TEST_F(ResourceOffersTest, ResourcesGetReofferedWhenUnused)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
Owned<MasterDetector> detector = master.get()->createDetector();
Try<Owned<cluster::Slave>> slave = StartSlave(detector.get());
ASSERT_SOME(slave);
MockScheduler sched1;
MesosSchedulerDriver driver1(
&sched1, DEFAULT_FRAMEWORK_INFO, master.get()->pid, DEFAULT_CREDENTIAL);
EXPECT_CALL(sched1, registered(&driver1, _, _));
Future<vector<Offer>> offers;
EXPECT_CALL(sched1, resourceOffers(&driver1, _))
.WillOnce(FutureArg<1>(&offers));
driver1.start();
AWAIT_READY(offers);
ASSERT_FALSE(offers->empty());
vector<TaskInfo> tasks; // Use nothing!
driver1.launchTasks(offers.get()[0].id(), tasks);
MockScheduler sched2;
MesosSchedulerDriver driver2(
&sched2, DEFAULT_FRAMEWORK_INFO, master.get()->pid, DEFAULT_CREDENTIAL);
EXPECT_CALL(sched2, registered(&driver2, _, _));
EXPECT_CALL(sched2, resourceOffers(&driver2, _))
.WillOnce(FutureArg<1>(&offers));
driver2.start();
AWAIT_READY(offers);
// Stop first framework before second so no offers are sent.
driver1.stop();
driver1.join();
driver2.stop();
driver2.join();
}
TEST_F(ResourceOffersTest, ResourcesGetReofferedAfterFrameworkStops)
{
Try<PID<Master>> master = StartMaster();
ASSERT_SOME(master);
Try<PID<Slave>> slave = StartSlave();
ASSERT_SOME(slave);
MockScheduler sched1;
MesosSchedulerDriver driver1(
&sched1, DEFAULT_FRAMEWORK_INFO, master.get(), DEFAULT_CREDENTIAL);
EXPECT_CALL(sched1, registered(&driver1, _, _))
.Times(1);
Future<vector<Offer>> offers;
EXPECT_CALL(sched1, resourceOffers(&driver1, _))
.WillOnce(FutureArg<1>(&offers));
driver1.start();
AWAIT_READY(offers);
EXPECT_NE(0u, offers.get().size());
driver1.stop();
driver1.join();
MockScheduler sched2;
MesosSchedulerDriver driver2(
&sched2, DEFAULT_FRAMEWORK_INFO, master.get(), DEFAULT_CREDENTIAL);
EXPECT_CALL(sched2, registered(&driver2, _, _))
.Times(1);
EXPECT_CALL(sched2, resourceOffers(&driver2, _))
.WillOnce(FutureArg<1>(&offers));
driver2.start();
AWAIT_READY(offers);
driver2.stop();
driver2.join();
Shutdown();
}
// This is a simple end to end test that makes sure a master using log
// storage with ZooKeeper can successfully launch a task.
TEST_F(RegistrarZooKeeperTest, TaskRunning)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
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);
MockScheduler sched;
MesosSchedulerDriver driver(
&sched, DEFAULT_FRAMEWORK_INFO, master.get()->pid, DEFAULT_CREDENTIAL);
EXPECT_CALL(sched, registered(&driver, _, _));
Future<vector<Offer>> offers;
EXPECT_CALL(sched, resourceOffers(&driver, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driver.start();
AWAIT_READY(offers);
EXPECT_NE(0u, offers->size());
TaskInfo task = createTask(offers.get()[0], "dummy", DEFAULT_EXECUTOR_ID);
EXPECT_CALL(exec, registered(_, _, _, _));
EXPECT_CALL(exec, launchTask(_, _))
.WillOnce(SendStatusUpdateFromTask(TASK_RUNNING));
Future<Nothing> resourcesUpdated;
EXPECT_CALL(containerizer,
update(_, Resources(offers.get()[0].resources())))
.WillOnce(DoAll(FutureSatisfy(&resourcesUpdated),
Return(Nothing())));
Future<TaskStatus> status;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&status));
driver.launchTasks(offers.get()[0].id(), {task});
AWAIT_READY(status);
EXPECT_EQ(TASK_RUNNING, status->state());
AWAIT_READY(resourcesUpdated);
EXPECT_CALL(exec, shutdown(_))
.Times(AtMost(1));
driver.stop();
driver.join();
}
// This test ensures that the command executor does not send
// TASK_KILLING to frameworks that do not support the capability.
TEST_P_TEMP_DISABLED_ON_WINDOWS(CommandExecutorTest, NoTaskKillingCapability)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
Owned<MasterDetector> detector = master.get()->createDetector();
slave::Flags flags = CreateSlaveFlags();
flags.http_command_executor = GetParam();
Try<Owned<cluster::Slave>> slave = StartSlave(detector.get(), flags);
ASSERT_SOME(slave);
// Start the framework without the task killing capability.
MockScheduler sched;
MesosSchedulerDriver driver(
&sched, DEFAULT_FRAMEWORK_INFO, master.get()->pid, DEFAULT_CREDENTIAL);
EXPECT_CALL(sched, registered(&driver, _, _));
Future<vector<Offer>> offers;
EXPECT_CALL(sched, resourceOffers(&driver, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driver.start();
AWAIT_READY(offers);
EXPECT_EQ(1u, offers->size());
// Launch a task with the command executor.
TaskInfo task = createTask(
offers->front().slave_id(),
offers->front().resources(),
"sleep 1000");
Future<TaskStatus> statusRunning;
EXPECT_CALL(sched, statusUpdate(_, _))
.WillOnce(FutureArg<1>(&statusRunning));
driver.launchTasks(offers->front().id(), {task});
AWAIT_READY(statusRunning);
EXPECT_EQ(TASK_RUNNING, statusRunning->state());
// There should only be a TASK_KILLED update.
Future<TaskStatus> statusKilled;
EXPECT_CALL(sched, statusUpdate(_, _))
.WillOnce(FutureArg<1>(&statusKilled));
driver.killTask(task.task_id());
AWAIT_READY(statusKilled);
EXPECT_EQ(TASK_KILLED, statusKilled->state());
driver.stop();
driver.join();
}
// This test verifies that when the scheduler calls stop() before
// abort(), no pending acknowledgements are sent.
TEST_F(MesosSchedulerDriverTest, DropAckIfStopCalledBeforeAbort)
{
Try<PID<Master>> master = StartMaster();
ASSERT_SOME(master);
MockExecutor exec(DEFAULT_EXECUTOR_ID);
TestContainerizer containerizer(&exec);
Try<PID<Slave>> slave = StartSlave(&containerizer);
ASSERT_SOME(slave);
MockScheduler sched;
MesosSchedulerDriver driver(
&sched, DEFAULT_FRAMEWORK_INFO, master.get(), DEFAULT_CREDENTIAL);
EXPECT_CALL(sched, registered(&driver, _, _));
EXPECT_CALL(sched, resourceOffers(&driver, _))
.WillOnce(LaunchTasks(DEFAULT_EXECUTOR_INFO, 1, 1, 16, "*"))
.WillRepeatedly(Return()); // Ignore subsequent offers.
// When an update is received, stop the driver and then abort it.
Future<Nothing> statusUpdate;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(DoAll(StopAndAbort(),
FutureSatisfy(&statusUpdate)));
// Ensure no status update acknowledgements are sent from the driver
// to the master.
EXPECT_NO_FUTURE_CALLS(
mesos::scheduler::Call(),
mesos::scheduler::Call::ACKNOWLEDGE,
_ ,
master.get());
EXPECT_CALL(exec, registered(_, _, _, _));
EXPECT_CALL(exec, launchTask(_, _))
.WillOnce(SendStatusUpdateFromTask(TASK_RUNNING));
EXPECT_CALL(exec, shutdown(_))
.Times(AtMost(1));
driver.start();
AWAIT_READY(statusUpdate);
// Settle the clock to ensure driver finishes processing the status
// update and sends acknowledgement if necessary. In this test it
// shouldn't send an acknowledgement.
Clock::pause();
Clock::settle();
driver.stop();
driver.join();
Shutdown();
}
TEST_F(ResourceOffersTest, ResourcesGetReofferedAfterFrameworkStops)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
Owned<MasterDetector> detector = master.get()->createDetector();
Try<Owned<cluster::Slave>> slave = StartSlave(detector.get());
ASSERT_SOME(slave);
MockScheduler sched1;
MesosSchedulerDriver driver1(
&sched1, DEFAULT_FRAMEWORK_INFO, master.get()->pid, DEFAULT_CREDENTIAL);
EXPECT_CALL(sched1, registered(&driver1, _, _));
Future<vector<Offer>> offers;
EXPECT_CALL(sched1, resourceOffers(&driver1, _))
.WillOnce(FutureArg<1>(&offers));
driver1.start();
AWAIT_READY(offers);
EXPECT_FALSE(offers->empty());
driver1.stop();
driver1.join();
MockScheduler sched2;
MesosSchedulerDriver driver2(
&sched2, DEFAULT_FRAMEWORK_INFO, master.get()->pid, DEFAULT_CREDENTIAL);
EXPECT_CALL(sched2, registered(&driver2, _, _));
EXPECT_CALL(sched2, resourceOffers(&driver2, _))
.WillOnce(FutureArg<1>(&offers));
driver2.start();
AWAIT_READY(offers);
driver2.stop();
driver2.join();
}
TEST_F_TEMP_DISABLED_ON_WINDOWS(
ResourceOffersTest,
ResourceOfferWithMultipleSlaves)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
Owned<MasterDetector> detector = master.get()->createDetector();
vector<Owned<cluster::Slave>> slaves;
// Start 10 slaves.
for (int i = 0; i < 10; i++) {
slave::Flags flags = CreateSlaveFlags();
flags.launcher = "posix";
flags.resources = Option<std::string>("cpus:2;mem:1024");
Try<Owned<cluster::Slave>> slave = StartSlave(detector.get(), flags);
ASSERT_SOME(slave);
slaves.push_back(slave.get());
}
MockScheduler sched;
MesosSchedulerDriver driver(
&sched, DEFAULT_FRAMEWORK_INFO, master.get()->pid, DEFAULT_CREDENTIAL);
EXPECT_CALL(sched, registered(&driver, _, _));
Future<vector<Offer>> offers;
EXPECT_CALL(sched, resourceOffers(&driver, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // All 10 slaves might not be in first offer.
driver.start();
AWAIT_READY(offers);
ASSERT_FALSE(offers->empty());
EXPECT_GE(10u, offers->size());
Resources resources(offers.get()[0].resources());
EXPECT_EQ(2, resources.get<Value::Scalar>("cpus")->value());
EXPECT_EQ(1024, resources.get<Value::Scalar>("mem")->value());
driver.stop();
driver.join();
}
TEST_F(ResourceOffersTest, ResourceOfferWithMultipleSlaves)
{
Try<PID<Master>> master = StartMaster();
ASSERT_SOME(master);
// Start 10 slaves.
for (int i = 0; i < 10; i++) {
slave::Flags flags = CreateSlaveFlags();
flags.resources = Option<std::string>("cpus:2;mem:1024");
Try<PID<Slave>> slave = StartSlave(flags);
ASSERT_SOME(slave);
}
MockScheduler sched;
MesosSchedulerDriver driver(
&sched, DEFAULT_FRAMEWORK_INFO, master.get(), DEFAULT_CREDENTIAL);
EXPECT_CALL(sched, registered(&driver, _, _))
.Times(1);
Future<vector<Offer>> offers;
EXPECT_CALL(sched, resourceOffers(&driver, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // All 10 slaves might not be in first offer.
driver.start();
AWAIT_READY(offers);
EXPECT_NE(0u, offers.get().size());
EXPECT_GE(10u, offers.get().size());
Resources resources(offers.get()[0].resources());
EXPECT_EQ(2, resources.get<Value::Scalar>("cpus").get().value());
EXPECT_EQ(1024, resources.get<Value::Scalar>("mem").get().value());
driver.stop();
driver.join();
Shutdown();
}
// 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();
}
// The purpose of this test is to ensure that when slaves are removed
// from the master, and then attempt to re-register, we deny the
// re-registration by sending 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 will attempt to
// re-register with its running tasks. We've already notified
// frameworks that these tasks were LOST, so we have to have the slave
// slave shut down.
TEST_F(PartitionTest, PartitionedSlaveReregistration)
{
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(), _, _);
MockExecutor exec(DEFAULT_EXECUTOR_ID);
TestContainerizer containerizer(&exec);
StandaloneMasterDetector detector(master.get()->pid);
Try<Owned<cluster::Slave>> slave = StartSlave(&detector, &containerizer);
ASSERT_SOME(slave);
MockScheduler sched;
MesosSchedulerDriver driver(
&sched, DEFAULT_FRAMEWORK_INFO, master.get()->pid, DEFAULT_CREDENTIAL);
EXPECT_CALL(sched, registered(&driver, _, _));
Future<vector<Offer>> offers;
EXPECT_CALL(sched, resourceOffers(&driver, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return());
driver.start();
AWAIT_READY(offers);
ASSERT_NE(0u, offers.get().size());
// Launch a task. This is to ensure the task is killed by the slave,
// during shutdown.
TaskID taskId;
taskId.set_value("1");
TaskInfo task;
task.set_name("");
task.mutable_task_id()->MergeFrom(taskId);
task.mutable_slave_id()->MergeFrom(offers.get()[0].slave_id());
task.mutable_resources()->MergeFrom(offers.get()[0].resources());
task.mutable_executor()->MergeFrom(DEFAULT_EXECUTOR_INFO);
task.mutable_executor()->mutable_command()->set_value("sleep 60");
// Set up the expectations for launching the task.
EXPECT_CALL(exec, registered(_, _, _, _));
EXPECT_CALL(exec, launchTask(_, _))
.WillOnce(SendStatusUpdateFromTask(TASK_RUNNING));
Future<TaskStatus> runningStatus;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&runningStatus));
Future<Nothing> statusUpdateAck = FUTURE_DISPATCH(
slave.get()->pid, &Slave::_statusUpdateAcknowledgement);
driver.launchTasks(offers.get()[0].id(), {task});
AWAIT_READY(runningStatus);
EXPECT_EQ(TASK_RUNNING, runningStatus.get().state());
// Wait for the slave to have handled the acknowledgment prior
// to pausing the clock.
AWAIT_READY(statusUpdateAck);
// Drop the first shutdown message from the master (simulated
// partition), allow the second shutdown message to pass when
// the slave re-registers.
Future<ShutdownMessage> shutdownMessage =
DROP_PROTOBUF(ShutdownMessage(), _, slave.get()->pid);
Future<TaskStatus> lostStatus;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&lostStatus));
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();
// The master will have notified the framework of the lost task.
AWAIT_READY(lostStatus);
EXPECT_EQ(TASK_LOST, lostStatus.get().state());
// 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);
Clock::resume();
// We now complete the partition on the slave side as well. This
// is done by simulating a master loss event which would normally
// occur during a network partition.
detector.appoint(None());
Future<Nothing> shutdown;
EXPECT_CALL(exec, shutdown(_))
.WillOnce(FutureSatisfy(&shutdown));
shutdownMessage = FUTURE_PROTOBUF(ShutdownMessage(), _, slave.get()->pid);
// Have the slave re-register with the master.
detector.appoint(master.get()->pid);
// Upon re-registration, the master will shutdown the slave.
// The slave will then shut down the executor.
AWAIT_READY(shutdownMessage);
AWAIT_READY(shutdown);
driver.stop();
driver.join();
}
// This is an end-to-end test that verfies that the slave returns the
// correct ResourceUsage based on the currently running executors, and
// the values get from the statistics endpoint are as expected.
TEST_F(MonitorIntegrationTest, RunningExecutor)
{
Try<PID<Master>> master = StartMaster();
ASSERT_SOME(master);
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<vector<Offer>> offers;
EXPECT_CALL(sched, resourceOffers(&driver, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driver.start();
AWAIT_READY(offers);
EXPECT_FALSE(offers.get().empty());
const Offer& offer = offers.get()[0];
// Launch a task and wait until it is in RUNNING status.
TaskInfo task = createTask(
offer.slave_id(),
Resources::parse("cpus:1;mem:32").get(),
"sleep 1000");
Future<TaskStatus> status;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&status));
driver.launchTasks(offer.id(), {task});
AWAIT_READY(status);
EXPECT_EQ(task.task_id(), status.get().task_id());
EXPECT_EQ(TASK_RUNNING, status.get().state());
// Hit the statistics endpoint and expect the response contains the
// resource statistics for the running container.
UPID upid("monitor", process::address());
Future<http::Response> response = http::get(upid, "statistics");
AWAIT_READY(response);
AWAIT_EXPECT_RESPONSE_STATUS_EQ(http::OK().status, response);
AWAIT_EXPECT_RESPONSE_HEADER_EQ(
"application/json",
"Content-Type",
response);
// Verify that the statistics in the response contains the proper
// resource limits for the container.
Try<JSON::Value> value = JSON::parse(response.get().body);
ASSERT_SOME(value);
Try<JSON::Value> expected = JSON::parse(strings::format(
"[{"
"\"statistics\":{"
"\"cpus_limit\":%g,"
"\"mem_limit_bytes\":%lu"
"}"
"}]",
1 + slave::DEFAULT_EXECUTOR_CPUS,
(Megabytes(32) + slave::DEFAULT_EXECUTOR_MEM).bytes()).get());
ASSERT_SOME(expected);
EXPECT_TRUE(value.get().contains(expected.get()));
driver.stop();
driver.join();
Shutdown();
}
TEST_F(MemoryPressureMesosTest, CGROUPS_ROOT_Statistics)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
slave::Flags flags = CreateSlaveFlags();
// We only care about memory cgroup for this test.
flags.isolation = "cgroups/mem";
flags.agent_subsystems = None();
Fetcher fetcher;
Try<MesosContainerizer*> _containerizer =
MesosContainerizer::create(flags, true, &fetcher);
ASSERT_SOME(_containerizer);
Owned<MesosContainerizer> containerizer(_containerizer.get());
Owned<MasterDetector> detector = master.get()->createDetector();
Try<Owned<cluster::Slave>> slave =
StartSlave(detector.get(), containerizer.get(), flags);
ASSERT_SOME(slave);
MockScheduler sched;
MesosSchedulerDriver driver(
&sched, DEFAULT_FRAMEWORK_INFO, master.get()->pid, DEFAULT_CREDENTIAL);
EXPECT_CALL(sched, registered(_, _, _));
Future<vector<Offer>> offers;
EXPECT_CALL(sched, resourceOffers(_, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driver.start();
AWAIT_READY(offers);
EXPECT_NE(0u, offers.get().size());
Offer offer = offers.get()[0];
// Run a task that triggers memory pressure event. We request 1G
// disk because we are going to write a 512 MB file repeatedly.
TaskInfo task = createTask(
offer.slave_id(),
Resources::parse("cpus:1;mem:256;disk:1024").get(),
"while true; do dd count=512 bs=1M if=/dev/zero of=./temp; done");
Future<TaskStatus> running;
Future<TaskStatus> killed;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&running))
.WillOnce(FutureArg<1>(&killed))
.WillRepeatedly(Return()); // Ignore subsequent updates.
driver.launchTasks(offer.id(), {task});
AWAIT_READY(running);
EXPECT_EQ(task.task_id(), running.get().task_id());
EXPECT_EQ(TASK_RUNNING, running.get().state());
Future<hashset<ContainerID>> containers = containerizer->containers();
AWAIT_READY(containers);
ASSERT_EQ(1u, containers.get().size());
ContainerID containerId = *(containers.get().begin());
// Wait a while for some memory pressure events to occur.
Duration waited = Duration::zero();
do {
Future<ResourceStatistics> usage = containerizer->usage(containerId);
AWAIT_READY(usage);
if (usage.get().mem_low_pressure_counter() > 0) {
// We will check the correctness of the memory pressure counters
// later, because the memory-hammering task is still active
// and potentially incrementing these counters.
break;
}
os::sleep(Milliseconds(100));
waited += Milliseconds(100);
} while (waited < Seconds(5));
EXPECT_LE(waited, Seconds(5));
// Pause the clock to ensure that the reaper doesn't reap the exited
// command executor and inform the containerizer/slave.
Clock::pause();
Clock::settle();
// Stop the memory-hammering task.
driver.killTask(task.task_id());
AWAIT_READY_FOR(killed, Seconds(120));
EXPECT_EQ(task.task_id(), killed->task_id());
EXPECT_EQ(TASK_KILLED, killed->state());
// Now check the correctness of the memory pressure counters.
Future<ResourceStatistics> usage = containerizer->usage(containerId);
AWAIT_READY(usage);
EXPECT_GE(usage.get().mem_low_pressure_counter(),
usage.get().mem_medium_pressure_counter());
EXPECT_GE(usage.get().mem_medium_pressure_counter(),
usage.get().mem_critical_pressure_counter());
Clock::resume();
driver.stop();
driver.join();
}
// This test ensures that the command executor sends TASK_KILLING
// to frameworks that support the capability.
TEST_F(CommandExecutorTest, TaskKillingCapability)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
Owned<MasterDetector> detector = master.get()->createDetector();
Try<Owned<cluster::Slave>> slave = StartSlave(detector.get());
ASSERT_SOME(slave);
// Start the framework with the task killing capability.
FrameworkInfo::Capability capability;
capability.set_type(FrameworkInfo::Capability::TASK_KILLING_STATE);
FrameworkInfo frameworkInfo = DEFAULT_FRAMEWORK_INFO;
frameworkInfo.add_capabilities()->CopyFrom(capability);
MockScheduler sched;
MesosSchedulerDriver driver(
&sched, frameworkInfo, master.get()->pid, DEFAULT_CREDENTIAL);
EXPECT_CALL(sched, registered(&driver, _, _));
Future<vector<Offer>> offers;
EXPECT_CALL(sched, resourceOffers(&driver, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driver.start();
AWAIT_READY(offers);
EXPECT_EQ(1u, offers->size());
// Launch a task with the command executor.
TaskInfo task = createTask(
offers->front().slave_id(),
offers->front().resources(),
"sleep 1000");
Future<TaskStatus> statusRunning;
EXPECT_CALL(sched, statusUpdate(_, _))
.WillOnce(FutureArg<1>(&statusRunning));
driver.launchTasks(offers->front().id(), {task});
AWAIT_READY(statusRunning);
EXPECT_EQ(TASK_RUNNING, statusRunning->state());
Future<TaskStatus> statusKilling, statusKilled;
EXPECT_CALL(sched, statusUpdate(_, _))
.WillOnce(FutureArg<1>(&statusKilling))
.WillOnce(FutureArg<1>(&statusKilled));
driver.killTask(task.task_id());
AWAIT_READY(statusKilling);
EXPECT_EQ(TASK_KILLING, statusKilling->state());
AWAIT_READY(statusKilled);
EXPECT_EQ(TASK_KILLED, statusKilled->state());
driver.stop();
driver.join();
}
// This test verifies if the executor is able to receive a Subscribed
// event in response to a Subscribe call request.
TEST_P(ExecutorHttpApiTest, Subscribe)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
ExecutorID executorId = DEFAULT_EXECUTOR_ID;
MockExecutor exec(executorId);
TestContainerizer containerizer(&exec);
Owned<MasterDetector> detector = master.get()->createDetector();
Try<Owned<cluster::Slave>> slave = StartSlave(detector.get(), &containerizer);
ASSERT_SOME(slave);
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));
Future<vector<Offer>> offers;
EXPECT_CALL(sched, resourceOffers(&driver, _))
.WillOnce(FutureArg<1>(&offers));
driver.start();
AWAIT_READY(frameworkId);
AWAIT_READY(offers);
ASSERT_EQ(1u, offers.get().size());
Future<Message> registerExecutorMessage =
DROP_MESSAGE(Eq(RegisterExecutorMessage().GetTypeName()), _, _);
TaskInfo taskInfo = createTask(offers.get()[0], "", executorId);
driver.launchTasks(offers.get()[0].id(), {taskInfo});
// Drop the `RegisterExecutorMessage` and then send a `Subscribe` request
// from the HTTP based executor.
AWAIT_READY(registerExecutorMessage);
Call call;
call.mutable_framework_id()->CopyFrom(evolve(frameworkId.get()));
call.mutable_executor_id()->CopyFrom(evolve(executorId));
call.set_type(Call::SUBSCRIBE);
call.mutable_subscribe();
// Retrieve the parameter passed as content type to this test.
const ContentType contentType = GetParam();
const string contentTypeString = stringify(contentType);
process::http::Headers headers;
headers["Accept"] = contentTypeString;
Future<Response> response = process::http::streaming::post(
slave.get()->pid,
"api/v1/executor",
headers,
serialize(contentType, call),
contentTypeString);
AWAIT_EXPECT_RESPONSE_STATUS_EQ(OK().status, response);
AWAIT_EXPECT_RESPONSE_HEADER_EQ("chunked", "Transfer-Encoding", response);
AWAIT_EXPECT_RESPONSE_HEADER_EQ(contentTypeString, "Content-Type", response);
ASSERT_EQ(Response::PIPE, response.get().type);
Option<Pipe::Reader> reader = response.get().reader;
ASSERT_SOME(reader);
auto deserializer =
lambda::bind(deserialize<Event>, contentType, lambda::_1);
Reader<Event> responseDecoder(
Decoder<Event>(deserializer),
reader.get());
Future<Result<Event>> event = responseDecoder.read();
AWAIT_READY(event);
ASSERT_SOME(event.get());
// Check event type is subscribed and if the ExecutorID matches.
ASSERT_EQ(Event::SUBSCRIBED, event.get().get().type());
ASSERT_EQ(event.get().get().subscribed().executor_info().executor_id(),
call.executor_id());
reader.get().close();
driver.stop();
driver.join();
}
// This test ensures that when explicit acknowledgements are enabled,
// acknowledgements for master-generated updates are dropped by the
// driver. We test this by creating an invalid task that uses no
// resources.
TEST_F(MesosSchedulerDriverTest, ExplicitAcknowledgementsMasterGeneratedUpdate)
{
Try<PID<Master>> master = StartMaster();
ASSERT_SOME(master);
Try<PID<Slave>> slave = StartSlave();
ASSERT_SOME(slave);
MockScheduler sched;
MesosSchedulerDriver driver(
&sched, DEFAULT_FRAMEWORK_INFO, master.get(), false, DEFAULT_CREDENTIAL);
EXPECT_CALL(sched, registered(&driver, _, _));
Future<vector<Offer>> offers;
EXPECT_CALL(sched, resourceOffers(&driver, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
// Ensure no status update acknowledgements are sent to the master.
EXPECT_NO_FUTURE_CALLS(
mesos::scheduler::Call(),
mesos::scheduler::Call::ACKNOWLEDGE,
_ ,
master.get());
driver.start();
AWAIT_READY(offers);
EXPECT_NE(0u, offers.get().size());
// Launch a task using no resources.
TaskInfo task;
task.set_name("");
task.mutable_task_id()->set_value("1");
task.mutable_slave_id()->MergeFrom(offers.get()[0].slave_id());
task.mutable_executor()->MergeFrom(DEFAULT_EXECUTOR_INFO);
vector<TaskInfo> tasks;
tasks.push_back(task);
Future<TaskStatus> status;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&status));
driver.launchTasks(offers.get()[0].id(), tasks);
AWAIT_READY(status);
ASSERT_EQ(TASK_ERROR, status.get().state());
ASSERT_EQ(TaskStatus::SOURCE_MASTER, status.get().source());
ASSERT_EQ(TaskStatus::REASON_TASK_INVALID, status.get().reason());
// Now send the acknowledgement.
driver.acknowledgeStatusUpdate(status.get());
// Settle the clock to ensure driver processes the acknowledgement,
// which should get dropped due to having come from the master.
Clock::pause();
Clock::settle();
driver.stop();
driver.join();
Shutdown();
}
// This test ensures that the HTTP command executor can self terminate
// after it gets the ACK for the terminal status update from agent.
TEST_F_TEMP_DISABLED_ON_WINDOWS(HTTPCommandExecutorTest, TerminateWithACK)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
slave::Flags flags = CreateSlaveFlags();
flags.http_command_executor = true;
Fetcher fetcher;
Try<MesosContainerizer*> _containerizer =
MesosContainerizer::create(flags, false, &fetcher);
CHECK_SOME(_containerizer);
Owned<MesosContainerizer> containerizer(_containerizer.get());
StandaloneMasterDetector detector(master.get()->pid);
MockSlave slave(flags, &detector, containerizer.get());
spawn(slave);
MockScheduler sched;
MesosSchedulerDriver driver(
&sched, DEFAULT_FRAMEWORK_INFO, master.get()->pid, DEFAULT_CREDENTIAL);
EXPECT_CALL(sched, registered(&driver, _, _));
Future<vector<Offer>> offers;
EXPECT_CALL(sched, resourceOffers(&driver, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driver.start();
AWAIT_READY(offers);
EXPECT_EQ(1u, offers->size());
// Launch a short lived task.
TaskInfo task = createTask(
offers->front().slave_id(),
offers->front().resources(),
"sleep 1");
Future<TaskStatus> statusRunning;
Future<TaskStatus> statusFinished;
EXPECT_CALL(sched, statusUpdate(_, _))
.WillOnce(FutureArg<1>(&statusRunning))
.WillOnce(FutureArg<1>(&statusFinished));
Future<Future<Option<ContainerTermination>>> termination;
EXPECT_CALL(slave, executorTerminated(_, _, _))
.WillOnce(FutureArg<2>(&termination));
driver.launchTasks(offers->front().id(), {task});
// Scheduler should first receive TASK_RUNNING followed by TASK_FINISHED.
AWAIT_READY(statusRunning);
EXPECT_EQ(TASK_RUNNING, statusRunning->state());
AWAIT_READY(statusFinished);
EXPECT_EQ(TASK_FINISHED, statusFinished->state());
// The executor should self terminate with 0 as exit status once
// it gets the ACK for the terminal status update from agent.
AWAIT_READY(termination);
ASSERT_TRUE(termination.get().isReady());
EXPECT_EQ(0, termination.get().get().get().status());
driver.stop();
driver.join();
terminate(slave);
wait(slave);
}
TEST_P(CpuIsolatorTest, ROOT_UserCpuUsage)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
slave::Flags flags = CreateSlaveFlags();
flags.isolation = GetParam();
Fetcher fetcher(flags);
Try<MesosContainerizer*> _containerizer =
MesosContainerizer::create(flags, true, &fetcher);
ASSERT_SOME(_containerizer);
Owned<MesosContainerizer> containerizer(_containerizer.get());
Owned<MasterDetector> detector = master.get()->createDetector();
Try<Owned<cluster::Slave>> slave = StartSlave(
detector.get(),
containerizer.get());
ASSERT_SOME(slave);
MockScheduler sched;
MesosSchedulerDriver driver(
&sched,
DEFAULT_FRAMEWORK_INFO,
master.get()->pid,
DEFAULT_CREDENTIAL);
EXPECT_CALL(sched, registered(&driver, _, _));
Future<vector<Offer>> offers;
EXPECT_CALL(sched, resourceOffers(&driver, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driver.start();
AWAIT_READY(offers);
ASSERT_FALSE(offers->empty());
// Max out a single core in userspace. This will run for at most one
// second.
TaskInfo task = createTask(
offers.get()[0],
"while true ; do true ; done & sleep 60");
Future<TaskStatus> statusRunning;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&statusRunning));
driver.launchTasks(offers.get()[0].id(), {task});
AWAIT_READY(statusRunning);
EXPECT_EQ(TASK_RUNNING, statusRunning->state());
Future<hashset<ContainerID>> containers = containerizer->containers();
AWAIT_READY(containers);
ASSERT_EQ(1u, containers->size());
ContainerID containerId = *(containers->begin());
// Wait up to 1 second for the child process to induce 1/8 of a
// second of user cpu time.
ResourceStatistics statistics;
Duration waited = Duration::zero();
do {
Future<ResourceStatistics> usage = containerizer->usage(containerId);
AWAIT_READY(usage);
statistics = usage.get();
// If we meet our usage expectations, we're done!
if (statistics.cpus_user_time_secs() >= 0.125) {
break;
}
os::sleep(Milliseconds(200));
waited += Milliseconds(200);
} while (waited < Seconds(1));
EXPECT_LE(0.125, statistics.cpus_user_time_secs());
driver.stop();
driver.join();
}
// This test verifies that docker image default cmd is executed correctly.
// This corresponds to the case in runtime isolator logic table: sh=0,
// value=0, argv=1, entrypoint=0, cmd=1.
TEST_F(DockerRuntimeIsolatorTest, ROOT_DockerDefaultCmdLocalPuller)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
const string directory = path::join(os::getcwd(), "archives");
Future<Nothing> testImage =
DockerArchive::create(directory, "alpine", "null", "[\"sh\"]");
AWAIT_READY(testImage);
ASSERT_TRUE(os::exists(path::join(directory, "alpine.tar")));
slave::Flags flags = CreateSlaveFlags();
flags.isolation = "docker/runtime,filesystem/linux";
flags.image_providers = "docker";
flags.docker_registry = directory;
// Make docker store directory as a temparary directory. Because the
// manifest of the test image is changeable, the image cached on
// previous tests should never be used.
flags.docker_store_dir = path::join(os::getcwd(), "store");
Owned<MasterDetector> detector = master.get()->createDetector();
Try<Owned<cluster::Slave>> slave = StartSlave(detector.get(), flags);
ASSERT_SOME(slave);
MockScheduler sched;
MesosSchedulerDriver driver(
&sched, DEFAULT_FRAMEWORK_INFO, master.get()->pid, DEFAULT_CREDENTIAL);
EXPECT_CALL(sched, registered(&driver, _, _));
Future<vector<Offer>> offers;
EXPECT_CALL(sched, resourceOffers(&driver, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driver.start();
AWAIT_READY(offers);
ASSERT_EQ(1u, offers->size());
const Offer& offer = offers.get()[0];
TaskInfo task;
task.set_name("test-task");
task.mutable_task_id()->set_value(UUID::random().toString());
task.mutable_slave_id()->CopyFrom(offer.slave_id());
task.mutable_resources()->CopyFrom(Resources::parse("cpus:1;mem:128").get());
task.mutable_command()->set_shell(false);
task.mutable_command()->add_arguments("-c");
task.mutable_command()->add_arguments("echo 'hello world'");
Image image;
image.set_type(Image::DOCKER);
image.mutable_docker()->set_name("alpine");
ContainerInfo* container = task.mutable_container();
container->set_type(ContainerInfo::MESOS);
container->mutable_mesos()->mutable_image()->CopyFrom(image);
Future<TaskStatus> statusRunning;
Future<TaskStatus> statusFinished;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&statusRunning))
.WillOnce(FutureArg<1>(&statusFinished));
driver.launchTasks(offer.id(), {task});
AWAIT_READY_FOR(statusRunning, Seconds(60));
EXPECT_EQ(task.task_id(), statusRunning->task_id());
EXPECT_EQ(TASK_RUNNING, statusRunning->state());
AWAIT_READY(statusFinished);
EXPECT_EQ(task.task_id(), statusFinished->task_id());
EXPECT_EQ(TASK_FINISHED, statusFinished->state());
driver.stop();
driver.join();
}
// This test verifies that docker image default entrypoint is executed
// correctly using registry puller. This corresponds to the case in runtime
// isolator logic table: sh=0, value=0, argv=1, entrypoint=1, cmd=0.
TEST_F(DockerRuntimeIsolatorTest,
ROOT_CURL_INTERNET_DockerDefaultEntryptRegistryPuller)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
slave::Flags flags = CreateSlaveFlags();
flags.isolation = "docker/runtime,filesystem/linux";
flags.image_providers = "docker";
flags.docker_store_dir = path::join(os::getcwd(), "store");
Owned<MasterDetector> detector = master.get()->createDetector();
Try<Owned<cluster::Slave>> slave = StartSlave(detector.get(), flags);
ASSERT_SOME(slave);
MockScheduler sched;
MesosSchedulerDriver driver(
&sched, DEFAULT_FRAMEWORK_INFO, master.get()->pid, DEFAULT_CREDENTIAL);
EXPECT_CALL(sched, registered(&driver, _, _));
Future<vector<Offer>> offers;
EXPECT_CALL(sched, resourceOffers(&driver, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driver.start();
AWAIT_READY(offers);
ASSERT_EQ(1u, offers->size());
const Offer& offer = offers.get()[0];
TaskInfo task;
task.set_name("test-task");
task.mutable_task_id()->set_value(UUID::random().toString());
task.mutable_slave_id()->CopyFrom(offer.slave_id());
task.mutable_resources()->CopyFrom(Resources::parse("cpus:1;mem:128").get());
task.mutable_command()->set_shell(false);
task.mutable_command()->add_arguments("hello world");
Image image;
image.set_type(Image::DOCKER);
// 'mesosphere/inky' image is used in docker containerizer test, which
// contains entrypoint as 'echo' and cmd as null.
image.mutable_docker()->set_name("mesosphere/inky");
ContainerInfo* container = task.mutable_container();
container->set_type(ContainerInfo::MESOS);
container->mutable_mesos()->mutable_image()->CopyFrom(image);
Future<TaskStatus> statusRunning;
Future<TaskStatus> statusFinished;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&statusRunning))
.WillOnce(FutureArg<1>(&statusFinished));
driver.launchTasks(offer.id(), {task});
AWAIT_READY_FOR(statusRunning, Seconds(60));
EXPECT_EQ(task.task_id(), statusRunning->task_id());
EXPECT_EQ(TASK_RUNNING, statusRunning->state());
AWAIT_READY(statusFinished);
EXPECT_EQ(task.task_id(), statusFinished->task_id());
EXPECT_EQ(TASK_FINISHED, statusFinished->state());
driver.stop();
driver.join();
}
// Test that memory pressure listening is restarted after recovery.
TEST_F(MemoryPressureMesosTest, CGROUPS_ROOT_SlaveRecovery)
{
Try<PID<Master>> master = StartMaster();
ASSERT_SOME(master);
slave::Flags flags = CreateSlaveFlags();
// We only care about memory cgroup for this test.
flags.isolation = "cgroups/mem";
flags.slave_subsystems = None();
Fetcher fetcher;
Try<MesosContainerizer*> containerizer1 =
MesosContainerizer::create(flags, true, &fetcher);
ASSERT_SOME(containerizer1);
Try<PID<Slave>> slave = StartSlave(containerizer1.get(), flags);
ASSERT_SOME(slave);
MockScheduler sched;
// Enable checkpointing for the framework.
FrameworkInfo frameworkInfo = DEFAULT_FRAMEWORK_INFO;
frameworkInfo.set_checkpoint(true);
MesosSchedulerDriver driver(
&sched, frameworkInfo, master.get(), DEFAULT_CREDENTIAL);
EXPECT_CALL(sched, registered(_, _, _));
Future<vector<Offer>> offers;
EXPECT_CALL(sched, resourceOffers(_, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driver.start();
AWAIT_READY(offers);
EXPECT_NE(0u, offers.get().size());
Offer offer = offers.get()[0];
// Run a task that triggers memory pressure event. We request 1G
// disk because we are going to write a 512 MB file repeatedly.
TaskInfo task = createTask(
offer.slave_id(),
Resources::parse("cpus:1;mem:256;disk:1024").get(),
"while true; do dd count=512 bs=1M if=/dev/zero of=./temp; done");
Future<TaskStatus> status;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&status))
.WillRepeatedly(Return()); // Ignore subsequent updates.
driver.launchTasks(offers.get()[0].id(), {task});
AWAIT_READY(status);
EXPECT_EQ(task.task_id(), status.get().task_id());
EXPECT_EQ(TASK_RUNNING, status.get().state());
// We restart the slave to let it recover.
Stop(slave.get());
delete containerizer1.get();
// Set up so we can wait until the new slave updates the container's
// resources (this occurs after the executor has re-registered).
Future<Nothing> update =
FUTURE_DISPATCH(_, &MesosContainerizerProcess::update);
// Use the same flags.
Try<MesosContainerizer*> containerizer2 =
MesosContainerizer::create(flags, true, &fetcher);
ASSERT_SOME(containerizer2);
slave = StartSlave(containerizer2.get(), flags);
ASSERT_SOME(slave);
// Wait until the containerizer is updated.
AWAIT_READY(update);
Future<hashset<ContainerID>> containers = containerizer2.get()->containers();
AWAIT_READY(containers);
ASSERT_EQ(1u, containers.get().size());
ContainerID containerId = *(containers.get().begin());
// Wait a while for some memory pressure events to occur.
Duration waited = Duration::zero();
do {
Future<ResourceStatistics> usage = containerizer2.get()->usage(containerId);
AWAIT_READY(usage);
if (usage.get().mem_low_pressure_counter() > 0) {
// We will check the correctness of the memory pressure counters
// later, because the memory-hammering task is still active
// and potentially incrementing these counters.
break;
}
os::sleep(Milliseconds(100));
waited += Milliseconds(100);
} while (waited < Seconds(5));
EXPECT_LE(waited, Seconds(5));
// Stop the memory-hammering task.
driver.killTask(task.task_id());
// Process any queued up events through before proceeding.
process::Clock::pause();
process::Clock::settle();
process::Clock::resume();
// Now check the correctness of the memory pressure counters.
Future<ResourceStatistics> usage = containerizer2.get()->usage(containerId);
AWAIT_READY(usage);
EXPECT_GE(usage.get().mem_low_pressure_counter(),
usage.get().mem_medium_pressure_counter());
EXPECT_GE(usage.get().mem_medium_pressure_counter(),
usage.get().mem_critical_pressure_counter());
driver.stop();
driver.join();
Shutdown();
delete containerizer2.get();
}
// Test that memory pressure listening is restarted after recovery.
TEST_F(MemoryPressureMesosTest, CGROUPS_ROOT_SlaveRecovery)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
slave::Flags flags = CreateSlaveFlags();
// We only care about memory cgroup for this test.
flags.isolation = "cgroups/mem";
Fetcher fetcher(flags);
Try<MesosContainerizer*> _containerizer =
MesosContainerizer::create(flags, true, &fetcher);
ASSERT_SOME(_containerizer);
Owned<MesosContainerizer> containerizer(_containerizer.get());
Owned<MasterDetector> detector = master.get()->createDetector();
Try<Owned<cluster::Slave>> slave =
StartSlave(detector.get(), containerizer.get(), flags);
ASSERT_SOME(slave);
MockScheduler sched;
// Enable checkpointing for the framework.
FrameworkInfo frameworkInfo = DEFAULT_FRAMEWORK_INFO;
frameworkInfo.set_checkpoint(true);
MesosSchedulerDriver driver(
&sched, frameworkInfo, master.get()->pid, DEFAULT_CREDENTIAL);
EXPECT_CALL(sched, registered(_, _, _));
Future<vector<Offer>> offers;
EXPECT_CALL(sched, resourceOffers(_, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driver.start();
AWAIT_READY(offers);
ASSERT_FALSE(offers->empty());
Offer offer = offers.get()[0];
// Run a task that triggers memory pressure event. We request 1G
// disk because we are going to write a 512 MB file repeatedly.
TaskInfo task = createTask(
offer.slave_id(),
Resources::parse("cpus:1;mem:256;disk:1024").get(),
"while true; do dd count=512 bs=1M if=/dev/zero of=./temp; done");
Future<TaskStatus> starting;
Future<TaskStatus> running;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&starting))
.WillOnce(FutureArg<1>(&running))
.WillRepeatedly(Return()); // Ignore subsequent updates.
Future<Nothing> runningAck =
FUTURE_DISPATCH(_, &Slave::_statusUpdateAcknowledgement);
Future<Nothing> startingAck =
FUTURE_DISPATCH(_, &Slave::_statusUpdateAcknowledgement);
driver.launchTasks(offers.get()[0].id(), {task});
AWAIT_READY(starting);
EXPECT_EQ(task.task_id(), starting->task_id());
EXPECT_EQ(TASK_STARTING, starting->state());
AWAIT_READY(startingAck);
AWAIT_READY(running);
EXPECT_EQ(task.task_id(), running->task_id());
EXPECT_EQ(TASK_RUNNING, running->state());
// Wait for the ACK to be checkpointed.
AWAIT_READY_FOR(runningAck, Seconds(120));
// We restart the slave to let it recover.
slave.get()->terminate();
// Set up so we can wait until the new slave updates the container's
// resources (this occurs after the executor has re-registered).
Future<Nothing> update =
FUTURE_DISPATCH(_, &MesosContainerizerProcess::update);
// Use the same flags.
_containerizer = MesosContainerizer::create(flags, true, &fetcher);
ASSERT_SOME(_containerizer);
containerizer.reset(_containerizer.get());
Future<SlaveReregisteredMessage> reregistered =
FUTURE_PROTOBUF(SlaveReregisteredMessage(), master.get()->pid, _);
slave = StartSlave(detector.get(), containerizer.get(), flags);
ASSERT_SOME(slave);
AWAIT_READY(reregistered);
// Wait until the containerizer is updated.
AWAIT_READY(update);
Future<hashset<ContainerID>> containers = containerizer->containers();
AWAIT_READY(containers);
ASSERT_EQ(1u, containers->size());
ContainerID containerId = *(containers->begin());
// Wait a while for some memory pressure events to occur.
Duration waited = Duration::zero();
do {
Future<ResourceStatistics> usage = containerizer->usage(containerId);
AWAIT_READY(usage);
if (usage->mem_low_pressure_counter() > 0) {
// We will check the correctness of the memory pressure counters
// later, because the memory-hammering task is still active
// and potentially incrementing these counters.
break;
}
os::sleep(Milliseconds(100));
waited += Milliseconds(100);
} while (waited < Seconds(5));
EXPECT_LE(waited, Seconds(5));
// Pause the clock to ensure that the reaper doesn't reap the exited
// command executor and inform the containerizer/slave.
Clock::pause();
Clock::settle();
Future<TaskStatus> killed;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&killed));
// Stop the memory-hammering task.
driver.killTask(task.task_id());
AWAIT_READY_FOR(killed, Seconds(120));
EXPECT_EQ(task.task_id(), killed->task_id());
EXPECT_EQ(TASK_KILLED, killed->state());
// Now check the correctness of the memory pressure counters.
Future<ResourceStatistics> usage = containerizer->usage(containerId);
AWAIT_READY(usage);
EXPECT_GE(usage->mem_low_pressure_counter(),
usage->mem_medium_pressure_counter());
EXPECT_GE(usage->mem_medium_pressure_counter(),
usage->mem_critical_pressure_counter());
Clock::resume();
driver.stop();
driver.join();
}
// The purpose of this test is to ensure that when slaves are removed
// from the master, and then attempt to send exited executor messages,
// 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 exited executor messages if
// it was unaware that the master removed it. We've already
// notified frameworks that the tasks under the executors were LOST,
// so we have to have the slave shut down.
TEST_F(PartitionTest, PartitionedSlaveExitedExecutor)
{
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(), _, _);
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);
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));\
Future<vector<Offer>> offers;
EXPECT_CALL(sched, resourceOffers(&driver, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return());
driver.start();
AWAIT_READY(frameworkId);
AWAIT_READY(offers);
ASSERT_NE(0u, offers.get().size());
// Launch a task. This allows us to have the slave send an
// ExitedExecutorMessage.
TaskID taskId;
taskId.set_value("1");
TaskInfo task;
task.set_name("");
task.mutable_task_id()->MergeFrom(taskId);
task.mutable_slave_id()->MergeFrom(offers.get()[0].slave_id());
task.mutable_resources()->MergeFrom(offers.get()[0].resources());
task.mutable_executor()->MergeFrom(DEFAULT_EXECUTOR_INFO);
task.mutable_executor()->mutable_command()->set_value("sleep 60");
// Set up the expectations for launching the task.
EXPECT_CALL(exec, registered(_, _, _, _));
EXPECT_CALL(exec, launchTask(_, _))
.WillOnce(SendStatusUpdateFromTask(TASK_RUNNING));
// Drop all the status updates from the slave, so that we can
// ensure the ExitedExecutorMessage is what triggers the slave
// shutdown.
DROP_PROTOBUFS(StatusUpdateMessage(), _, master.get()->pid);
driver.launchTasks(offers.get()[0].id(), {task});
// Drop the first shutdown message from the master (simulated
// partition) and allow the second shutdown message to pass when
// triggered by the ExitedExecutorMessage.
Future<ShutdownMessage> shutdownMessage =
DROP_PROTOBUF(ShutdownMessage(), _, slave.get()->pid);
Future<TaskStatus> lostStatus;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&lostStatus));
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();
// The master will have notified the framework of the lost task.
AWAIT_READY(lostStatus);
EXPECT_EQ(TASK_LOST, lostStatus.get().state());
// 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);
// Induce an ExitedExecutorMessage from the slave.
containerizer.destroy(
frameworkId.get(), DEFAULT_EXECUTOR_INFO.executor_id());
// Upon receiving the message, the master will shutdown the slave.
AWAIT_READY(shutdownMessage);
Clock::resume();
driver.stop();
driver.join();
}
// This test does not set any Accept header for the subscribe call.
// The default response media type should be "application/json" in
// this case.
TEST_P(ExecutorHttpApiTest, NoAcceptHeader)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
ExecutorID executorId = DEFAULT_EXECUTOR_ID;
MockExecutor exec(executorId);
TestContainerizer containerizer(&exec);
Owned<MasterDetector> detector = master.get()->createDetector();
Try<Owned<cluster::Slave>> slave = StartSlave(detector.get(), &containerizer);
ASSERT_SOME(slave);
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));
Future<vector<Offer>> offers;
EXPECT_CALL(sched, resourceOffers(&driver, _))
.WillOnce(FutureArg<1>(&offers));
Future<Nothing> statusUpdate;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureSatisfy(&statusUpdate));
driver.start();
AWAIT_READY(frameworkId);
AWAIT_READY(offers);
ASSERT_EQ(1u, offers.get().size());
EXPECT_CALL(exec, registered(_, _, _, _))
.Times(1);
EXPECT_CALL(exec, launchTask(_, _))
.WillOnce(SendStatusUpdateFromTask(TASK_RUNNING));
TaskInfo taskInfo = createTask(offers.get()[0], "", executorId);
driver.launchTasks(offers.get()[0].id(), {taskInfo});
// Wait until status update is received on the scheduler before sending
// an executor subscribe request.
AWAIT_READY(statusUpdate);
// Only subscribe needs to 'Accept' JSON or protobuf.
Call call;
call.mutable_framework_id()->CopyFrom(evolve(frameworkId.get()));
call.mutable_executor_id()->CopyFrom(evolve(executorId));
call.set_type(Call::SUBSCRIBE);
call.mutable_subscribe();
// Retrieve the parameter passed as content type to this test.
const ContentType contentType = GetParam();
// No 'Accept' header leads to all media types considered
// acceptable. JSON will be chosen by default.
process::http::Headers headers;
Future<Response> response = process::http::streaming::post(
slave.get()->pid,
"api/v1/executor",
headers,
serialize(contentType, call),
stringify(contentType));
AWAIT_EXPECT_RESPONSE_STATUS_EQ(OK().status, response);
AWAIT_EXPECT_RESPONSE_HEADER_EQ(APPLICATION_JSON, "Content-Type", response);
driver.stop();
driver.join();
}
// Ensures that when a scheduler enables explicit acknowledgements
// on the driver, there are no implicit acknowledgements sent, and
// the call to 'acknowledgeStatusUpdate' sends the ack to the master.
TEST_F(MesosSchedulerDriverTest, ExplicitAcknowledgements)
{
Try<PID<Master>> master = StartMaster();
ASSERT_SOME(master);
MockExecutor exec(DEFAULT_EXECUTOR_ID);
TestContainerizer containerizer(&exec);
Try<PID<Slave>> slave = StartSlave(&containerizer);
ASSERT_SOME(slave);
MockScheduler sched;
MesosSchedulerDriver driver(
&sched, DEFAULT_FRAMEWORK_INFO, master.get(), false, DEFAULT_CREDENTIAL);
EXPECT_CALL(sched, registered(&driver, _, _));
EXPECT_CALL(sched, resourceOffers(&driver, _))
.WillOnce(LaunchTasks(DEFAULT_EXECUTOR_INFO, 1, 1, 16, "*"))
.WillRepeatedly(Return()); // Ignore subsequent offers.
Future<TaskStatus> status;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&status));
// Ensure no status update acknowledgements are sent from the driver
// to the master until the explicit acknowledgement is sent.
EXPECT_NO_FUTURE_CALLS(
mesos::scheduler::Call(),
mesos::scheduler::Call::ACKNOWLEDGE,
_ ,
master.get());
EXPECT_CALL(exec, registered(_, _, _, _));
EXPECT_CALL(exec, launchTask(_, _))
.WillOnce(SendStatusUpdateFromTask(TASK_RUNNING));
EXPECT_CALL(exec, shutdown(_))
.Times(AtMost(1));
driver.start();
AWAIT_READY(status);
// Settle the clock to ensure driver finishes processing the status
// update, we want to ensure that no implicit acknowledgement gets
// sent.
Clock::pause();
Clock::settle();
// Now send the acknowledgement.
Future<mesos::scheduler::Call> acknowledgement = FUTURE_CALL(
mesos::scheduler::Call(),
mesos::scheduler::Call::ACKNOWLEDGE,
_,
master.get());
driver.acknowledgeStatusUpdate(status.get());
AWAIT_READY(acknowledgement);
driver.stop();
driver.join();
Shutdown();
}
// This test ensures that a task will transition straight from `TASK_KILLING` to
// `TASK_KILLED`, even if the health check begins to fail during the kill policy
// grace period.
//
// TODO(gkleiman): this test takes about 7 seconds to run, consider using mock
// tasks and health checkers to speed it up.
TEST_P(CommandExecutorTest, NoTransitionFromKillingToRunning)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
Owned<MasterDetector> detector = master.get()->createDetector();
slave::Flags flags = CreateSlaveFlags();
flags.http_command_executor = GetParam();
Try<Owned<cluster::Slave>> slave = StartSlave(detector.get(), flags);
ASSERT_SOME(slave);
// Start the framework with the task killing capability.
FrameworkInfo::Capability capability;
capability.set_type(FrameworkInfo::Capability::TASK_KILLING_STATE);
FrameworkInfo frameworkInfo = DEFAULT_FRAMEWORK_INFO;
frameworkInfo.add_capabilities()->CopyFrom(capability);
MockScheduler sched;
MesosSchedulerDriver driver(
&sched, frameworkInfo, master.get()->pid, DEFAULT_CREDENTIAL);
EXPECT_CALL(sched, registered(&driver, _, _));
Future<vector<Offer>> offers;
EXPECT_CALL(sched, resourceOffers(&driver, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driver.start();
AWAIT_READY(offers);
EXPECT_EQ(1u, offers->size());
const string command = strings::format(
"%s %s --sleep_duration=15",
getTestHelperPath("test-helper"),
KillPolicyTestHelper::NAME).get();
TaskInfo task = createTask(offers->front(), command);
// Create a health check that succeeds until a temporary file is removed.
Try<string> temporaryPath = os::mktemp(path::join(os::getcwd(), "XXXXXX"));
ASSERT_SOME(temporaryPath);
const string tmpPath = temporaryPath.get();
HealthCheck healthCheck;
healthCheck.set_type(HealthCheck::COMMAND);
healthCheck.mutable_command()->set_value("ls " + tmpPath + " >/dev/null");
healthCheck.set_delay_seconds(0);
healthCheck.set_grace_period_seconds(0);
healthCheck.set_interval_seconds(0);
task.mutable_health_check()->CopyFrom(healthCheck);
// Set the kill policy grace period to 5 seconds.
KillPolicy killPolicy;
killPolicy.mutable_grace_period()->set_nanoseconds(Seconds(5).ns());
task.mutable_kill_policy()->CopyFrom(killPolicy);
vector<TaskInfo> tasks;
tasks.push_back(task);
Future<TaskStatus> statusRunning;
Future<TaskStatus> statusHealthy;
Future<TaskStatus> statusKilling;
Future<TaskStatus> statusKilled;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&statusRunning))
.WillOnce(FutureArg<1>(&statusHealthy))
.WillOnce(FutureArg<1>(&statusKilling))
.WillOnce(FutureArg<1>(&statusKilled));
driver.launchTasks(offers->front().id(), tasks);
AWAIT_READY(statusRunning);
EXPECT_EQ(TASK_RUNNING, statusRunning.get().state());
AWAIT_READY(statusHealthy);
EXPECT_EQ(TASK_RUNNING, statusHealthy.get().state());
EXPECT_TRUE(statusHealthy.get().has_healthy());
EXPECT_TRUE(statusHealthy.get().healthy());
driver.killTask(task.task_id());
AWAIT_READY(statusKilling);
EXPECT_EQ(TASK_KILLING, statusKilling->state());
EXPECT_FALSE(statusKilling.get().has_healthy());
// Remove the temporary file, so that the health check fails.
os::rm(tmpPath);
AWAIT_READY(statusKilled);
EXPECT_EQ(TASK_KILLED, statusKilled->state());
EXPECT_FALSE(statusKilled.get().has_healthy());
driver.stop();
driver.join();
}
TEST_F(ResourceOffersTest, ResourcesGetReofferedAfterTaskInfoError)
{
Try<PID<Master>> master = StartMaster();
ASSERT_SOME(master);
Try<PID<Slave>> slave = StartSlave();
ASSERT_SOME(slave);
MockScheduler sched1;
MesosSchedulerDriver driver1(
&sched1, DEFAULT_FRAMEWORK_INFO, master.get(), DEFAULT_CREDENTIAL);
EXPECT_CALL(sched1, registered(&driver1, _, _))
.Times(1);
Future<vector<Offer>> offers;
EXPECT_CALL(sched1, resourceOffers(&driver1, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driver1.start();
AWAIT_READY(offers);
EXPECT_NE(0u, offers.get().size());
TaskInfo task;
task.set_name("");
task.mutable_task_id()->set_value("1");
task.mutable_slave_id()->MergeFrom(offers.get()[0].slave_id());
task.mutable_executor()->MergeFrom(DEFAULT_EXECUTOR_INFO);
Resource* cpus = task.add_resources();
cpus->set_name("cpus");
cpus->set_type(Value::SCALAR);
cpus->mutable_scalar()->set_value(-1);
Resource* mem = task.add_resources();
mem->set_name("mem");
mem->set_type(Value::SCALAR);
mem->mutable_scalar()->set_value(Gigabytes(1).bytes());
vector<TaskInfo> tasks;
tasks.push_back(task);
Future<TaskStatus> status;
EXPECT_CALL(sched1, statusUpdate(&driver1, _))
.WillOnce(FutureArg<1>(&status));
driver1.launchTasks(offers.get()[0].id(), tasks);
AWAIT_READY(status);
EXPECT_EQ(task.task_id(), status.get().task_id());
EXPECT_EQ(TASK_ERROR, status.get().state());
EXPECT_EQ(TaskStatus::REASON_TASK_INVALID, status.get().reason());
EXPECT_TRUE(status.get().has_message());
EXPECT_TRUE(strings::startsWith(
status.get().message(), "Task uses invalid resources"));
MockScheduler sched2;
MesosSchedulerDriver driver2(
&sched2, DEFAULT_FRAMEWORK_INFO, master.get(), DEFAULT_CREDENTIAL);
EXPECT_CALL(sched2, registered(&driver2, _, _))
.Times(1);
EXPECT_CALL(sched2, resourceOffers(&driver2, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driver2.start();
AWAIT_READY(offers);
driver1.stop();
driver1.join();
driver2.stop();
driver2.join();
Shutdown();
}
// This test ensures that driver based schedulers using explicit
// acknowledgements can acknowledge status updates sent from
// HTTP based executors.
TEST_F_TEMP_DISABLED_ON_WINDOWS(
HTTPCommandExecutorTest,
ExplicitAcknowledgements)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
Owned<MasterDetector> detector = master.get()->createDetector();
slave::Flags flags = CreateSlaveFlags();
flags.http_command_executor = true;
Try<Owned<cluster::Slave>> slave = StartSlave(detector.get(), flags);
ASSERT_SOME(slave);
MockScheduler sched;
MesosSchedulerDriver driver(
&sched,
DEFAULT_FRAMEWORK_INFO,
master.get()->pid,
false,
DEFAULT_CREDENTIAL);
EXPECT_CALL(sched, registered(&driver, _, _));
Future<vector<Offer>> offers;
EXPECT_CALL(sched, resourceOffers(&driver, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driver.start();
AWAIT_READY(offers);
EXPECT_EQ(1u, offers->size());
// Launch a task with the command executor.
TaskInfo task = createTask(
offers->front().slave_id(),
offers->front().resources(),
"sleep 1000");
Future<TaskStatus> statusRunning;
EXPECT_CALL(sched, statusUpdate(_, _))
.WillOnce(FutureArg<1>(&statusRunning));
// Ensure no status update acknowledgements are sent from the driver
// to the master until the explicit acknowledgement is sent.
EXPECT_NO_FUTURE_CALLS(
mesos::scheduler::Call(),
mesos::scheduler::Call::ACKNOWLEDGE,
_ ,
master.get()->pid);
driver.launchTasks(offers->front().id(), {task});
AWAIT_READY(statusRunning);
EXPECT_TRUE(statusRunning->has_slave_id());
EXPECT_EQ(TASK_RUNNING, statusRunning->state());
// Now send the acknowledgement.
Future<mesos::scheduler::Call> acknowledgement = FUTURE_CALL(
mesos::scheduler::Call(),
mesos::scheduler::Call::ACKNOWLEDGE,
_,
master.get()->pid);
driver.acknowledgeStatusUpdate(statusRunning.get());
AWAIT_READY(acknowledgement);
driver.stop();
driver.join();
}
TEST_F(ResourceOffersTest, ResourcesGetReofferedAfterTaskInfoError)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
Owned<MasterDetector> detector = master.get()->createDetector();
Try<Owned<cluster::Slave>> slave = StartSlave(detector.get());
ASSERT_SOME(slave);
MockScheduler sched1;
MesosSchedulerDriver driver1(
&sched1, DEFAULT_FRAMEWORK_INFO, master.get()->pid, DEFAULT_CREDENTIAL);
EXPECT_CALL(sched1, registered(&driver1, _, _));
Future<vector<Offer>> offers;
EXPECT_CALL(sched1, resourceOffers(&driver1, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driver1.start();
AWAIT_READY(offers);
ASSERT_FALSE(offers->empty());
TaskInfo task;
task.set_name("");
task.mutable_task_id()->set_value("1");
task.mutable_slave_id()->MergeFrom(offers.get()[0].slave_id());
task.mutable_executor()->MergeFrom(DEFAULT_EXECUTOR_INFO);
Resource* cpus = task.add_resources();
cpus->set_name("cpus");
cpus->set_type(Value::SCALAR);
cpus->mutable_scalar()->set_value(-1);
Resource* mem = task.add_resources();
mem->set_name("mem");
mem->set_type(Value::SCALAR);
mem->mutable_scalar()->set_value(static_cast<double>(Gigabytes(1).bytes()));
vector<TaskInfo> tasks;
tasks.push_back(task);
Future<TaskStatus> status;
EXPECT_CALL(sched1, statusUpdate(&driver1, _))
.WillOnce(FutureArg<1>(&status));
driver1.launchTasks(offers.get()[0].id(), tasks);
AWAIT_READY(status);
EXPECT_EQ(task.task_id(), status->task_id());
EXPECT_EQ(TASK_ERROR, status->state());
EXPECT_EQ(TaskStatus::REASON_TASK_INVALID, status->reason());
EXPECT_TRUE(status->has_message());
EXPECT_TRUE(strings::contains(status->message(), "Invalid scalar resource"))
<< status->message();
MockScheduler sched2;
MesosSchedulerDriver driver2(
&sched2, DEFAULT_FRAMEWORK_INFO, master.get()->pid, DEFAULT_CREDENTIAL);
EXPECT_CALL(sched2, registered(&driver2, _, _));
EXPECT_CALL(sched2, resourceOffers(&driver2, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driver2.start();
AWAIT_READY(offers);
driver1.stop();
driver1.join();
driver2.stop();
driver2.join();
}
// This test has been temporarily disabled due to MESOS-1257.
TEST_F(ExternalContainerizerTest, DISABLED_Launch)
{
Try<PID<Master> > master = this->StartMaster();
ASSERT_SOME(master);
Flags testFlags;
slave::Flags flags = this->CreateSlaveFlags();
flags.isolation = "external";
flags.containerizer_path =
testFlags.build_dir + "/src/examples/python/test-containerizer";
MockExternalContainerizer containerizer(flags);
Try<PID<Slave> > slave = this->StartSlave(&containerizer, flags);
ASSERT_SOME(slave);
MockScheduler sched;
MesosSchedulerDriver driver(
&sched, DEFAULT_FRAMEWORK_INFO, master.get(), DEFAULT_CREDENTIAL);
Future<FrameworkID> frameworkId;
EXPECT_CALL(sched, registered(&driver, _, _))
.WillOnce(FutureArg<1>(&frameworkId));
Future<vector<Offer> > offers;
EXPECT_CALL(sched, resourceOffers(&driver, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driver.start();
AWAIT_READY(frameworkId);
AWAIT_READY(offers);
EXPECT_NE(0u, offers.get().size());
TaskInfo task;
task.set_name("isolator_test");
task.mutable_task_id()->set_value("1");
task.mutable_slave_id()->CopyFrom(offers.get()[0].slave_id());
task.mutable_resources()->CopyFrom(offers.get()[0].resources());
Resources resources(offers.get()[0].resources());
Option<Bytes> mem = resources.mem();
ASSERT_SOME(mem);
Option<double> cpus = resources.cpus();
ASSERT_SOME(cpus);
const std::string& file = path::join(flags.work_dir, "ready");
// This task induces user/system load in a child process by
// running top in a child process for ten seconds.
task.mutable_command()->set_value(
#ifdef __APPLE__
// Use logging mode with 30,000 samples with no interval.
"top -l 30000 -s 0 2>&1 > /dev/null & "
#else
// Batch mode, with 30,000 samples with no interval.
"top -b -d 0 -n 30000 2>&1 > /dev/null & "
#endif
"touch " + file + "; " // Signals that the top command is running.
"sleep 60");
Future<TaskStatus> status;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&status))
.WillRepeatedly(Return()); // Ignore rest for now.
Future<ContainerID> containerId;
EXPECT_CALL(containerizer, launch(_, _, _, _, _, _, _, _))
.WillOnce(DoAll(FutureArg<0>(&containerId),
Invoke(&containerizer,
&MockExternalContainerizer::_launch)));
driver.launchTasks(offers.get()[0].id(), {task});
AWAIT_READY(containerId);
AWAIT_READY(status);
EXPECT_EQ(TASK_RUNNING, status.get().state());
// Wait for the task to begin inducing cpu time.
while (!os::exists(file));
ExecutorID executorId;
executorId.set_value(task.task_id().value());
// We'll wait up to 10 seconds for the child process to induce
// 1/8 of a second of user and system cpu time in total.
// TODO(bmahler): Also induce rss memory consumption, by re-using
// the balloon framework.
ResourceStatistics statistics;
Duration waited = Duration::zero();
do {
Future<ResourceStatistics> usage = containerizer.usage(containerId.get());
AWAIT_READY(usage);
statistics = usage.get();
// If we meet our usage expectations, we're done!
// NOTE: We are currently getting dummy-data from the test-
// containerizer python script matching these expectations.
// TODO(tillt): Consider working with real data.
if (statistics.cpus_user_time_secs() >= 0.120 &&
statistics.cpus_system_time_secs() >= 0.05 &&
statistics.mem_rss_bytes() >= 1024u) {
break;
}
os::sleep(Milliseconds(100));
waited += Milliseconds(100);
} while (waited < Seconds(10));
EXPECT_GE(statistics.cpus_user_time_secs(), 0.120);
EXPECT_GE(statistics.cpus_system_time_secs(), 0.05);
EXPECT_EQ(statistics.cpus_limit(), cpus.get());
EXPECT_GE(statistics.mem_rss_bytes(), 1024u);
EXPECT_EQ(statistics.mem_limit_bytes(), mem.get().bytes());
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&status));
driver.killTask(task.task_id());
AWAIT_READY(status);
EXPECT_EQ(TASK_KILLED, status.get().state());
driver.stop();
driver.join();
this->Shutdown();
}
// This test verifies that a re-registering slave sends the terminal
// unacknowledged tasks for a terminal executor. This is required
// for the master to correctly reconcile its view with the slave's
// view of tasks. This test drops a terminal update to the master
// and then forces the slave to re-register.
TEST_F(MasterSlaveReconciliationTest, SlaveReregisterTerminatedExecutor)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
MockExecutor exec(DEFAULT_EXECUTOR_ID);
TestContainerizer containerizer(&exec);
StandaloneMasterDetector detector(master.get()->pid);
Try<Owned<cluster::Slave>> slave = StartSlave(&detector, &containerizer);
ASSERT_SOME(slave);
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, _))
.WillOnce(LaunchTasks(DEFAULT_EXECUTOR_INFO, 1, 1, 512, "*"))
.WillRepeatedly(Return()); // Ignore subsequent offers.
ExecutorDriver* execDriver;
EXPECT_CALL(exec, registered(_, _, _, _))
.WillOnce(SaveArg<0>(&execDriver));
EXPECT_CALL(exec, launchTask(_, _))
.WillOnce(SendStatusUpdateFromTask(TASK_RUNNING));
Future<TaskStatus> status;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&status));
Future<StatusUpdateAcknowledgementMessage> statusUpdateAcknowledgementMessage
= FUTURE_PROTOBUF(
StatusUpdateAcknowledgementMessage(),
master.get()->pid,
slave.get()->pid);
driver.start();
AWAIT_READY(status);
EXPECT_EQ(TASK_RUNNING, status.get().state());
// Make sure the acknowledgement reaches the slave.
AWAIT_READY(statusUpdateAcknowledgementMessage);
// Drop the TASK_FINISHED status update sent to the master.
Future<StatusUpdateMessage> statusUpdateMessage =
DROP_PROTOBUF(StatusUpdateMessage(), _, master.get()->pid);
Future<ExitedExecutorMessage> executorExitedMessage =
FUTURE_PROTOBUF(ExitedExecutorMessage(), _, _);
TaskStatus finishedStatus;
finishedStatus = status.get();
finishedStatus.set_state(TASK_FINISHED);
execDriver->sendStatusUpdate(finishedStatus);
// Ensure the update was sent.
AWAIT_READY(statusUpdateMessage);
EXPECT_CALL(sched, executorLost(&driver, DEFAULT_EXECUTOR_ID, _, _));
// Now kill the executor.
containerizer.destroy(frameworkId.get(), DEFAULT_EXECUTOR_ID);
Future<TaskStatus> status2;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&status2));
// We drop the 'UpdateFrameworkMessage' from the master to slave to
// stop the status update manager from retrying the update that was
// already sent due to the new master detection.
DROP_PROTOBUFS(UpdateFrameworkMessage(), _, _);
detector.appoint(master.get()->pid);
AWAIT_READY(status2);
EXPECT_EQ(TASK_FINISHED, status2.get().state());
driver.stop();
driver.join();
}
