net/http: scale rstAvoidanceDelay to reduce test flakiness
As far as I can tell, some flakiness is unavoidable in tests
that race a large client request write against a server's response
when the server doesn't read the full request.
It does not appear to be possible to simultaneously ensure that
well-behaved clients see EOF instead of ECONNRESET and also prevent
misbehaving clients from consuming arbitrary server resources.
(See RFC 7230 ยง6.6 for more detail.)
Since there doesn't appear to be a way to cleanly eliminate
this source of flakiness, we can instead work around it:
we can allow the test to adjust the hard-coded delay if it
sees a plausibly-related failure, so that the test can retry
with a longer delay.
As a nice side benefit, this also allows the tests to run more quickly
in the typical case: since the test will retry in case of spurious
failures, we can start with an aggressively short delay, and only back
off to a longer one if it is really needed on the specific machine
running the test.
Fixes #57084.
Fixes #51104.
For #58398.
Change-Id: Ia4050679f0777e5eeba7670307a77d93cfce856f
Cq-Include-Trybots: luci.golang.try:gotip-linux-amd64-longtest-race,gotip-linux-amd64-race,gotip-windows-amd64-race
Reviewed-on: https://go-review.googlesource.com/c/go/+/527196
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com> Reviewed-by: Damien Neil <dneil@google.com>
Auto-Submit: Bryan Mills <bcmills@google.com>