call for data on HTTP routing

[Jonathan Amsterdam]

At https://github.com/golang/go/discussions/60227, we're discussing enhancements to the standard library's http.ServeMux routing. Two issues have arisen that we could use your help on.

The first is easy to state: does fast routing matter? Are there Go web servers for which http.ServeMux is too slow, where it consumes a significant fraction of the serving latency? My position (which is also the null hypothesis) is that no, for the vast majority of applications routing time is noise. Some evidence for that: the gorilla/mux router  (https://github.com/gorilla/mux) is about 60 times slower than the fastest routers, but as far as we can tell it's still widely used (despite being unmaintained for two years). Change my view! Do you have a production service for which you had to abandon http.ServeMux because it was too slow? Tell me about it. Specific routing schemes and benchmarks are a plus.

The second question is about a specific routing scheme, and here I have no preconceived notions. Here is the context: given two routing patterns that might otherwise conflict, we'd like to allow both by preferring one over the other, where this makes sense. Currently, we prefer a pattern that specifies a method over one that doesn't; and after that, we prefer a pattern with a more specific path. For example,

    GET /foo/

wins over

    /foo/

because the first specifies a method, but it loses to

    GET /foo/bar

because the latter's path is more specific (it matches only "/foo/bar", while "/foo/" matches any path beginning "/foo").

Those three patterns make sense together:

    GET /foo/bar                     match a specific GET request

    GET /foo/                           match any GET request beginning /foo

    /foo/                                    match a request with any method beginning /foo

But what about these two patterns:

    GET /foo/

    /foo/bar

The first pattern specifies a method, but the second has a more specific path. Which should win? Or should this combination be disallowed because it's too confusing? My question is, do any pairs of patterns like this show up in practice? And if so, which one should take precedence?

Thanks in advance for your help.

[Ian Lance Taylor]

On Thu, Jun 1, 2023 at 2:23 PM 'Jonathan Amsterdam' via golang-nuts
<golan...@googlegroups.com> wrote:
>
> The second question is about a specific routing scheme, and here I have no preconceived notions. Here is the context: given two routing patterns that might otherwise conflict, we'd like to allow both by preferring one over the other, where this makes sense. Currently, we prefer a pattern that specifies a method over one that doesn't; and after that, we prefer a pattern with a more specific path. For example,
>
> GET /foo/
>
> wins over
>
> /foo/
>
> because the first specifies a method, but it loses to
>
> GET /foo/bar
>
> because the latter's path is more specific (it matches only "/foo/bar", while "/foo/" matches any path beginning "/foo").
>
> Those three patterns make sense together:
>
> GET /foo/bar match a specific GET request
> GET /foo/ match any GET request beginning /foo
> /foo/ match a request with any method beginning /foo
>
> But what about these two patterns:
>
> GET /foo/
> /foo/bar
>
> The first pattern specifies a method, but the second has a more specific path. Which should win? Or should this combination be disallowed because it's too confusing? My question is, do any pairs of patterns like this show up in practice? And if so, which one should take precedence?

This is an obvious point, but in general for any case where two
different patterns can match an HTTP request it is possible to write a
single pattern that matches both and then let the function
differentiate based on the precise request. On the other hand, when
it is unclear which pattern is going to match a specific HTTP request,
it seems quite possible for the programmer to make a mistake as to
which one will match, and since the cases are by construction obscure
it seems easy to fail to detect such a case in testing.

This suggests that one possible approach is to use a very simple
request matching rule that no reasonable person can misunderstand, and
for the router to give an error on ambiguous cases.

Ian

[Andrew Harris]

I wanted to make a quick observation w/r/t Ian's suggestion of not registering ambiguous patterns. The tradeoff isn't ultimately about what behaviors can be associated with what routes, but around what is required to register routes. To intentionally use the kinds of routes we'd interpret as ambiguous, it might be plausible to (1) register a methodless pattern, and move the method-matching into the associated handler. Or, (2) register a methodless pattern with e.g. a blanket 404 or 405 handler, and then register a the same pattern with a method and a more active handler. Or, something else - it would add friction to setting up this kind of routing logic (maybe more verbose, maybe more explicit), but not prohibit it.

(1)
mux.HandleFunc("/foo/", func...

    ...

    if req.Method == http.MethodGet ...

    ...

(2)
mux.Hanlde("/foo/", notReallyFooHandler)
mux.Handle("GET /foo/", fooHandler)