[MESOS-9419] Executor to framework message crashes master if framework has not re-registered.

• Type: Bug
• Status: Resolved
• Priority: Blocker
• Resolution: Fixed
• Affects Version/s: 1.2.0, 1.2.1, 1.2.2, 1.2.3, 1.3.0, 1.3.1, 1.3.2, 1.4.0, 1.4.1, 1.4.2, 1.5.0, 1.5.1, 1.6.0, 1.6.1, 1.7.0
• Fix Version/s: 1.4.3, 1.5.2, 1.6.2, 1.7.1, 1.8.0
• Component/s: master
• Target Version/s: 1.5.2, 1.6.2, 1.7.1, 1.8.0
• Sprint: Storage R8 Sprint 34
• Story Points: 3
• Assignee: Chun-Hung Hsiao
• Reporter: Benjamin Mahler
• Shepherd: Benjamin Mahler
• Created: 26/Nov/18 23:21
• Updated: 23/Jan/19 22:29
• Resolved: 28/Nov/18 23:37

Description

If the executor sends a framework message after a master failover, and the framework has not yet re-registered with the master, this will crash the master:

W20181105 22:02:48.782819 172709 master.hpp:2304] Master attempted to send message to disconnected framework 03dc2603-acd6-491e-\ 8717-3f03e5ee37f4-0000 (Cook-1.24.0-9299b474217db499c9d28738050b359ac8dd55bb)
F20181105 22:02:48.782830 172709 master.hpp:2314] CHECK_SOME(pid): is NONE
*** Check failure stack trace: ***
*** @ 0x7f09e016b6cd google::LogMessage::Fail()
*** @ 0x7f09e016d38d google::LogMessage::SendToLog()
*** @ 0x7f09e016b2b3 google::LogMessage::Flush()
*** @ 0x7f09e016de09 google::LogMessageFatal::~LogMessageFatal()
*** @ 0x7f09df086228 _CheckFatal::~_CheckFatal()
*** @ 0x7f09df3a403d mesos::internal::master::Framework::send<>()
*** @ 0x7f09df2f4886 mesos::internal::master::Master::executorMessage()
*** @ 0x7f09df3b06a4 _ZN15ProtobufProcessIN5mesos8internal6master6MasterEE8handlerNINS1_26ExecutorToFrameworkMessageEJRKNS0\ _7SlaveIDERKNS0_11FrameworkIDERKNS0_10ExecutorIDERKNSt7__cxx1112basic_stringIcSt11char_traitsIcESaIcEEEEJS9_SC_SF_SN_EEEvPS3_MS3\ _FvRKN7process4UPIDEDpT1_ESS_SN_DpMT_KFT0_vE @ 0x7f09df345b43 std::_Function_handler<>::_M_invoke()
*** @ 0x7f09df36930f ProtobufProcess<>::consume()
*** @ 0x7f09df2e0ff5 mesos::internal::master::Master::_consume()
*** @ 0x7f09df2f5542 mesos::internal::master::Master::consume()
*** @ 0x7f09e00d9c7a process::ProcessManager::resume()
*** @ 0x7f09e00dd836 _ZNSt6thread5_ImplISt12_Bind_simpleIFZN7process14ProcessManager12init_threadsEvEUlvE_vEEE6_M_runEv
*** @ 0x7f09dd467ac8 execute_native_thread_routine
*** @ 0x7f09dd6f6b50 start_thread
*** @ 0x7f09dcc7030d (unknown)

This is because Framework::send proceeds if the framework is disconnected. In the case of a recovered framework, it will not have a pid or http connection yet:

https://github.com/apache/mesos/blob/9b889a10927b13510a1d02e7328925dba3438a0b/src/master/master.hpp#L2590-L2610

// Sends a message to the connected framework.
template <typename Message>
void Framework::send(const Message& message)
{
  if (!connected()) {
    LOG(WARNING) << "Master attempted to send message to disconnected"
                 << " framework " << *this;
    // XXX proceeds!
  }

  metrics.incrementEvent(message);

  if (http.isSome()) {
    if (!http->send(message)) {
      LOG(WARNING) << "Unable to send event to framework " << *this << ":"
                   << " connection closed";
    }
  } else {
    CHECK_SOME(pid); // XXX Will crash.
    master->send(pid.get(), message);
  }
}

The executor to framework path does not guard against the framework being disconnected, unlike the status update path:

https://github.com/apache/mesos/blob/9b889a10927b13510a1d02e7328925dba3438a0b/src/master/master.cpp#L6472-L6495

vs.

https://github.com/apache/mesos/blob/9b889a10927b13510a1d02e7328925dba3438a0b/src/master/master.cpp#L8371-L8373

It was reported that this crash didn't occur for the user on 1.2.0, however the issue appears to present there as well, so we will try to backport a test to see if it's indeed not occurring in 1.2.0.

Issue Links

• causes: MESOS-8623 Crashed framework brings down the whole Mesos cluster