net/http/httptrace: add ServerTrace hooks

[joeshaw]

Problem

A very common pattern in Go HTTP servers is to implement an http.ResponseWriter that wraps another http.ResponseWriter and captures the status code. This is often used for logging and metrics collection.

For example,

type statusCaptureWriter struct {
    http.ResponseWriter
    status int
}

func (scw *statusCaptureWriter) WriteHeader(status int) {
    scw.status = status
    scw.ResponseWriter.WriteHeader(status)
}

type loggedHandler struct {
    handler http.Handler
    logger SomeLogger
}

func (h *loggedHandler) ServeHTTP(w http.ResponseWriter, r *http.Request) {
    scw := &statusCaptureWriter{ResponseWriter: w}
    h.handler.ServeHTTP(scw, r)
    h.logger.Logf("status=%d ...", scw.status)
}

I've written something like this a bunch of times. You can find examples in nearly every Go web framework out there. One example is https://github.com/urfave/negroni/blob/master/response_writer.go#L13-L26.

There are some issues with this approach. For instance, my statusCaptureWriter doesn't implement other interfaces like http.Flusher, http.CloseNotifier or http.Pusher. I can't determine at compile time whether the underlying http.ResponseWriter implementation implements any of these interfaces, so if I choose to implement them I might lie to callers at higher levels of the stack and inadvertently break things. (This particularly a problem with CloseNotifier.)

Proposal (rejected, see below)

I'd like to propose an additional interface, http.Statuser (better name welcome) that exposes the status code within a http.ResponseWriter implementation. The internal http.(*response) implementation already tracks the status code written, so this can just be exposed and it will automatically implement this interface.

Software could avoid wrapping the http.ResponseWriter by instead type asserting it to http.Statuser and getting the status as needed there. (And it could optionally continue to wrap the ResponseWriter as needed until this is widely deployed.)

type loggedHandler struct {
    handler http.Handler
    logger SomeLogger
}

func (h *loggedHandler) ServeHTTP(w http.ResponseWriter, r *http.Request) {
    // Could check for http.Statuser implementation here and wrap http.ResponseWriter 
    // if necessary, but omitting for brevity

    h.handler.ServeHTTP(w, r)

    status := 0
    if s, ok := w.(http.Statuser); ok {
        status = s.Status()
    }

    h.logger.Logf("status=%d ...", status)
}

Alternative proposal

Implement an httptrace.ServerTrace struct that is analogous to the ClientTrace already there. See #18997 (comment) for more info.

[joeshaw]

More examples:

https://github.com/gobuffalo/buffalo/blob/5b5e01469cfd064cad0a56b1eee4728a41d2ec46/response.go

https://github.com/gin-gonic/gin/blob/2dae550eb5392006a4582ce9c90016a9b5a74e8b/response_writer.go#L19-L48

https://github.com/goji/glogrus2/blob/master/writer_proxy.go

https://github.com/rikonor/go-status-code-tracker

[bradfitz]

These exist apparently, but why?

Code should know what it did previously.

[joeshaw]

@bradfitz The example I posted is a simplified instance: you create a wrapper that handles logging of all requests and then the individual handlers don't need to do or understand logging themselves. They are free to just call w.WriteHeader(http.StatusOK) and things get logged at a higher level. This plays nicely with middleware, too.

Other things that might get logged there are the URL/path, the duration of the call, etc. It would be a pain to have to do that at essentially every exit site of every middleware or handler. I think it's also a valid separation of concerns.

[joeshaw]

Another example beyond logging is collection of metrics. You need to know when you serve 5xx errors, for instance. Again, you don't necessarily want to be doing this everywhere it happens (and in the case of panics, I'm not sure offhand it's possible.)

[bradfitz]

Rather than a dozen optional methods, then, I'd prefer to see something like a State() method that returned a concrete struct type to which we could add fields over time.

[gopherbot]

CL https://golang.org/cl/36647 mentions this issue.

[urandom]

Given that the primary target of this feature is logging, I suggest that the length of the response body would be a good addition to the output of State()

[bradfitz]

What if we just provide one canonical ResponseWriter wrapper instead?

And if this is about logging, we should (re-)consider the HTTP logging issues/proposals.

[joeshaw]

@bradfitz the thing that worries me about providing a wrapper type is also wrapping the other interfaces that may optionally be supported by the underlying type. For instance, http.Pusher is something that's only supported by the http2 ResponseWriters. This wrapper type either has to eschew support for Pusher or it has to "lie" about supporting it.

[tombergan]

I've written a wrapper type like this as well, except instead of just logging the response code, my wrapper type also logged the number of response body bytes written, an md5sum of the response body, the time between the first and last Write call, and a few other miscellaneous things. Skimming through the examples in this comment, I see that I'm not alone -- many of the wrappers provide more than just the response code. I worry that adding types like http.Statuser is an endless game of whack-a-mole. Even with Brad's State() idea, I wouldn't expect the http library to compute an md5sum of the response body.

@joeshaw, have you considered designing a server-side version of httptrace? We considered this at one time, but never worked through the details. Also see #18095 and #3344.

These exist apparently, but why?
Code should know what it did previously.

It is convenient to put all logging code in one place. One way to do this is by wrapping the http.ResponseWriter. This approach becomes more compelling when the logger tracks a bunch of things such as timing info or a hash of the response body, rather than just the response code (which is easy to remember by itself).

[joeshaw]

@tombergan i hadn't considered an httptrace-equivalent, but perhaps that is a better way to go in order to support different use cases like hashing response bodies.

[gopherbot]

CL https://golang.org/cl/37010 mentions this issue.

[joeshaw]

I just uploaded a CL that has the very start of this by introducing an httputil.ServerTrace struct. At the moment it only implements a WroteHeader hook. It'd be nice if people could take a look and give an opinion on the general approach.

I expect to add StartHandler and EndHandler hooks as well. Since it's implemented in the server, it could contain connection-level hooks as well.

One things that's a little strange about it is that the data isn't accessible from the handlers, which I imagine could be a problem if logging code wanted to include additional request-specific data. One possibility might be to pass the http.Request into the hooks, have handlers add values to the req.Context and have an EndHandler hook extract and log there?

[tombergan]

I think you're right that the server has both connection-level events and request-level events, and those may need to be handled separately. We could start by supporting request-level events only, since that seems to be what is wanted the most.

A nice thing about ClientTrace is that it's easy to create a new ClientTrace struct for each request. For example, here's how I use ClientTrace:

info := &reqInfo{}
req = req.WithContext(httptrace.WithClientTrace(req.Context(), &httptrace.ClientTrace{
  GotConn: info.GotConn,
  ...
}))
transport.RoundTrip(req)
...
log(info)

This lets me accumulate all info about a single request into a single struct (reqInfo). It would be ideal if I could a similar pattern for server traces. I think we could do this be changing your proposal slightly:

packge http
type Server struct {
  ...
  NewTrace func() *httptrace.ServerTrace   // optional callback to create a trace for a new request
}

One things that's a little strange about it is that the data isn't accessible from the handlers, which I imagine could be a problem if logging code wanted to include additional request-specific data. One possibility might be to pass the http.Request into the hooks, have handlers add values to the req.Context and have an EndHandler hook extract and log there?

That would work. Another idea is to add an opaquely-typed field to ServerTrace that can be set by NewTrace. Perhaps UserData interface{}. Then, the request handler could extract the trace info using httptrace.ContextServerTrace(req.Context()).

[Merovius]

@joeshaw It is theoretically possible to dynamically decide, whether to wrap pusher or not:

func wrap(wrap, with http.ResponseWriter) http.ResponseWriter {
    if p, ok := wrap.(http.Pusher); ok {
        return struct {
            http.ResponseWriter
            http.Pusher
        }{ with, p }
    }
    return with
}

The trouble is a) the combinatoric explosion. You have to provide a branch for every combination of interface that could exist. And b) that the wrapping code must know about all possible optional interfaces. You can get around these by providing generic wrapping-code, but that also only solves the issue for all the optional interfaces in net/http itself, not potential third-party packages using the same trick.

This is why I'm so troubled by "optional interfaces" and the fact that they are used in net/http so pervasively. It's a very leaky abstraction and sets a bad example.

[joeshaw]

@Merovius wow! that is a very interesting hack, and one i hadn't ever considered. I agree that it's not scalable, but thanks for pointing that out.

[bradfitz]

@tombergan,

Then, the request handler could extract the trace info using httptrace.ContextServerTrace(req.Context())

Putting things on the context might be interesting. I don't know whether that crosses the line of acceptable uses of context, though. I'm fine with it, but I think others might balk. I think it's better than more optional interfaces at least.

[rsc]

Context makes some sense for client tracing, in that the client causes it to trace all requests being made, but that's arguably OK. On the server, though, you want to scope it to one request, not all the requests being handled by the server. It's unclear that context can be made to work here.

It sounds like there's not a concrete proposal here at this point and that someone needs to write a design doc that incorporates the feedback and feature discussion above.

Or we can move this issue out of the proposal process and leave it open as a regular issue until someone does want to write a proposal.

[joeshaw]

I think it makes sense to move it out of the proposal process for now, since my original proposal wasn't well-liked.

Using httptrace sounds like a better way to go, but I think more prototyping needs to be done before it's ready to be a formal proposal.

[joeshaw]

@rhysh This is an interesting proposal. There are a few things I would tweak about it, like returning interface values for the individual items in the ResponseOptions struct. (ie, Flush func() would be Flusher http.Flusher, but in both cases the values would be nil if the interface weren't implemented.)

There are a few things I like about it:

• It makes it explicit what interfaces the ResponseWriter implementation supports.
• ResponseWriter wrappers need only to implement the ResponseWriter and ResponseOptionser interfaces.

The things I am not sure how I feel about:

• It does not address the very widespread use of ResponseWriter wrappers for common problems like storing status codes, calculating response sizes, or handler durations. (More on this below.)
• It effectively deprecates type asserting the ResponseWriter to other interfaces, except ResponseOptionser. In other words, in a handler it's no longer sufficient to do:

if p, ok := rw.(http.Pusher); ok {
    p.Push(...)
}

which is a very standard, idiomatic thing to do. (It would still succeed much of the time, since the underlying private ResponseWriter implementation would support it, but wrappers would not.)

Instead, you have to do:

if ro, ok := rw.(http.ResponseOptionser); ok {
    if ro.ResponseOptions().Pusher != nil {
        ro.ResponseOptions().Pusher.Push(...)
    }
}

which is a little longer and less natural.

Back to the proliferation of ResponseWriter wrappers. Is this a real problem?

It's likely impossible to eliminate the need for ResponseWriter wrappers entirely in all cases. But the vast majority of the duplication out there are to track status codes and, to a lesser extent, things like timings and response body sizes. Perhaps the standard library should provide these in a new State method and interface as @bradfitz first suggested?

This counterproposal doesn't address the proliferation of wrappers, whereas the ServerTrace one likely would.

Unless people are vehemently opposed to this, I may take a stab at it.

[joeshaw]

Back to the ServerTrace implementation, the separation of http and httptrace packages is really making the usability of ServerTrace ugly because of circular dependencies. GotRequest would really benefit from having access to http.Request, but instead most of the fields on it have to duplicated. Same for http.Header. It makes me worried about the long-term viability of doing it this way.

To see @pteichman's and my progress on ServerTrace thus far: https://github.com/pteichman/go/tree/servertrace

[Merovius]

An alternative would be, to enable reflect to dynamically create methods (via #15924 and #16522). That way, people who want to create wrappers with optional interfaces could do so, e.g. by

func wrap(res http.ResponseWriter, statusCallback func(int)) http.ResponseWriter {
    st := reflect.StructOf{
        reflect.StructField{
            Type: reflect.TypeOf(res),
            Anonymous: true,
        },
    }
    nt := reflect.NamedOf(st, "wrapped", []reflect.Method{
        Name: "WriteHeader",
        Type: reflect.TypeOf(func(interface{}, int) {}),
        Func: reflect.ValueOf(func(_ interface{}, code int) { statusCallback(code) }),
    }
    v := nt.New()
    v.Field(0).Set(reflect.ValueOf(st))
    return v.Interface().(http.ResponseWriter)
}

Not terribly ergonomic (and I papered over the detail of how NamedOf would work and what the argument of that function would be, given that the struct-type is dynamically created); but it would work.

[tombergan]

Stepping back a bit, the reason we originally added httptrace.ClientTrace was to expose info that could not be tracked by simply wrapping a RoundTripper. If we focus solely on better ways to wrap a ResponseWriter in this issue, then we lose sight of exposing info that is not available to ResponseWriter. For example, see #18095, which was dup'd into this issue. That issue is about logging bad requests, which we cannot do with ResponseWriter wrappers because bad requests are never forwarded to request handlers in the first place. Further, we might imagine adding HTTP/2 specific hooks in the future, or hooks to get the exact headers written on the wire, which can change slightly compared to what the caller passes to WriteHeader.

There is also #16220, which was not very well motivated in isolation, but would be fixed as a side-effect of adding UpdateRequestContext.

There is also #20956, which is yet another example of people trying to get at the underlying net.Conn (especially see this comment).

It used to do a bunch more, before Go 1.7 included context and before Go 1.8 fixed the lifespan of the Context to match the request (vs matching the lifespan of the call to ServeHTTP).

@rhysh, can you say more about this? It sounds like you're currently unable to do something that you'd like to do as a result of the Context change.

But the vast majority of the duplication out there are to track status codes and, to a lesser extent, things like timings and response body sizes. Perhaps the standard library should provide these in a new State method and interface as @bradfitz first suggested?

I am strongly opposed to this. There is no way the standard library will cover all kinds of stats that every user wants to track. Rather than getting ourselves into the business of tracking stats, we should provide APIs that allow users to track whatever stats they need.

[rhysh]

can you say more about this? It sounds like you're currently unable to do something that you'd like to do as a result of the Context change.

On the contrary: Many of the features that had to be implemented from scratch by that internal package are now done by the standard library. For instance it used to use http.CloseNotifier to learn that the client had disconnected prematurely; as of Go 1.8, the request-scoped Context now provides that signal.

As of Go 1.8, Context in net/http has been working well. The focus of the internal package has narrowed, thanks to improvements in the standard library.

---

httptrace provides a lot of information about the request lifecycle and very little of it overlaps with information available to http.RoundTripper implementations.

The partial overlap I've encountered is that my internal package would additionally wrap the request body to learn when the full request had been written. This forced chunked transfers, which unfortunately didn't work with HTTP/1.0 servers, but otherwise worked well enough. After the addition of httptrace, I was able to use the WroteRequest hook to get that signal.

Because there's so little overlap, what are the problem statements that this issue should address? Here are some possibilities:

1. It's not possible [for *http.Server / http.Handler code] to learn about invalid/malformed requests
2. It's not possible to measure how long it took to read the request headers, from first byte to full headers received
3. It's not possible to measure how long a client's connection was idle before it sent the first byte of headers, or to determine if the connection was reused from a previous (or concurrent http/2) request.
4. Measuring the status code returned by a handler is inconvenient
5. Measuring the wall clock time at which the status code was set is inconvenient
6. Writing code to measure the status code returned by a handler is likely to break features like http/2 push or sendfile (for http/1.1).

The first three seem like they'd be addressed well with a server-side equivalent to the current httptrace.ClientTrace hooks. The later three can be done by wrapping http.ResponseWriter values as middleware, but it's hard currently too hard to do well.

[tombergan]

That is a nice summary. I think all six of your listed use cases could be handled by httptrace.ServerTrace. For 4 & 5, we'd pass through the status code (and possibly headers) from the ResponseWriter to the ServerTrace. 6 should "just work" since ServerTrace doesn't wrap the ResponseWriter.

[rhysh]

I think all six of your listed use cases could be handled by httptrace.ServerTrace
It could do any of them. Where should it focus? Where should the discussion focus?

httptrace.ClientTrace doesn't allow getting the status code; that's trivial to do in a http.RoundTripper. But it also doesn't allow getting the sizes of the request body or the response body—both possible, but inconvenient, to get from a http.RoundTripper wrapper.

ClientTrace focused on giving access to net/http internals that it alone could provide.

In the current implementation draft of ServerTrace, the GotBadRequest hook gives access to information that would not otherwise be available to http.Handler authors. This looks like a good way to give access to that information.

The WroteHeader and WroteBodyChunk callbacks give access to data that http.Handler authors already have. It's not very convenient access because of the pattern of adding secret bonus methods on http.ResponseWriters, but it's access.

Should ServerTrace have a callback for when Flush is called and for when http/2 push is used? If sendfile(2) is activated, should a hook be called at the beginning/middle/end of that transfer?

I think it would be helpful to split the discussion into 1) information present in the internals of net/http that is not exposed to users, and 2) information that users have, but need better access to.

---

Use of the http.Server.UpdateRequestContext field looks clunky. It's called before the http.Request is available, meaning that users would need to coordinate between the code that runs during that hook and a top-level http.Handler in order to match the ServerTrace hooks to an http.Request.

Since it's attached to the http.Server, it can't be used by middlewares. It's set once at the top level by the owner of the service, making it inaccessible to the maintainers of middleware packages. They're likely to continue wrapping http.ResponseWriters, since that would be the only way the code they control is able to access the data they need.

I have a proposal for an alternate design, if we split the functionality more or less into the two types of data I identified above:

1. A new field on http.Server points to a struct that contains the GotBadRequest hook (and maybe some others). Consider this logically as a structured extension of the http.Server.ErrorLog field.

2. A function in net/http that takes the Context value from a http.Request (or a *http.Request directly) and something like a httptrace.ServerTrace struct, registering it for callbacks as the request is served. Any layer of http.Handler middleware could call the function, just as any layer of http.RoundTripper middleware can with ClientTrace.

(The function wouldn't return a Context like the ClientTrace equivalent does; the http.Server implementation would use a private context key to store and access a function of type func(httptrace.ServerTrace). The net/http package's function would extract and call it.)

[tombergan]

[UpdateRequestContext is] called before the http.Request is available, meaning that users would need to coordinate between the code that runs during that hook and a top-level http.Handler in order to match the ServerTrace hooks to an http.Request.

Not sure I followed? The hooks will be available via http.Request.Context(). A pattern like the following should work:

s := &http.Server{
  UpdateRequestContext: func(ctx context.Context) context.Context {
    sh := newServerHooks()  // a user-defined struct
    ctx = context.WithValue(ctx, key, sh)
    return httptrace.WithServerTrace(ctx, &httptrace.ServerTrace{
      GotRequest: sh.GotRequest,
      ...
    })
}

The handler can then use ctx.Value(key) to extract sh.

Since [UpdateRequestContext is] attached to the http.Server, it can't be used by middlewares. It's set once at the top level by the owner of the service, making it inaccessible to the maintainers of middleware packages.

The middleware needs some entry point in order to apply its http.ResponseWriter wrapper. Couldn't the middleware use a similar entry point to install its UpdateRequestContext wrapper? I am admittedly not familiar with many such libraries (aside from the few I've written myself for internal use), so I am very interested in hearing about libraries where UpdateRequestContext will be difficult to use.

A function in net/http that takes the Context value from a http.Request ...

Can you explain this with a pseudocode example? I did not follow.

A new field on http.Server points to a struct that contains the GotBadRequest hook ...
I think it would be helpful to split the discussion into 1) information present in the internals of net/http that is not exposed to users, and 2) information that users have, but need better access to.

That is a useful distinction for this discussion, however, my main concern is API clutter. I would like to minimize the API surface and avoid cluttering http.Server with callback fields, if possible. I fear we are heading towards API clutter unless the new functions/callbacks/fields we add are sufficiently general.

I am promoting two APIs: (1) Wrapping http.ResponseWriter. This will be the canonical way to affect how the response is produced. (2) UpdateRequestContext. This will be the canonical way to get a read-only view of each request for logging. There may need to be a long list of read-only hooks, but at least those hooks will be hidden in net/http/httptrace rather than cluttered in net/http.

[rhysh]

Considering the case where a user wants to log the status code of all responses, here are three possible approaches:

1. No changes to net/http. Logging package authors wrap the http.ResponseWriter, providing a function with signature func(inner http.Handler) http.Handler. Application owners can call it any time they have a http.Handler, sometime before setting their http.Server.Handler field.

2. Add a field to http.Server, UpdateRequestContext. Logging package authors write two pieces of code: their usual middleware (to access the URI path from the http.Request), and a function to add to the UpdateRequestContext field (to access the status code). Application owners call one of the functions from their http.Handler, and the other where they set up the http.Server (which may be in a different file or package).

3. Add a function in http designed to be called from an http.Handler. Logging package authors convert their http.ResponseWriter wrappers to use the new function with no change in the API they provide to users. Applications owners can call it any time they have a http.Handler, sometime before setting their http.Server.Handler field.

The shortcoming of (1) is that it blocks access to additional methods on the http.ResponseWriter. The shortcoming of (2) is that it requires users to coordinate between two different parts of the net/http API. Users who only have access to an http.Handler cannot use it (and would be forced to use the existing option 1).

Code for each follows. Does example 2 match the API design you have in mind? Can you give a more concise demonstration of correct use of the new field?

1. No changes to net/http:

type statusRecorder struct {
	status int
	w      http.ResponseWriter
}

func (srw *statusRecorder) Header() Header {
	return srw.w.Header()
}

func (srw *statusRecorder) Write(p []byte) (int, error) {
	if srw.status == 0 {
		srw.status = http.StatusOK
	}
	return srw.w.Write(p)
}

func (srw *statusRecorder) WriteHeader(status int) {
	srw.status = status
	srw.w.WriteHeader(status)
}

func logStatusMiddleware_ResponseWriter(inner http.Handler) http.Handler {
	return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
		srw := &statusRecorder{w: w}

		defer func() {
			log.Printf("path=%q status=%d", r.URL.Path, srw.status)
		}()

		inner.ServeHTTP(srw, r)
	})
}

2. Add a field to http.Server, UpdateRequestContext:

func logStatusMiddleware_AndServerUpdate(inner http.Handler, srv *http.Server) http.Handler {
	ctxKey := new(int)

	prevFn := srv.UpdateRequestContext
	newFn := func(ctx context.Context) context.Context {
		if prevFn != nil {
			ctx = prevFn(ctx)
		}
		var (
			mu   sync.Mutex
			info httptrace.WroteHeaderInfo
		)
		ctx = httptrace.WithServerTrace(ctx, &httptrace.ServerTrace{
			WroteHeader: func(i httptrace.WroteHeaderInfo) {
				mu.Lock()
				defer mu.Unlock()
				info = i
			},
		})
		ctx = context.WithValue(ctx, ctxKey, func() httptrace.WroteHeaderInfo {
			mu.Lock()
			defer mu.Unlock()
			return info
		})
	}
	srv.UpdateRequestContext = newFn

	return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
		defer func() {
			info := r.Context().Value(ctxKey).(func() httptrace.WroteHeaderInfo)()
			log.Printf("path=%q status=%d", r.URL.Path, info.StatusCode)
		}()

		inner.ServeHTTP(w, r)
	})
}

3. Add a function in http designed to be called from an http.Handler:

func logStatusMiddleware(inner http.Handler) http.Handler {
	return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
		http.AddServerTrace(r.Context(), &httptrace.ServerTrace{
			WroteHeader: func(info httptrace.WroteHeaderInfo) {
				log.Printf("path=%q status=%d", r.URL.Path, info.StatusCode)
			},
		})

		inner.ServeHTTP(w, r)
	})
}

[urandom]

If I'm reading this correctly, ServerTrace proposal doesn't allow modification of the response body prior to writing. Therefore it will not allow things like compressing the response body prior to it being sent, for example.

Presumably, the ServerTrace struct will need to define another handler: WillWriteBody(body []byte, header http.Header) []byte

Such a handler would need to receive the body before it's being written, as well as the headers, if there's a need to modify them. And it should return the new body for actual writing. I assume that the request headers are already somehow available from within that handler already, since they also usually play a part here.

[rhysh]

You make a good point @urandom , that packages like https://github.com/NYTimes/gziphandler are currently implemented by wrapping the ResponseWriter with special behavior for Write and Flush. If packages like that one continue to wrap the ResponseWriter, new features exposed as methods will continue to be shadowed.

That package also has special behavior for the Push method: it ensures that the synthetic request will have the "Accept-Encoding: gzip" header. And, the handler disables connection hijacking. It needs to intercept several methods in order to change their behavior.

Will the ServerTrace struct expand to allow modifying all additional ResponseWriter methods? When new behaviors are added to the ResponseWriter, will they be added to the ServerTrace struct as well? That seems like a lot of API duplication.

[bradfitz]

Looks like this didn't happen before the Go 1.10 freeze. Bumping to Go 1.11.

[stapelberg]

Another data point of a package which has to deal with the combinatorial explosion of optional interfaces:
https://github.com/prometheus/client_golang/blob/1cdba8fdde81f444626530cce66e138d2f68ae1c/prometheus/promhttp/delegator.go#L100

[keegancsmith]

I have an alternative proposal for avoiding the combinatorial explosion of optional interfaces. It is inspired by the causer interface from github.com/pkg/errors (interface{ Cause() error }). The stdlib documents that if you wrap a http.ResponseWriter you should also implement this interface

type Wrapper interface {
    Wrap() http.ResponseWriter
}

Then for example, if we want to get at the http.Hijacker we can use code like this:

func GetHijacker(w http.ResponseWriter) http.Hijacker {
    for w != nil {
        if h, ok := w.(http.Hijacker); ok {
            return h
        }
        wrapper, ok := w.(http.Wrapper)
        if !ok {
            break
        }
        w = wrapper.Wrap()
    }
    return nil
}

This doesn't solve the problem of getting at additional server event data that ServerTrace would, but would help with the original problem proposed in this issue and the issues with wrapping ResponseWriter in general.