The "test_rosbag2_record_end_to_end record_end_to_end_test_with_zstd_file_compression" is tend to be flaky

[MichaelOrlov]

From CI runs at Apex.AI, it was found that the rosbag2/rosbag2_tests:test_rosbag2_record_end_to_end record_end_to_end_test_with_zstd_file_compression/0 test is tends to be flaky.
Sometimes the tests fail with messages like:

rosbag2/rosbag2_tests/test/rosbag2_tests/test_rosbag2_record_end_to_end.cpp:88 Value of: pub_manager.wait_for_matched(topic_name) Actual: false Expected: true Expected find rosbag subscription

more detailed log:

2025-11-27T04:16:18.225672Z 01O [ RUN      ] TestRecordEndToEnd/RecordFixture.record_end_to_end_test_with_zstd_file_compression/0
2025-11-27T04:16:18.225676Z 01O stdin is not a terminal device. Keyboard handling disabled.[INFO] [1764216365.594788401] [rosbag2_recorder]: Event publisher thread: Started
2025-11-27T04:16:18.225681Z 01O [INFO] [1764216365.595704303] [rosbag2_recorder]: Press SPACE for pausing/resuming
2025-11-27T04:16:18.225685Z 01O [INFO] [1764216365.728459186] [rosbag2_recorder]: Starting recording to '/bb/exchange/worker-build/255/root/_tmp/6203efcd52a01311d4488813eb683adf/tmp_test_dir_7ee76b/record_end_to_end_test_with_zstd_file_compression_0_sqlite3'
2025-11-27T04:16:18.225690Z 01O [WARN] [1764216365.729871799] [rosbag2_recorder]: No input serialization format specified, using default rmw serialization format: 'cdr'.
2025-11-27T04:16:18.225697Z 01O [INFO] [1764216365.974652317] [rosbag2_storage]: Opened database '/bb/exchange/worker-build/255/root/_tmp/6203efcd52a01311d4488813eb683adf/tmp_test_dir_7ee76b/record_end_to_end_test_with_zstd_file_compression_0_sqlite3/record_end_to_end_test_with_zstd_file_compression_0_sqlite3_0.db3' for READ_WRITE.
2025-11-27T04:16:18.225701Z 01O [WARN] [1764216365.974771730] [rosbag2_cpp]: Environment variable ROS_DISTRO not set, can't store value in bag metadata.
2025-11-27T04:16:18.225706Z 01O grace/recording/rosbag2/rosbag2_tests/test/rosbag2_tests/test_rosbag2_record_end_to_end.cpp:88: Failure
2025-11-27T04:16:18.225710Z 01O Value of: pub_manager.wait_for_matched(topic_name)
2025-11-27T04:16:18.225711Z 01O   Actual: false
2025-11-27T04:16:18.225714Z 01O Expected: true
2025-11-27T04:16:18.225717Z 01O Expected find rosbag subscription
2025-11-27T04:16:18.225719Z 01O [INFO] [1764216366.808885262] [rosbag2_recorder]: Subscribed to topic '/test_topic'
2025-11-27T04:16:18.225729Z 01O [INFO] [1764216366.813860979] [rosbag2_recorder]: Listening for topics...
2025-11-27T04:16:18.225733Z 01O [INFO] [1764216366.813922721] [rosbag2_recorder]: Topics discovery started.
2025-11-27T04:16:18.225736Z 01O [INFO] [1764216366.813965362] [rosbag2_recorder]: All requested topics are subscribed. Stopping discovery...
2025-11-27T04:16:18.225739Z 01O [INFO] [1764216366.813974162] [rosbag2_recorder]: Topics discovery stopped.
2025-11-27T04:16:18.225747Z 01O [INFO] [1764216366.818906048] [rosbag2_recorder]: Recording...

[MichaelOrlov]

RCA

We failed by timeout waiting for matched publishers with subscription pub_manager.wait_for_matched(topic_name). The default timeout is 10 Seconds, which should be fairly enough.

Root Cause Analysis from the Codex gpt-5.4

The intermittent failure was caused by a race in test synchronization, not by zstd compression itself.

The failing test starts "'ros2 bag record and then waits for PublicationManager::wait_for_matched() to observe a subscriber on the publisher side. Recorder startup is asynchronous: the writer and storage are opened first, topic discovery starts afterward, and the actual subscription to /test_topic is created later in the recorder discovery thread.

The previous PublicationManager::wait_for_matched() implementation only sleep-polled get_subscription_count(). Under CI load or slower DDS graph propagation, the recorder could still be in topic discovery while the helper timed out waiting for the publisher-side match to become visible.

Proposed fix

Update PublicationManager::wait_for_matched() to use a hybrid wait strategy for all callers:

• wait on ROS graph events so discovery changes wake the helper promptly
• keep a short bounded periodic re-check so the helper still makes progress even if a useful graph wakeup does not arrive

This preserves the existing API while reducing recorder test flakiness across the suite.