// The purpose of this test is to ensure that when slaves are removed
// from the master, and then attempt to send status updates, we send
// a ShutdownMessage to the slave. Why? Because during a network
// partition, the master will remove a partitioned slave, thus sending
// its tasks to LOST. At this point, when the partition is removed,
// the slave may attempt to send updates if it was unaware that the
// master removed it. We've already notified frameworks that these
// tasks were LOST, so we have to have the slave shut down.
TEST_F(PartitionTest, PartitionedSlaveStatusUpdates)
{
master::Flags masterFlags = CreateMasterFlags();
Try<Owned<cluster::Master>> master = StartMaster(masterFlags);
ASSERT_SOME(master);
// Allow the master to PING the slave, but drop all PONG messages
// from the slave. Note that we don't match on the master / slave
// PIDs because it's actually the SlaveObserver Process that sends
// the pings.
Future<Message> ping = FUTURE_MESSAGE(
Eq(PingSlaveMessage().GetTypeName()), _, _);
DROP_PROTOBUFS(PongSlaveMessage(), _, _);
Future<SlaveRegisteredMessage> slaveRegisteredMessage =
FUTURE_PROTOBUF(SlaveRegisteredMessage(), _, _);
MockExecutor exec(DEFAULT_EXECUTOR_ID);
TestContainerizer containerizer(&exec);
Owned<MasterDetector> detector = master.get()->createDetector();
Try<Owned<cluster::Slave>> slave = StartSlave(detector.get(), &containerizer);
ASSERT_SOME(slave);
AWAIT_READY(slaveRegisteredMessage);
SlaveID slaveId = slaveRegisteredMessage.get().slave_id();
MockScheduler sched;
MesosSchedulerDriver driver(
&sched, DEFAULT_FRAMEWORK_INFO, master.get()->pid, DEFAULT_CREDENTIAL);
Future<FrameworkID> frameworkId;
EXPECT_CALL(sched, registered(&driver, _, _))
.WillOnce(FutureArg<1>(&frameworkId));
EXPECT_CALL(sched, resourceOffers(&driver, _))
.WillRepeatedly(Return());
driver.start();
AWAIT_READY(frameworkId);
// Drop the first shutdown message from the master (simulated
// partition), allow the second shutdown message to pass when
// the slave sends an update.
Future<ShutdownMessage> shutdownMessage =
DROP_PROTOBUF(ShutdownMessage(), _, slave.get()->pid);
EXPECT_CALL(sched, offerRescinded(&driver, _))
.WillRepeatedly(Return());
Future<Nothing> slaveLost;
EXPECT_CALL(sched, slaveLost(&driver, _))
.WillOnce(FutureSatisfy(&slaveLost));
Clock::pause();
// Now, induce a partition of the slave by having the master
// timeout the slave.
size_t pings = 0;
while (true) {
AWAIT_READY(ping);
pings++;
if (pings == masterFlags.max_slave_ping_timeouts) {
break;
}
ping = FUTURE_MESSAGE(Eq(PingSlaveMessage().GetTypeName()), _, _);
Clock::advance(masterFlags.slave_ping_timeout);
Clock::settle();
}
Clock::advance(masterFlags.slave_ping_timeout);
Clock::settle();
// Wait for the master to attempt to shut down the slave.
AWAIT_READY(shutdownMessage);
// The master will notify the framework that the slave was lost.
AWAIT_READY(slaveLost);
shutdownMessage = FUTURE_PROTOBUF(ShutdownMessage(), _, slave.get()->pid);
// At this point, the slave still thinks it's registered, so we
// simulate a status update coming from the slave.
TaskID taskId;
taskId.set_value("task_id");
const StatusUpdate& update = protobuf::createStatusUpdate(
frameworkId.get(),
slaveId,
taskId,
TASK_RUNNING,
TaskStatus::SOURCE_SLAVE,
UUID::random());
StatusUpdateMessage message;
message.mutable_update()->CopyFrom(update);
message.set_pid(stringify(slave.get()->pid));
process::post(master.get()->pid, message);
// The master should shutdown the slave upon receiving the update.
AWAIT_READY(shutdownMessage);
Clock::resume();
driver.stop();
driver.join();
}
// This test checks that a scheduler gets a slave lost
// message for a partitioned slave.
TEST_F(PartitionTest, PartitionedSlave)
{
master::Flags masterFlags = CreateMasterFlags();
Try<Owned<cluster::Master>> master = StartMaster(masterFlags);
ASSERT_SOME(master);
// Set these expectations up before we spawn the slave so that we
// don't miss the first PING.
Future<Message> ping = FUTURE_MESSAGE(
Eq(PingSlaveMessage().GetTypeName()), _, _);
// Drop all the PONGs to simulate slave partition.
DROP_PROTOBUFS(PongSlaveMessage(), _, _);
Owned<MasterDetector> detector = master.get()->createDetector();
Try<Owned<cluster::Slave>> slave = StartSlave(detector.get());
ASSERT_SOME(slave);
MockScheduler sched;
MesosSchedulerDriver driver(
&sched, DEFAULT_FRAMEWORK_INFO, master.get()->pid, DEFAULT_CREDENTIAL);
EXPECT_CALL(sched, registered(&driver, _, _));
Future<Nothing> resourceOffers;
EXPECT_CALL(sched, resourceOffers(&driver, _))
.WillOnce(FutureSatisfy(&resourceOffers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driver.start();
// Need to make sure the framework AND slave have registered with
// master. Waiting for resource offers should accomplish both.
AWAIT_READY(resourceOffers);
Clock::pause();
EXPECT_CALL(sched, offerRescinded(&driver, _))
.Times(AtMost(1));
Future<Nothing> slaveLost;
EXPECT_CALL(sched, slaveLost(&driver, _))
.WillOnce(FutureSatisfy(&slaveLost));
// Now advance through the PINGs.
size_t pings = 0;
while (true) {
AWAIT_READY(ping);
pings++;
if (pings == masterFlags.max_slave_ping_timeouts) {
break;
}
ping = FUTURE_MESSAGE(Eq(PingSlaveMessage().GetTypeName()), _, _);
Clock::advance(masterFlags.slave_ping_timeout);
}
Clock::advance(masterFlags.slave_ping_timeout);
AWAIT_READY(slaveLost);
slave.get()->terminate();
slave->reset();
JSON::Object stats = Metrics();
EXPECT_EQ(1, stats.values["master/slave_removals"]);
EXPECT_EQ(1, stats.values["master/slave_removals/reason_unhealthy"]);
driver.stop();
driver.join();
Clock::resume();
}
// This test checks that a scheduler gets a slave lost
// message for a partitioned slave.
TEST_F(PartitionTest, PartitionedSlave)
{
Try<PID<Master> > master = StartMaster();
ASSERT_SOME(master);
// Set these expectations up before we spawn the slave so that we
// don't miss the first PING.
Future<Message> ping = FUTURE_MESSAGE(Eq("PING"), _, _);
// Drop all the PONGs to simulate slave partition.
DROP_MESSAGES(Eq("PONG"), _, _);
Try<PID<Slave> > slave = StartSlave();
ASSERT_SOME(slave);
MockScheduler sched;
MesosSchedulerDriver driver(
&sched, DEFAULT_FRAMEWORK_INFO, master.get(), DEFAULT_CREDENTIAL);
EXPECT_CALL(sched, registered(&driver, _, _));
Future<Nothing> resourceOffers;
EXPECT_CALL(sched, resourceOffers(&driver, _))
.WillOnce(FutureSatisfy(&resourceOffers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driver.start();
// Need to make sure the framework AND slave have registered with
// master. Waiting for resource offers should accomplish both.
AWAIT_READY(resourceOffers);
Clock::pause();
EXPECT_CALL(sched, offerRescinded(&driver, _))
.Times(AtMost(1));
Future<Nothing> slaveLost;
EXPECT_CALL(sched, slaveLost(&driver, _))
.WillOnce(FutureSatisfy(&slaveLost));
// Now advance through the PINGs.
uint32_t pings = 0;
while (true) {
AWAIT_READY(ping);
pings++;
if (pings == master::MAX_SLAVE_PING_TIMEOUTS) {
break;
}
ping = FUTURE_MESSAGE(Eq("PING"), _, _);
Clock::advance(master::SLAVE_PING_TIMEOUT);
}
Clock::advance(master::SLAVE_PING_TIMEOUT);
AWAIT_READY(slaveLost);
this->Stop(slave.get());
JSON::Object stats = Metrics();
EXPECT_EQ(1, stats.values["master/slave_removals"]);
EXPECT_EQ(1, stats.values["master/slave_removals/reason_unhealthy"]);
driver.stop();
driver.join();
Shutdown();
Clock::resume();
}
