API design: Which version of versioning is right for you? | Hacker News Reader

API design: Which version of versioning is right for you?

cloud.google.com

51 pointsmoks15 comments

One issue I see with HTTP header versioning, is that now I have to add the HTTP version header into my routing logic. I already have to take domain, sub-domain, protocol, HTTP-verb, application context, and path into account when considering routing of HTTP requests. When defining my API signature, I would now need to include more than just the URL path and verb, I would need to include the version header value. I can imagine troubleshooting with other engineers.. "wait can you look through the logs and verify the http version header you used for that request."

Also, I really like when one piece of data can serve dual purpose use. In this case, the URL, which is highly visible, contains the version in it. So the one string of data reveals more information to me.

   func metrics(w http.ResponseWriter, r *http.Request) {
   }
   func AddV1Routes(r *mux.Router) {
    r.HandleFunc("/metrics/", metrics).Methods("GET")
    r.HandleFunc("/v1func/", v1func).Methods("GET")
   }
   func AddV2Routes(r *mux.Router) {
    r.HandleFunc("/metrics/", metrics).Methods("GET")
    r.HandleFunc("/v2func/", v1func).Methods("GET")
   }


  r := mux.NewRouter()
  AddV1Routes(r.PathPrefix("/api/v1").Subrouter())
  AddV2Routes(r.PathPrefix("/api/v2").Subrouter())
  handler := c.Handler(r)
  http.ListenAndServe(":5000", handler

With a pattern like this, you can just decide which functions to group by version.

Or, you could just deploy a copy of each version of the code using a reverse proxy.

So many texts no one line of example? I am totally confused...

His talk at Google Next 18 was much more clear:

https://www.youtube.com/watch?v=P0a7PwRNLVU

It might feel a bit aimless at first, but he pulls it together about halfway through. By the end makes good, insightful points. Definitely worth the watch.

A lot of the problems related to organization of entities and type description can be solved by using out-of-band type metadata (i.e. GraphQL). It’s not a panacea by any means, and may not be the right fit for many APIs, but I’m surprised it wasn’t mentioned at all, considering that the article is about API design.

Or no versioning.

Create new endpoint URIs for new semantics.

Use vocabulary-based (property-based) entities and never change word (property) semantics.

Contrary to its unpopularity (or because of?), this “strategy”* is far far simpler to maintain and evolve than is “versioning”.

*I say “strategy” because this is more commonsense: it’s how we do things in the real world anyway.

I think this is still versioning. You just end up versioning with the endpoint URI. Eg many names that end -2 or -3.

If the semantics change, why do you need a “-2”?

If the semantics don’t change, why do you need a new URI.

Keep in mind you don’t have to focus on the suffix or the last part of the path name when you are naming things by URI.

Also, why is it versioning? You have two endpoints in use. One happened to exist before the other one. That’s an incidental property. And there are scenarios where you may deprecate the latter before the former.

Thinking in “versions” is not an accurate/ergonomic mapping of the domain anyway; we’ve just been tricked into thinking that way because... anthropology something.

> If the semantics change, why do you need a “-2”?

Because the API may serve the same purpose as the original, with some modified semantics.

> Also, why is it versioning?

Because the API may serve the same purpose as the original, with some modified semantics.

Sharing a common name with versioning helps identify the relationships between the versions. It also avoids polluting the API namespace with a set of different names for different versions of essentially the same operations.

> Thinking in “versions” is not an accurate/ergonomic mapping of the domain anyway; we’ve just been tricked into thinking that way because... anthropology something.

It feels to me like you're trying to hard to be contrarian, or perhaps you're dealing with some special case scenario where your strategy makes sense.

You'd need to explain the argument in significantly more detail to have a chance of making it convincing.

See the link to Rich Hickey’s talk in my other comment. It covers the issues pretty well.

I’m not being contrarian, but I am expressing frustration with a common yet unnecessarily complicated approach to a general problem. But I stand by my statements, they’ve paid off tremendously in real world contexts.

Could you please give an example of these property-based entities, or link to an API that follows this approach? I think I understand your suggestion but want to be sure. New endpoints for new semantics seems rock-solid to me along same lines as content-addressable / immutable build artifacts which I endorse.

Rich Hickey explains a similar thinking in this talk https://youtu.be/oyLBGkS5ICk but it’s more directed at code libraries. The same arguments apply.

I 100% think this is the way to go for internal APIs.

For public APIs I may change my opinion here for more sociological/political reasons. The rational/technical reasons all still remain.

Thanks! Helpful link, and args simply / well-put. :)

Google's great at fast tracking APIs to their end of life.

Reply on news.ycombinator.com