TestServer needs a IHttpResponseStart feature to avoid flushing

[JunTaoLuo]

To mimic the Kestrel behaviour, a IHttpResponseStartFeature needs to be added by the TestServer. Otherwise it falls back to the flushing behaviour: https://github.com/aspnet/AspNetCore/blob/4646ed56189a8ee36ced6397dd4f86fce9e6a8e8/src/Http/Http/src/Internal/DefaultHttpResponse.cs#L154-L159.

[jkotalik]

So a few questions:

• Do you agree that flushing if the StartFeature isn't implemented is desired behavior?
• If so, why do you need TestServer to implement the IHttpResponseStartFeature? Are you trying to verify that StartAsync doesn't flush headers to the client?

I agree that having inconsistent behavior based on if the feature is implemented isn't great. Maybe we should throw if the feature isn't implemented instead.

[JamesNK]

This area came up when when we upgraded gRPC to the latest version of preview 3. A combination of circumstances combined:

• We had heard that StartAsync is required before using Response.BodyPipe, and that calling it multiple times is fine because the feature will no-op once the response has started
• We changed our write message code to start with a call to StartAsync
• We have a unit test that does buffering and calls write message multiple times
• Because our test is running in TestServer there was no IHttpResponseStartFeature so writing a message would flush previous messages

The main problem here is the difference in behavior between a real server and the TestServer. The test server is flushing each time StartAsync is called. That could be solved by #7778

[Tratcher]

The fallback behavior is OK, but we should go ahead and implement IHttpResponseStartFeature and pipes. There's already a response body pipe.

[Tratcher]

#7815

[muratg]

Triage: We won't be doing this. Instead we'll do #7779

[JunTaoLuo]

Why are we not adding the feature to TestServer?

[JunTaoLuo]

I think I named the issue poorly and it was misunderstood? We should still add the feature to the test server.

[JamesNK]

Related: TestServer should be using real pipes instead of the stream adapters