Don't call it `*_id`!

In a distributed system, we often need to store data that we do not own. We might use it as a unique identifier across domains or the system we do own needs to proxy it to yet another service.

In many cases, the origin system exposes it in the same way it is represented internally, for instance: a foreign key name of a relational database:

Fig. 1: local relational table, rev. 1

+----+--------------+----------------------+
| id | name         | backend              |
+----+--------------+----------------------+
| 42 | Marley Spoon | elixir, ruby, python |
+----+--------------+----------------------+

Translated to JSON, we could send it over the wire as follows:

Fig. 2: example JSON payload

{ 
  "id": 42, 
  "name": "Marley Spoon",
  "backend": ["elixir", "ruby", "python"]
}

A service consuming the information may want to add prefixes to scope it for "companies" and store/cache it like so:

Fig. 3: local relational table, rev. 2

+------------+--------------+----------------------+
| company_id | company_name | backend              |
+------------+--------------+----------------------+
| 42         | Marley Spoon | elixir, ruby, python |
+------------+--------------+----------------------+

But here's the problem: as engineers, we look at _id fields and immediately think of it as integers. However, the consuming service has no control over the data it receives and the data type is only assumed.

If you use that distributed ID field as a local foreign key: some external system controls the value and an unforeseen change might break our setup.

Identification

It has proven valuable to us to use the pattern *_identifier to indicate that…

1. it is some kind of a unique identifier
2. some other system has control over it

If the *_identifier value is to be stored, it should always be saved as a string type. Almost anything can be coerced into a string, and that way we guarantee that the origin system can choose whatever they want for their unique identifier.

This is particularly true if the origin system decided to move to using UUIDs. A final version of the local relational table above could look like this:

Fig. 4: local relational table, rev. 4

+--------------------------------------+--------------+----
| company_identifier                   | company_name | …
+--------------------------------------+--------------+----
| 328129ae-df4e-4168-94d3-2572b4b343ef | Marley Spoon | …
+--------------------------------------+--------------+----

Payload

If the system exposing the data is controlled by your organisation, we can support this at the source.

It is a common pitfall of API designs, especially RESTful APIs, to expose a resource exactly like you represent it in your database. This makes sense to reduce the cognitive load of the team maintaining the API. However, the data layer will inevitably change, rendering this point moot: the DB representation has to be translated to maintain a stable contract. Why not abstract from the data layer to begin with and name the keys in the payload in a system-agnostic way?

Fig. 5: JSON payload abstracted from persistence

{ 
  "company_identifier": "328129ae-df4e-4168-94d3-2572b4b343ef",
  "company_name": "Marley Spoon",
  "backend": ["elixir", "ruby", "python"]
}

Summary

1. Use *_identifier instead of *_id fields for externally owned data
2. Prefer a string type over integer for *_identifier values
3. Avoid a 1:1-map of your persistence model to your external API

Comments

[José Pablo Ramírez Vargas]

While I understand the motivation for this, I would disagree. First and foremost, we have the data size. Using a non-integer primary key will most likely increase your database size very fast and by a lot. Why? Because of foreign keys. A small table of 100 records could be serving dozens of foreign keys in tables that hold tens of millions of records. If we used a tinyint (SQL Server type for 1 byte), we could easily cover the 100 records, but if we were to use, say, a 5-letter primary key, we are now forcing foreign keys to consume 5 times more bytes. Multiply that by the number of records and you have a mega waste of storage space, which in turn puts pressure on the database server's RAM: Now less data can be stored in RAM, making queries slower.

The second one is performance around the clustered index: Indexing and ordering numbers is faster than indexing and ordering strings.

Finally, the "_id evokes number in people's mind" argument is something that is easily corrected: Don't assume a data type unless the company that provides the data has dictated a column-naming scheme that allows assumptions around the column's data type.

Where I work, for example, a suffix is mandated that most of the time directly relates to the data type. For example, _dtm for date/time values, or qty for quantities (would be a real number). Personally, I hate this kind of conventions. Whoever came up with this naming convention has never coded a backend in their lives.