Flaky: a_handout_on_the_install_pass_listens_out_its_window_and_the_loop_reports_it can lose its ephemeral port#4121

opened23h agobyNNick
GitHub writeback is off. Issues and comments you change here stay in GitCafe. Sync settings

checkin::tests::a_handout_on_the_install_pass_listens_out_its_window_and_the_loop_reports_it failed once in a 200-iteration stress run of test(checkin), at core/crates/ployz/src/checkin/checkin_tests.rs:1095.

text
1panicked at crates/ployz/src/checkin/checkin_tests.rs:1095:5:
2assertion `left == right` failed
3 left: 0
4 right: 22

The test takes its port from a plain TcpListener::bind("127.0.0.1:0") at checkin_tests.rs:1067. That port is in the kernel's ephemeral range, so another test process can be handed the same number before the probe binds it. #4033's fix moves probe_tests.rs and ployzd's fake-cluster ports below the ephemeral range and reserves them across processes. This file's ports were left as they are.

On the #4033 branch it failed 1 in 200; on main the test passed 100 of 100 run alone, which can't tell the two apart. The likely fix is to take this port from the same reserved, below-ephemeral pool probe_tests.rs uses after #4033.

Scores per AGENTS.md: F 2 (seen once under stress, CI rerun clears it), S 2 (CI only, no user effect). F×S 4, so filed, not fixed in #4033's PR.