You were probably affected by #24138

You were probably affected by #24138

Ugh. Thanks. This is about (not) having timeouts on the client side, right? It's unfortunate that it affects brand new requests and that it permanently breaks all new requests.

Could we add a (when-broken-only?) connection re-use timeout without breaking the compatibility promise? It seems like connection re-use decisions are entirely up to net/http.

Any suggestions for workarounds? It's tricky since I am actually doing long polling requests, where I actively want to keep connections open for hours, so I don't want to set aggressive timeouts. Would it help to call http.DefaultClient.CloseIdleConnections() whenever an http request fails? Or start using a new http.Client whenever an http request fails?

A different culprit than lack of timeout might be lack of resume-after-suspend notification. That would let us ping the server on resume and reconnect if nec. Also discussed in #36141 (comment)

Maybe unrelated, but the CL for #23459 (net.Dial TCP keepalive on by default) didn't touch net/http. Does http.Client not use net.Dial, or is this disabled for http.Client by a different CL?

Given that we don't know whether 1.13 was affected, I suspect this should not be milestoned to 1.14 at this point.

@bcmills I milestoned based on superficial similarity to #34978, which has outstanding backports to 1.13. Feel free to re-milestone, although it might be better to understand a bit more first.

@josharian Did you enable http2 on the Server? If so, curious what happens with http/1.1.

I used net/http out of the box. I will run for a bit with http2 disabled and see what happens.

In 1.13 (and presumably earlier; not sure), http.DefaultTransport uses TCP keepalive. On client suspend/resume where the server drops the link (e.g. due to no traffic), subsequent client Write() disables TCP keepalive, preventing a timeout error from the last client Read() -- see #31490. I don't know whether this is preventing your http.DefaultClient from recovering, but thought I should mention it.

Setting a short keepalive helped in my case (a plain TCP/TLS context, not HTTP). Eventually I plan to disable keepalive and use a server pulse to detect a dead link.

From net/http/transport.go:

// DefaultTransport is the default implementation of Transport and is
// used by DefaultClient. It establishes network connections as needed
// and caches them for reuse by subsequent calls. It uses HTTP proxies
// as directed by the $HTTP_PROXY and $NO_PROXY (or $http_proxy and
// $no_proxy) environment variables.
var DefaultTransport RoundTripper = &Transport{
        Proxy: ProxyFromEnvironment,
        DialContext: (&net.Dialer{
                Timeout:   30 * time.Second,
                KeepAlive: 30 * time.Second,
                DualStack: true,
        }).DialContext,
        ForceAttemptHTTP2:     true,
        MaxIdleConns:          100,
        IdleConnTimeout:       90 * time.Second,
        TLSHandshakeTimeout:   10 * time.Second,
        ExpectContinueTimeout: 1 * time.Second,
}

@networkimprov, that seems unrelated.

I finally tracked this down.

It turns out that there was a code path in which in I didn't close resp.Body. Timeouts didn't help.

Although this was user error (sorry and thanks!), I do wonder whether there's a way to detect this, since it was not easy to find.

One option is a vet check that, similar to the lostcancel check, checks that resp.Body.Close gets called on all exits from a function. Another option would be some kind of check at runtime in the http package, although I don't know net/http well enough to know whether that is feasible.

Correction: timeouts were also necessary (in addition to properly closing resp.Body).

Closing as user error.

@josharian, could you provide some more detail about how this resulting from a missing resp.Body.Close?

The documentation for the Body field says:

The default HTTP client's Transport may not
reuse HTTP/1.x "keep-alive" TCP connections if the Body is
not read to completion and closed.

It does not say anything about what happens for HTTP/2 connections, so at the very least there seems to be a documentation issue here.

This also seems like something we could detect and report, using a finalizer-based approach similar to the one described in #24739: if the Body becomes unreachable without being closed, that always indicates a caller error, and could be reported as such (perhaps as a panic, or perhaps by some other means).

Yeah, I don't think forgetting to close resp.Body should cause anything worse than a TCP connection leak. It certainly shouldn't affect unrelated HTTP fetches.

I'm honestly not sure. I tried to reproduce using multiple simpler ssetup and failed. There was clearly something else going on as well that closing the Body helped with. I went looking to try to understand better, but I've already spent over a full day on this, and I'm not really inclined to spend more time digging.

I do think a finalizer-based unclosed unreachable Body detection would be a good idea, since that is definitely a bug, regardless of what precise consequences it has. I suspect it'd be better to log instead of panicking--net/http already logs when it notices an ineffectual WriteHeader call, so there's some precedent for it.

Possibly related: I just saw this error message in our CI system. I couldn't reproduce it, but I can't see why closing idle connections should cause this kind of error:

Delete "http://127.0.0.1:35617/api/v2/variables/020f755c3c082000": net/http: HTTP/1.x transport connection broken: http: CloseIdleConnections called

The above might have just been fixed via #41600 and #42026.

This makes so much sense.
I have two API calls, one has
c.Response().Header().Set("Connection", "keep-alive")
defer c.Request().Body.Close() // Added this one
and the other one
c.Response().Header().Set("Connection", "keep-alive")

Guess which one keeps giving me write: connection reset by peer?