workaround for graceful_exit / wait_closed

[abe-winter]

What changed

• add new _single_stage kwarg in graceful_exit which runs first_ and second_stage in sequence, rather than waiting for a second signal
• new test

  Why

  We use grpclib in a scenario where https://github.com/vmagamedov/grpclib/issues/107 is a problem. We currently workaround by SIGTERM-ing twice. I'm hoping to add this kwarg so that we can trigger more aggressive shutdown behavior.

[vmagamedov]

Thank you for your effort but graceful exit with immediate second stage doesn't look like graceful exit, it is just exit :)

I tried to implement graceful exit without GOAWAY frame in #190. Here the first stage is more aggressive, should exit predictably and as soon as possible. And the only way to reach the second stage should be some slow cancellation logic inside a method handler.

In this implementation when graceful exit begins server starts to reject incoming requests on existing connections using REFUSED_STREAM error code while connection is not completely drained.

Can you install grpclib from graceful-exit branch and check if it works for you?

[abe-winter]

Thank you very much for looking at this.

I tried your graceful-exit branch as well as a few other versions (current master, 0.4.6, 0.4.7).

It looks like our specific shutdown issue was fixed between 0.4.6 and 0.4.7.

Looked deeper: removing server.py L739 causes 0.4.7 to behave like 0.4.6

(removing await _server_closed_fut in Server.wait_closed)

I should have tested on the latest release before submitting a fix -- I'm sorry about that.

Do you want me to do any more investigation? Or should I close this because it only affects an old version.

[abe-winter]

closing, guessing not needed