[sentry-junior]

TestServer::startTestServer() started php -S on the fixed port 45884 and then polled /ping in a while (true) loop. If the server couldn't bind or died on startup, the loop never ended. It only checked whether the process was alive after the loop, which it never reached. The CI job then hung until GitHub's 6h job limit, with no hint of which test was stuck. That's the likely cause of the Tests (ubuntu-latest, 8.3, lowest) hang on #2329, which stopped somewhere between tests 672 and 732. Every test in that range that starts the test server is in AgentClientBuilderTest or HttpClientTest.

This gives the test server the same treatment the test agent got in #2222:

• A free port is picked on every start instead of 45884. That port is inside Linux's ephemeral range, so other connections on the runner can occasionally be holding it.
• The readiness loop gives up after 5 seconds, and fails right away if php -S has exited. The error message includes the server's stderr, e.g. Failed to listen on 127.0.0.1:39315 (reason: Address already in use).
• On a failed start, the process is killed and closed before the exception is thrown, so tearDown() doesn't trip over a half-started server.

The server now binds to 127.0.0.1 instead of localhost, so it listens on the same address the free port was reserved on.

via Martin Linzmayer.

--

View Session [Sentry]