proxysql

History

Rene Cannao e9b78ec572 Fix reg_test_3847_admin_lock for isolated CI runtime layout The legacy version of reg_test_3847_admin_lock assumed the old TAP execution model where everything effectively lived on 127.0.0.1 behind docker port mappings. In the new legacy-g2 infrastructure the test process and the primary ProxySQL run in separate containers, while the test still launches a second local ProxySQL instance inside the test container. That split exposed several test-side assumptions that were no longer valid and caused the child replica setup to fail before the actual deadlock coverage even started. The most important issue was how the secondary ProxySQL was started. The test passed a static proxysql_sec.cfg, but it did not give the child process an isolated datadir. On the new CI the child could therefore discover and reuse an existing proxysql.db from the default runtime location, which meant the tracked config file was not actually authoritative. The test would then wait for a replica on 127.0.0.1:26081 that never came up with the expected credentials or cluster topology. This change makes the secondary runtime explicit and self-contained. The test now creates a private runtime directory under test/tap/tests/reg_test_3847_node_datadir/runtime, generates the child config dynamically for the current environment, and launches the child with -D pointing at that runtime directory. The generated config keeps the local child replica on 127.0.0.1:26081/36081 while wiring the primary cluster peer to the real TAP admin endpoint discovered from CommandLine. That removes the stale single-host assumption and makes the test match how the isolated container environment is actually wired. The fix also cleans up endpoint usage inside the test body. Connections to the primary ProxySQL admin interface now use cl.admin_host/cl.admin_port instead of the generic frontend host, while the worker that talks to the locally spawned child explicitly targets 127.0.0.1:26081. This matters in the new CI because the primary lives in a different container but the child lives next to the test process. There was also a correctness bug in the thread bookkeeping. The previous code could overwrite a real worker error with success, and failed worker connections did not stop execution immediately. The worker status flags and launcher status are now atomic, connect failures return immediately, main-loop checks only treat positive errno values as failures, and shutdown waits report the live state instead of silently racing past it. Since this class of failure is expensive to diagnose remotely, the TAP diagnostics are intentionally much more verbose now. The test logs the primary and replica endpoints it is using, the generated config path, runtime directory, launcher capture log, internal ProxySQL log path, worker startup messages, periodic progress from the two background loops, wait-loop heartbeats, and shutdown progress. That should make future CI failures explain themselves from the TAP log instead of requiring ad hoc reproduction. Finally, add a local .gitignore entry for reg_test_3847_node_datadir/runtime/. That directory is generated by the test and contains transient databases, logs, and certificates for the spawned child ProxySQL instance; it should not be committed.	1 month ago
..
.gitignore	Fix reg_test_3847_admin_lock for isolated CI runtime layout	1 month ago
proxysql_sec.cfg	Add new regression test for issue #3847	4 years ago

Rene Cannao e9b78ec572 Fix reg_test_3847_admin_lock for isolated CI runtime layout

The legacy version of reg_test_3847_admin_lock assumed the old TAP execution model where everything effectively lived on 127.0.0.1 behind docker port mappings. In the new legacy-g2 infrastructure the test process and the primary ProxySQL run in separate containers, while the test still launches a second local ProxySQL instance inside the test container. That split exposed several test-side assumptions that were no longer valid and caused the child replica setup to fail before the actual deadlock coverage even started.

The most important issue was how the secondary ProxySQL was started. The test passed a static proxysql_sec.cfg, but it did not give the child process an isolated datadir. On the new CI the child could therefore discover and reuse an existing proxysql.db from the default runtime location, which meant the tracked config file was not actually authoritative. The test would then wait for a replica on 127.0.0.1:26081 that never came up with the expected credentials or cluster topology.

This change makes the secondary runtime explicit and self-contained. The test now creates a private runtime directory under test/tap/tests/reg_test_3847_node_datadir/runtime, generates the child config dynamically for the current environment, and launches the child with -D pointing at that runtime directory. The generated config keeps the local child replica on 127.0.0.1:26081/36081 while wiring the primary cluster peer to the real TAP admin endpoint discovered from CommandLine. That removes the stale single-host assumption and makes the test match how the isolated container environment is actually wired.

The fix also cleans up endpoint usage inside the test body. Connections to the primary ProxySQL admin interface now use cl.admin_host/cl.admin_port instead of the generic frontend host, while the worker that talks to the locally spawned child explicitly targets 127.0.0.1:26081. This matters in the new CI because the primary lives in a different container but the child lives next to the test process.

There was also a correctness bug in the thread bookkeeping. The previous code could overwrite a real worker error with success, and failed worker connections did not stop execution immediately. The worker status flags and launcher status are now atomic, connect failures return immediately, main-loop checks only treat positive errno values as failures, and shutdown waits report the live state instead of silently racing past it.

Since this class of failure is expensive to diagnose remotely, the TAP diagnostics are intentionally much more verbose now. The test logs the primary and replica endpoints it is using, the generated config path, runtime directory, launcher capture log, internal ProxySQL log path, worker startup messages, periodic progress from the two background loops, wait-loop heartbeats, and shutdown progress. That should make future CI failures explain themselves from the TAP log instead of requiring ad hoc reproduction.

Finally, add a local .gitignore entry for reg_test_3847_node_datadir/runtime/. That directory is generated by the test and contains transient databases, logs, and certificates for the spawned child ProxySQL instance; it should not be committed.

.gitignore

Fix reg_test_3847_admin_lock for isolated CI runtime layout

proxysql_sec.cfg

Add new regression test for issue #3847