Re: [PATCH 0/16] http test bug potpourri

[Junio C Hamano <gitster@pobox.com>]

Jeff King <peff@peff.net> writes:

> So all I wanted to do was this one-liner:
>
> diff --git a/t/t5559-http-fetch-smart-http2.sh b/t/t5559-http-fetch-smart-http2.sh
> index 9eece71c2c..54aa9d3bff 100755
> --- a/t/t5559-http-fetch-smart-http2.sh
> +++ b/t/t5559-http-fetch-smart-http2.sh
> @@ -1,4 +1,5 @@
>  #!/bin/sh
>  
>  HTTP_PROTO=HTTP/2
> +LIB_HTTPD_SSL=1
>  . ./t5551-http-fetch-smart.sh
>
> but somehow I'm 16 patches deep. Let me back up.
>
> I got bit once again by the "oops, HTTP/2 tests in t5559 are sometimes
> flaky" bug. One thing that came up in earlier discussion is that HTTP/2
> over TLS should be much more reliable, because it doesn't have to go
> through the funny HTTP-upgrade path.
>
> Hence the patch above, which is also patch 16 here. And it does make the
> consistent failure of t5551.30 go away. And it even makes --stress work
> longer before a racy failure, though it still fails for me pretty
> consistently within a few dozen runs.
>
> But in doing so, I found out all sorts of neat things, like:
>
>   - when I tested with HTTP/2 and TLS before, I was accidentally not
>     using HTTP/2!
>
>   - we even have a test that should detect which version is used, but
>     it's a silent noop unless you set GIT_TEST_PROTOCOL_VERSION=0, which
>     clearly nobody does
>
>   - it turns out there are a bunch of tests which are skipped (some of
>     which even fail!) unless you set that variable
>
> So this series fixes the broken tests, adapts them to work with both v0
> and v2 Git protocol, makes them work with HTTP/2 when needed, sprinkles
> in a couple other fixes, and then finally does that one-liner.
>
> I'm actually not sure if the final patch is a good idea or not, but
> certainly all of the fixes leading up to it are worth doing.

Thanks; this must have been a lot of work.  From the "test what the
end users use, or at least something close to it" standpoint, 16/16
certainly is the right thing to do, I would think.