net/http: mention Context and Client.Do in docs for Get, Head, Post, PostForm

[rittneje]

The documentation for Client.Do says "Generally Get, Post, or PostForm will be used instead of Do." However, there is no way to pass a context.Context into those methods, so we cannot, for example, set a per-request timeout without resorting to Do. I propose adding the following new methods to http.Client.

• GetContext(ctx context.Context, url string) (resp *Response, err error)
• HeadContext(ctx context.Context, url string) (resp *Response, err error)
• PostContext(ctx context.Context, url, contentType string, body io.Reader) (resp *Response, err error)
• PostFormContext(ctx context.Context, url string, data url.Values) (resp *Response, err error)

This would effectively deprecate the existing Get, Head, Post, and PostForm methods.

We should also make a corresponding global function for each of these, and deprecate the existing http.Get, http.Head, http.Post, and http.PostForm functions.

[zikaeroh]

This doesn't resolve the proposal, but you might not be aware of ctxhttp in x/net: https://godoc.org/golang.org/x/net/context/ctxhttp

[odeke-em]

Hello @rittneje! Thank you for filing this issue. I see how this is inconvenient. However, history beat us to the punch, those functions were added years before context.Context was a thing. I am sympathetic towards adding Context on them. On the flip side though:

The highlighted functions are helper/convenience methods on the global/default HTTPClient.
To customize more things on those methods, we recommend taking a look at NewRequest et al for example https://golang.org/pkg/net/http/#Post

The already existing customizers are:

• NewRequestWithContext
• Request.WithContext

but perhaps we should add examples enumerating how to make the various requests corresponding to those highlighted functions showing how to accomplish them but with customization and NewRequestWithContext, but also add the notice to use NewRequestWithContext for better customization.

Adding more helper functions with a new flavor would be increasing the API surface for things that we already recommend alternatives to.

/cc @bradfitz

[rittneje]

@odeke-em It does kind of stink that there's no real convenience method for issuing a GET request anymore without dropping the context on the floor. I don't really consider that to be a customization at all, since dropping context is a code smell. In any case, if you do decide not to add the missing helpers, you should probably remove that statement from the documentation of Client.Do since it is no longer true.

[rsc]

I was writing some HTTP request code earlier today and I had to use http.NewRequest and client.Do because I had to set some headers. It's not a huge deal to have to do that. The convenience functions don't have to provide for every possible convenience, or else they stop being convenient.

[rittneje]

@rsc I agree they don't have to provide for every possible convenience, but the point is that the old convenience methods really should never be used anymore outside toy example code. So either new versions should be added to accept a context (to not encourage developers to drop the incoming context), or the comment on Client.Do should be revised.

[alercah]

I just was reviewing code today that was omitting contexts where it shouldn't have been, and these helper methods I imagine were a big factor in this. I would very much like to see this approved as it will save a lot of time and fix logic errors.

perhaps we should add examples enumerating how to make the various requests corresponding to those highlighted functions showing how to accomplish them but with customization and NewRequestWithContext, but also add the notice to use NewRequestWithContext for better customization.

I agree the documentation could be a little better, but I think that this characterization of contexts as a customization is wrong. In modern idiomatic Go, contexts are not customization. They're bread-and-butter mandatory parts of APIs.

the point is that the old convenience methods really should never be used anymore outside toy example code.

This hits the nail on the head. And the existing API to the package strongly implies that you want to use them, or else why are they there? It sucks to tell someone in code review "No, sorry, you can't use this convenient one-liner, you have to use a more complicated function instead just to add the mandatory context parameter."

[bradfitz]

Related: see #23707 for designing a new HTTP client package.

[rsc]

I think there is plenty of short-lived command-line use for which these functions are just fine. I do agree that in a long-running program or server or a library that might be used in one of those, the helpers are not appropriate. Let's just document that more clearly. @bradfitz volunteered to write a CL.

[rsc]

Retitled to match previous discussion about updating docs.
We can leave the more general API fixup as a separate proposal (likely based on #23707).

With the rescoping, this seems like a likely accept.

[alercah]

This is incredibly disappointing.

…

On Wed, 26 Feb 2020 at 13:12, Russ Cox ***@***.***> wrote: Retitled to match previous discussion about updating docs. We can leave the more general API fixup as a separate proposal (likely based on #23707 <#23707>). With the rescoping, this seems like a *likely accept*. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#35562?email_source=notifications&email_token=AE7AOVNAAO7BCZXRTF5AIS3RE2WKVA5CNFSM4JM664G2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOENBJHSI#issuecomment-591565769>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AE7AOVKZLPQVXQ2OGFZLJ73RE2WKVANCNFSM4JM664GQ> .

[cristaloleg]

@alercah I suspect #23707 will happen in upcoming years, so the context will be used as it should be.

[rsc]

No change in consensus, so accepting.