[MESOS-8468] `LAUNCH_GROUP` failure tears down the default executor.

• Type: Bug
• Status: Resolved
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: 1.2.0, 1.3.0, 1.4.0, 1.5.0
• Fix Version/s: 1.5.1, 1.6.0
• Labels: default-executor mesosphere
• Target Version/s: 1.6.0
• Epic Link: Default Executor improvements
• Sprint: Mesosphere Sprint 73, Mesosphere Sprint 74
• Story Points: 5
• Assignee: Gastón Kleiman
• Reporter: Chun-Hung Hsiao
• Shepherd: Qian Zhang
• Created: 19/Jan/18 21:59
• Updated: 14/Feb/18 12:51
• Resolved: 14/Feb/18 12:51

Description

The following code in the default executor (https://github.com/apache/mesos/blob/12be4ba002f2f5ff314fbc16af51d095b0d90e56/src/launcher/default_executor.cpp#L525-L535) shows that if a `LAUNCH_NESTED_CONTAINER` call is failed (say, due to a fetcher failure), the whole executor will be shut down:

// Check if we received a 200 OK response for all the
// `LAUNCH_NESTED_CONTAINER` calls. Shutdown the executor
// if this is not the case.
foreach (const Response& response, responses.get()) {
  if (response.code != process::http::Status::OK) {
    LOG(ERROR) << "Received '" << response.status << "' ("
               << response.body << ") while launching child container";
    _shutdown();
    return;
  }
}

This is not expected by a user. Instead, one would expect that a failed `LAUNCH_GROUP` won't affect other task groups launched by the same executor, similar to the case that a task failure only takes down its own task group. We should adjust the semantics to make a failed `LAUNCH_GROUP` not take down the executor and affect other task groups.