most reliable UIDs, because while you can still screw those up, it's in your power to avoid the mistake, and it's also in your power to fix any problem. Once you start using as keys "UIDs" that aren't owned by you, you lose both the ability to prevent bugs and to fix them.
For example, if you're making a library system and use a surrogate ID as a key, duplicates may happen - but they're always indicative of someone screwing up data entry, and can be trivially corrected by assigning a new ID for the entry. If you're using ISBN, or EAN-13, or some unholy hash of author/title/publication year, you're going to get a lot of corner cases, and you won't be able to fix them (you can't exactly change the barcode on everyone's copy of a printed book, just to undo the problems caused by the publisher playing funny games with SKUs).
> People say "natural key" when they're talking about a unique identifier (that's the "key" part) that is part of the domain model of an object (the "natural" part.)
Yes; under this phrasing, the point of TFA/my top comment is, your "natural key" is going to be only as good as your domain model of an object is - which, in practice, means it's going to be bad, because if your model defines some property correlated with real world to be unique, it's already broken.
Ultimately, the point is: there are no truly unique keys in the real world. Best you can do is to chose a key that's least likely to be duplicated in data. Perhaps so unlikely that, in non-critical systems, you can get away pretending it's unique.