Hi, I gave Santiago's patch a try on crystal 1.18.2+dfsg-1, and it works for me.
I could not use schroot on this machine, so I tested it with sbuild's unshare backend, with networking disabled. The full build completed successfully: compiler_spec: 13306 examples, 0 failures, 0 errors, 18 pending std_spec: 17491 examples, 0 failures, 0 errors, 28 pending primitives_spec: 717 examples, 0 failures, 0 errors, 0 pending For comparison, the unpatched std_spec ran 17563 examples, with 33 pending. The difference is exactly 72 examples and 5 pending examples, matching all the examples tagged "network" in the five socket spec files. The compiler and primitives counts did not change. One slightly confusing detail is that the unpatched build also passed with the unshare backend. This is because the resolver error from its isolated network namespace is accepted by the current tests. To check the actual network activity, I also ran those five socket spec files under strace in a network-enabled environment. Without the filter, strace recorded five connect(2) calls to DNS port 53 (15 system calls involving port 53 in total). With "--tag '~network'", no examples were selected and strace recorded no DNS or connect calls. So the patch appears to do exactly what we want: it skips only the 72 network-tagged examples, and the rest of the test suite still passes. As of today, crystal 1.18.2+dfsg-1 and its reverse dependency shards are still scheduled for autoremoval from testing on 22 August because of this bug. Could you please apply the patch and upload a fixed version before then? -- Best regards Cheng Li
