// This test confirms that setting no values for the soft and hard
// limits implies an unlimited resource.
TEST_F(PosixRLimitsIsolatorTest, UnsetLimits) {
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
slave::Flags flags = CreateSlaveFlags();
flags.isolation = "posix/rlimits";
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(_, _, _));
Future<vector<Offer>> offers;
EXPECT_CALL(sched, resourceOffers(_, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driver.start();
AWAIT_READY(offers);
ASSERT_NE(0u, offers->size());
TaskInfo task = createTask(
offers.get()[0].slave_id(),
offers.get()[0].resources(),
"exit `ulimit -c | grep -q unlimited`");
// Force usage of C locale as we interpret a potentially translated
// string in the task's command.
mesos::Environment::Variable* locale =
task.mutable_command()->mutable_environment()->add_variables();
locale->set_name("LC_ALL");
locale->set_value("C");
ContainerInfo* container = task.mutable_container();
container->set_type(ContainerInfo::MESOS);
// Setting rlimit for core without soft or hard limit signifies
// unlimited range.
RLimitInfo rlimitInfo;
RLimitInfo::RLimit* rlimit = rlimitInfo.add_rlimits();
rlimit->set_type(RLimitInfo::RLimit::RLMT_CORE);
container->mutable_rlimit_info()->CopyFrom(rlimitInfo);
Future<TaskStatus> statusRunning;
Future<TaskStatus> statusFinal;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&statusRunning))
.WillOnce(FutureArg<1>(&statusFinal));
driver.launchTasks(offers.get()[0].id(), {task});
AWAIT_READY(statusRunning);
EXPECT_EQ(task.task_id(), statusRunning->task_id());
EXPECT_EQ(TASK_RUNNING, statusRunning->state());
AWAIT_READY(statusFinal);
EXPECT_EQ(task.task_id(), statusFinal->task_id());
EXPECT_EQ(TASK_FINISHED, statusFinal->state());
driver.stop();
driver.join();
}
// In this test, the agent initially doesn't enable disk isolation
// but then restarts with XFS disk isolation enabled. We verify that
// the old container launched before the agent restart is
// successfully recovered.
TEST_F(ROOT_XFS_QuotaTest, RecoverOldContainers)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
Owned<MasterDetector> detector = master.get()->createDetector();
slave::Flags flags = CreateSlaveFlags();
// `CreateSlaveFlags()` enables `disk/xfs` so here we reset
// `isolation` to empty.
flags.isolation.clear();
Try<Owned<cluster::Slave>> slave = StartSlave(detector.get(), flags);
ASSERT_SOME(slave);
FrameworkInfo frameworkInfo = DEFAULT_FRAMEWORK_INFO;
frameworkInfo.set_checkpoint(true);
MockScheduler sched;
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];
TaskInfo task = createTask(
offer.slave_id(),
Resources::parse("cpus:1;mem:128;disk:1").get(),
"dd if=/dev/zero of=file bs=1024 count=1; sleep 1000");
Future<TaskStatus> startingStatus;
Future<TaskStatus> runningstatus;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&startingStatus))
.WillOnce(FutureArg<1>(&runningstatus));
driver.launchTasks(offer.id(), {task});
AWAIT_READY(startingStatus);
EXPECT_EQ(task.task_id(), startingStatus->task_id());
EXPECT_EQ(TASK_STARTING, startingStatus->state());
AWAIT_READY(runningstatus);
EXPECT_EQ(task.task_id(), runningstatus->task_id());
EXPECT_EQ(TASK_RUNNING, runningstatus->state());
{
Future<ResourceUsage> usage =
process::dispatch(slave.get()->pid, &Slave::usage);
AWAIT_READY(usage);
// We should have 1 executor using resources but it doesn't have
// disk limit enabled.
ASSERT_EQ(1, usage->executors().size());
const ResourceUsage_Executor& executor = usage->executors().Get(0);
ASSERT_TRUE(executor.has_statistics());
ASSERT_FALSE(executor.statistics().has_disk_limit_bytes());
}
// Restart the slave.
slave.get()->terminate();
Future<SlaveReregisteredMessage> slaveReregisteredMessage =
FUTURE_PROTOBUF(SlaveReregisteredMessage(), _, _);
// This time use the agent flags that include XFS disk isolation.
slave = StartSlave(detector.get(), CreateSlaveFlags());
ASSERT_SOME(slave);
// Wait for the slave to re-register.
AWAIT_READY(slaveReregisteredMessage);
{
Future<ResourceUsage> usage =
process::dispatch(slave.get()->pid, &Slave::usage);
AWAIT_READY(usage);
// We should still have 1 executor using resources but it doesn't
// have disk limit enabled.
ASSERT_EQ(1, usage->executors().size());
const ResourceUsage_Executor& executor = usage->executors().Get(0);
ASSERT_TRUE(executor.has_statistics());
ASSERT_FALSE(executor.statistics().has_disk_limit_bytes());
}
driver.stop();
driver.join();
}
// This test ensures that the command executor sends TASK_KILLING
// to frameworks that support the capability.
TEST_P(CommandExecutorTest, TaskKillingCapability)
{
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());
// 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();
}
// Verify that we can get accurate resource statistics from the XFS
// disk isolator.
TEST_F(ROOT_XFS_QuotaTest, ResourceStatistics)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
slave::Flags flags = CreateSlaveFlags();
Fetcher fetcher(flags);
Owned<MasterDetector> detector = master.get()->createDetector();
Try<MesosContainerizer*> _containerizer =
MesosContainerizer::create(flags, true, &fetcher);
ASSERT_SOME(_containerizer);
Owned<MesosContainerizer> containerizer(_containerizer.get());
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);
ASSERT_FALSE(offers->empty());
Offer offer = offers.get()[0];
// Create a task that uses 4 of 3MB disk but doesn't fail. We will verify
// that the allocated disk is filled.
TaskInfo task = createTask(
offer.slave_id(),
Resources::parse("cpus:1;mem:128;disk:3").get(),
"dd if=/dev/zero of=file bs=1048576 count=4 || sleep 1000");
Future<TaskStatus> startingStatus;
Future<TaskStatus> runningStatus;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&startingStatus))
.WillOnce(FutureArg<1>(&runningStatus))
.WillRepeatedly(Return()); // Ignore subsequent updates.
driver.launchTasks(offers.get()[0].id(), {task});
AWAIT_READY(startingStatus);
EXPECT_EQ(task.task_id(), startingStatus->task_id());
EXPECT_EQ(TASK_STARTING, startingStatus->state());
AWAIT_READY(runningStatus);
EXPECT_EQ(task.task_id(), runningStatus->task_id());
EXPECT_EQ(TASK_RUNNING, runningStatus->state());
Future<hashset<ContainerID>> containers = containerizer.get()->containers();
AWAIT_READY(containers);
ASSERT_EQ(1u, containers->size());
ContainerID containerId = *(containers->begin());
Timeout timeout = Timeout::in(Seconds(5));
while (true) {
Future<ResourceStatistics> usage = containerizer.get()->usage(containerId);
AWAIT_READY(usage);
ASSERT_TRUE(usage->has_disk_limit_bytes());
EXPECT_EQ(Megabytes(3), Bytes(usage->disk_limit_bytes()));
if (usage->has_disk_used_bytes()) {
// Usage must always be <= the limit.
EXPECT_LE(usage->disk_used_bytes(), usage->disk_limit_bytes());
// Usage might not be equal to the limit, but it must hit
// and not exceed the limit.
if (usage->disk_used_bytes() >= usage->disk_limit_bytes()) {
EXPECT_EQ(
usage->disk_used_bytes(), usage->disk_limit_bytes());
EXPECT_EQ(Megabytes(3), Bytes(usage->disk_used_bytes()));
break;
}
}
ASSERT_FALSE(timeout.expired());
os::sleep(Milliseconds(100));
}
driver.stop();
driver.join();
}
// In this test, the framework is not checkpointed. This ensures that when we
// stop the slave, the executor is killed and we will need to recover the
// working directories without getting any checkpointed recovery state.
TEST_F(ROOT_XFS_QuotaTest, NoCheckpointRecovery)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
slave::Flags flags = CreateSlaveFlags();
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;
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);
ASSERT_FALSE(offers->empty());
Offer offer = offers.get()[0];
TaskInfo task = createTask(
offer.slave_id(),
Resources::parse("cpus:1;mem:128;disk:1").get(),
"dd if=/dev/zero of=file bs=1048576 count=1; sleep 1000");
Future<TaskStatus> runningStatus;
Future<TaskStatus> startingStatus;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&startingStatus))
.WillOnce(FutureArg<1>(&runningStatus))
.WillOnce(Return());
driver.launchTasks(offer.id(), {task});
AWAIT_READY(startingStatus);
EXPECT_EQ(task.task_id(), startingStatus->task_id());
EXPECT_EQ(TASK_STARTING, startingStatus->state());
AWAIT_READY(runningStatus);
EXPECT_EQ(task.task_id(), runningStatus->task_id());
EXPECT_EQ(TASK_RUNNING, runningStatus->state());
Future<ResourceUsage> usage1 =
process::dispatch(slave.get()->pid, &Slave::usage);
AWAIT_READY(usage1);
// We should have 1 executor using resources.
ASSERT_EQ(1, usage1->executors().size());
Future<hashset<ContainerID>> containers = containerizer->containers();
AWAIT_READY(containers);
ASSERT_EQ(1u, containers->size());
ContainerID containerId = *containers->begin();
// Restart the slave.
slave.get()->terminate();
Future<SlaveReregisteredMessage> slaveReregisteredMessage =
FUTURE_PROTOBUF(SlaveReregisteredMessage(), _, _);
_containerizer = MesosContainerizer::create(flags, true, &fetcher);
ASSERT_SOME(_containerizer);
containerizer.reset(_containerizer.get());
slave = StartSlave(detector.get(), containerizer.get(), flags);
ASSERT_SOME(slave);
// Wait until slave recovery is complete.
Future<Nothing> _recover = FUTURE_DISPATCH(_, &Slave::_recover);
AWAIT_READY_FOR(_recover, Seconds(60));
// Wait until the orphan containers are cleaned up.
AWAIT_READY_FOR(containerizer.get()->wait(containerId), Seconds(60));
AWAIT_READY(slaveReregisteredMessage);
Future<ResourceUsage> usage2 =
process::dispatch(slave.get()->pid, &Slave::usage);
AWAIT_READY(usage2);
// We should have no executors left because we didn't checkpoint.
ASSERT_TRUE(usage2->executors().empty());
Try<std::list<string>> sandboxes = os::glob(path::join(
slave::paths::getSandboxRootDir(mountPoint.get()),
"*",
"frameworks",
"*",
"executors",
"*",
"runs",
"*"));
ASSERT_SOME(sandboxes);
// One sandbox and one symlink.
ASSERT_EQ(2u, sandboxes->size());
// Scan the remaining sandboxes and make sure that no projects are assigned.
foreach (const string& sandbox, sandboxes.get()) {
// Skip the "latest" symlink.
if (os::stat::islink(sandbox)) {
continue;
}
EXPECT_NONE(xfs::getProjectId(sandbox));
}
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_registry = "https://registry-1.docker.io";
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();
}
// Verify that a task that tries to consume more space than it has requested
// is only allowed to consume exactly the assigned resources. We tell dd
// to write 2MB but only give it 1MB of resources and (roughly) verify that
// it exits with a failure (that should be a write error).
TEST_F(ROOT_XFS_QuotaTest, DiskUsageExceedsQuota)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
Owned<MasterDetector> detector = master.get()->createDetector();
Try<Owned<cluster::Slave>> slave =
StartSlave(detector.get(), CreateSlaveFlags());
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());
const Offer& offer = offers.get()[0];
// Create a task which requests 1MB disk, but actually uses more
// than 2MB disk.
TaskInfo task = createTask(
offer.slave_id(),
Resources::parse("cpus:1;mem:128;disk:1").get(),
"dd if=/dev/zero of=file bs=1048576 count=2");
Future<TaskStatus> startingStatus;
Future<TaskStatus> runningStatus;
Future<TaskStatus> failedStatus;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&startingStatus))
.WillOnce(FutureArg<1>(&runningStatus))
.WillOnce(FutureArg<1>(&failedStatus));
driver.launchTasks(offer.id(), {task});
AWAIT_READY(startingStatus);
EXPECT_EQ(task.task_id(), startingStatus->task_id());
EXPECT_EQ(TASK_STARTING, startingStatus->state());
AWAIT_READY(runningStatus);
EXPECT_EQ(task.task_id(), runningStatus->task_id());
EXPECT_EQ(TASK_RUNNING, runningStatus->state());
AWAIT_READY(failedStatus);
EXPECT_EQ(task.task_id(), failedStatus->task_id());
EXPECT_EQ(TASK_FAILED, failedStatus->state());
// Unlike the 'disk/du' isolator, the reason for task failure
// should be that dd got an IO error.
EXPECT_EQ(TaskStatus::SOURCE_EXECUTOR, failedStatus->source());
EXPECT_EQ("Command exited with status 1", failedStatus->message());
driver.stop();
driver.join();
}
// This test verifies that the master reconciles tasks that are
// missing from a reregistering slave. In this case, we trigger
// a race between the slave re-registration message and the launch
// message. There should be no TASK_LOST / TASK_DROPPED.
// This was motivated by MESOS-1696.
TEST_F(MasterSlaveReconciliationTest, ReconcileRace)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
MockExecutor exec(DEFAULT_EXECUTOR_ID);
TestContainerizer containerizer(&exec);
StandaloneMasterDetector detector(master.get()->pid);
Future<SlaveRegisteredMessage> slaveRegisteredMessage =
FUTURE_PROTOBUF(SlaveRegisteredMessage(), master.get()->pid, _);
Try<Owned<cluster::Slave>> slave = StartSlave(&detector, &containerizer);
ASSERT_SOME(slave);
AWAIT_READY(slaveRegisteredMessage);
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();
// Since the agent may have retried registration, we want to
// ensure that any duplicate registrations are flushed before
// we appoint the master again. Otherwise, the agent may
// receive a stale registration message.
Clock::pause();
Clock::settle();
Clock::resume();
// Trigger a re-registration of the slave and capture the message
// so that we can spoof a race with a launch task message.
DROP_PROTOBUFS(ReregisterSlaveMessage(), slave.get()->pid, master.get()->pid);
Future<ReregisterSlaveMessage> reregisterSlaveMessage =
DROP_PROTOBUF(
ReregisterSlaveMessage(),
slave.get()->pid,
master.get()->pid);
detector.appoint(master.get()->pid);
AWAIT_READY(reregisterSlaveMessage);
AWAIT_READY(offers);
ASSERT_FALSE(offers->empty());
TaskInfo task;
task.set_name("test task");
task.mutable_task_id()->set_value("1");
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);
ExecutorDriver* executorDriver;
EXPECT_CALL(exec, registered(_, _, _, _))
.WillOnce(SaveArg<0>(&executorDriver));
// Leave the task in TASK_STAGING.
Future<Nothing> launchTask;
EXPECT_CALL(exec, launchTask(_, _))
.WillOnce(FutureSatisfy(&launchTask));
EXPECT_CALL(sched, statusUpdate(&driver, _))
.Times(0);
driver.launchTasks(offers.get()[0].id(), {task});
AWAIT_READY(launchTask);
// Send the stale re-registration message, which does not contain
// the task we just launched. This will trigger a reconciliation
// by the master.
Future<SlaveReregisteredMessage> slaveReregisteredMessage =
FUTURE_PROTOBUF(SlaveReregisteredMessage(), _, _);
// Prevent this from being dropped per the DROP_PROTOBUFS above.
FUTURE_PROTOBUF(
ReregisterSlaveMessage(),
slave.get()->pid,
master.get()->pid);
process::post(
slave.get()->pid,
master.get()->pid,
reregisterSlaveMessage.get());
AWAIT_READY(slaveReregisteredMessage);
// Neither the master nor the slave should send a TASK_LOST
// as part of the reconciliation. We check this by calling
// Clock::settle() to flush all pending events.
Clock::pause();
Clock::settle();
Clock::resume();
// Now send TASK_FINISHED and make sure it's the only message
// received by the scheduler.
Future<TaskStatus> status;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&status));
TaskStatus taskStatus;
taskStatus.mutable_task_id()->CopyFrom(task.task_id());
taskStatus.set_state(TASK_FINISHED);
executorDriver->sendStatusUpdate(taskStatus);
AWAIT_READY(status);
ASSERT_EQ(TASK_FINISHED, status->state());
EXPECT_CALL(exec, shutdown(_))
.Times(AtMost(1));
driver.stop();
driver.join();
}
// This test verifies that when the slave reregisters, the master
// does not send TASK_LOST update for a task that has reached terminal
// state but is waiting for an acknowledgement.
TEST_F(MasterSlaveReconciliationTest, SlaveReregisterTerminalTask)
{
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);
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());
TaskInfo task;
task.set_name("test task");
task.mutable_task_id()->set_value("1");
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);
EXPECT_CALL(exec, registered(_, _, _, _));
// Send a terminal update right away.
EXPECT_CALL(exec, launchTask(_, _))
.WillOnce(SendStatusUpdateFromTask(TASK_FINISHED));
// Drop the status update from slave to the master, so that
// the slave has a pending terminal update when it reregisters.
DROP_PROTOBUF(StatusUpdateMessage(), _, master.get()->pid);
Future<Nothing> _statusUpdate = FUTURE_DISPATCH(_, &Slave::_statusUpdate);
Future<TaskStatus> status;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&status))
.WillRepeatedly(Return()); // Ignore retried update due to update framework.
driver.launchTasks(offers.get()[0].id(), {task});
AWAIT_READY(_statusUpdate);
Future<SlaveReregisteredMessage> slaveReregisteredMessage =
FUTURE_PROTOBUF(SlaveReregisteredMessage(), _, _);
// Simulate a spurious master change event (e.g., due to ZooKeeper
// expiration) at the slave to force re-registration.
detector.appoint(master.get()->pid);
AWAIT_READY(slaveReregisteredMessage);
// The master should not send a TASK_LOST after the slave
// reregisters. We check this by calling Clock::settle() so that
// the only update the scheduler receives is the retried
// TASK_FINISHED update.
// NOTE: The task status update manager resends the status update
// when it detects a new master.
Clock::pause();
Clock::settle();
AWAIT_READY(status);
ASSERT_EQ(TASK_FINISHED, status->state());
EXPECT_CALL(exec, shutdown(_))
.Times(AtMost(1));
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<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 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.
// Ensure no status update acknowledgements are sent to the master.
EXPECT_NO_FUTURE_CALLS(
mesos::scheduler::Call(),
mesos::scheduler::Call::ACKNOWLEDGE,
_ ,
master.get()->pid);
driver.start();
AWAIT_READY(offers);
EXPECT_NE(0u, offers->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->state());
ASSERT_EQ(TaskStatus::SOURCE_MASTER, status->source());
ASSERT_EQ(TaskStatus::REASON_TASK_INVALID, status->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();
}
TEST_P(MemoryIsolatorTest, ROOT_MemUsage)
{
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());
TaskInfo task = createTask(offers.get()[0], "sleep 120");
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());
Future<ResourceStatistics> usage = containerizer->usage(containerId);
AWAIT_READY(usage);
// TODO(jieyu): Consider using a program that predictably increases
// RSS so that we can set more meaningful expectation here.
EXPECT_LT(0u, usage->mem_rss_bytes());
driver.stop();
driver.join();
}
// 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";
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;
// 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);
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;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&running));
Future<Nothing> _statusUpdateAcknowledgement =
FUTURE_DISPATCH(_, &Slave::_statusUpdateAcknowledgement);
driver.launchTasks(offers.get()[0].id(), {task});
AWAIT_READY(running);
EXPECT_EQ(task.task_id(), running.get().task_id());
EXPECT_EQ(TASK_RUNNING, running.get().state());
// Wait for the ACK to be checkpointed.
AWAIT_READY(_statusUpdateAcknowledgement);
// 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.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();
Future<TaskStatus> killed;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&killed));
// Stop the memory-hammering task.
driver.killTask(task.task_id());
AWAIT_READY(killed);
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 checks the behavior of passed invalid limits.
TEST_F(PosixRLimitsIsolatorTest, InvalidLimits)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
slave::Flags flags = CreateSlaveFlags();
flags.isolation = "posix/rlimits";
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(_, _, _));
Future<vector<Offer>> offers;
EXPECT_CALL(sched, resourceOffers(_, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driver.start();
AWAIT_READY(offers);
ASSERT_NE(0u, offers->size());
TaskInfo task = createTask(
offers.get()[0].slave_id(),
offers.get()[0].resources(),
"true");
ContainerInfo* container = task.mutable_container();
container->set_type(ContainerInfo::MESOS);
// Set impossible limit soft > hard.
RLimitInfo rlimitInfo;
RLimitInfo::RLimit* rlimit = rlimitInfo.add_rlimits();
rlimit->set_type(RLimitInfo::RLimit::RLMT_CPU);
rlimit->set_soft(100);
rlimit->set_hard(1);
container->mutable_rlimit_info()->CopyFrom(rlimitInfo);
Future<TaskStatus> taskStatus;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&taskStatus));
driver.launchTasks(offers.get()[0].id(), {task});
AWAIT_READY(taskStatus);
EXPECT_EQ(task.task_id(), taskStatus->task_id());
EXPECT_EQ(TASK_FAILED, taskStatus->state());
EXPECT_EQ(TaskStatus::REASON_EXECUTOR_TERMINATED, taskStatus->reason());
driver.stop();
driver.join();
}
// This test confirms that if a task exceeds configured resource
// limits it is forcibly terminated.
TEST_F(PosixRLimitsIsolatorTest, TaskExceedingLimit)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
slave::Flags flags = CreateSlaveFlags();
flags.isolation = "posix/rlimits";
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(_, _, _));
Future<vector<Offer>> offers;
EXPECT_CALL(sched, resourceOffers(_, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driver.start();
AWAIT_READY(offers);
ASSERT_NE(0u, offers->size());
// The task attempts to use an infinite amount of CPU time.
TaskInfo task = createTask(
offers.get()[0].slave_id(),
offers.get()[0].resources(),
"while true; do true; done");
ContainerInfo* container = task.mutable_container();
container->set_type(ContainerInfo::MESOS);
// Limit the process to use maximally 1 second of CPU time.
RLimitInfo rlimitInfo;
RLimitInfo::RLimit* cpuLimit = rlimitInfo.add_rlimits();
cpuLimit->set_type(RLimitInfo::RLimit::RLMT_CPU);
cpuLimit->set_soft(1);
cpuLimit->set_hard(1);
container->mutable_rlimit_info()->CopyFrom(rlimitInfo);
Future<TaskStatus> statusRunning;
Future<TaskStatus> statusFailed;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&statusRunning))
.WillOnce(FutureArg<1>(&statusFailed));
driver.launchTasks(offers.get()[0].id(), {task});
AWAIT_READY(statusRunning);
EXPECT_EQ(task.task_id(), statusRunning->task_id());
EXPECT_EQ(TASK_RUNNING, statusRunning->state());
AWAIT_READY(statusFailed);
EXPECT_EQ(task.task_id(), statusFailed->task_id());
EXPECT_EQ(TASK_FAILED, statusFailed->state());
driver.stop();
driver.join();
}
// This test verifies that the framework can launch a command task
// that specifies both container image and persistent volumes.
TEST_F(LinuxFilesystemIsolatorMesosTest,
ROOT_ChangeRootFilesystemCommandExecutorPersistentVolume)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
string registry = path::join(sandbox.get(), "registry");
AWAIT_READY(DockerArchive::create(registry, "test_image"));
slave::Flags flags = CreateSlaveFlags();
flags.resources = "cpus:2;mem:1024;disk(role1):1024";
flags.isolation = "filesystem/linux,docker/runtime";
flags.docker_registry = registry;
flags.docker_store_dir = path::join(sandbox.get(), "store");
flags.image_providers = "docker";
Owned<MasterDetector> detector = master.get()->createDetector();
Try<Owned<cluster::Slave>> slave = StartSlave(detector.get(), flags);
ASSERT_SOME(slave);
MockScheduler sched;
FrameworkInfo frameworkInfo = DEFAULT_FRAMEWORK_INFO;
frameworkInfo.set_roles(0, "role1");
MesosSchedulerDriver driver(
&sched,
frameworkInfo,
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()); // Ignore subsequent offers.
driver.start();
AWAIT_READY(frameworkId);
AWAIT_READY(offers);
ASSERT_FALSE(offers->empty());
Offer offer = offers.get()[0];
string dir1 = path::join(sandbox.get(), "dir1");
ASSERT_SOME(os::mkdir(dir1));
Resource persistentVolume = createPersistentVolume(
Megabytes(64),
"role1",
"id1",
"path1",
None(),
None(),
frameworkInfo.principal());
// We use the filter explicitly here so that the resources will not
// be filtered for 5 seconds (the default).
Filters filters;
filters.set_refuse_seconds(0);
TaskInfo task = createTask(
offer.slave_id(),
Resources::parse("cpus:1;mem:512").get() + persistentVolume,
"echo abc > path1/file");
task.mutable_container()->CopyFrom(createContainerInfo(
"test_image",
{createVolumeHostPath("/tmp", dir1, Volume::RW)}));
// Create the persistent volumes and launch task via `acceptOffers`.
driver.acceptOffers(
{offer.id()},
{CREATE(persistentVolume), LAUNCH({task})},
filters);
Future<TaskStatus> statusStarting;
Future<TaskStatus> statusRunning;
Future<TaskStatus> statusFinished;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&statusStarting))
.WillOnce(FutureArg<1>(&statusRunning))
.WillOnce(FutureArg<1>(&statusFinished));
AWAIT_READY(statusStarting);
EXPECT_EQ(TASK_STARTING, statusStarting->state());
AWAIT_READY(statusRunning);
EXPECT_EQ(TASK_RUNNING, statusRunning->state());
AWAIT_READY(statusFinished);
EXPECT_EQ(TASK_FINISHED, statusFinished->state());
// NOTE: The command executor's id is the same as the task id.
ExecutorID executorId;
executorId.set_value(task.task_id().value());
string directory = slave::paths::getExecutorLatestRunPath(
flags.work_dir,
offer.slave_id(),
frameworkId.get(),
executorId);
EXPECT_FALSE(os::exists(path::join(directory, "path1")));
string volumePath = slave::paths::getPersistentVolumePath(
flags.work_dir,
"role1",
"id1");
EXPECT_SOME_EQ("abc\n", os::read(path::join(volumePath, "file")));
driver.stop();
driver.join();
}
// This test verifies that when the slave reregisters, we correctly
// send the information about actively running frameworks.
TEST_F(MasterSlaveReconciliationTest, SlaveReregisterFrameworks)
{
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);
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());
TaskInfo task;
task.set_name("test task");
task.mutable_task_id()->set_value("1");
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);
EXPECT_CALL(exec, registered(_, _, _, _));
// Send an update right away.
EXPECT_CALL(exec, launchTask(_, _))
.WillOnce(SendStatusUpdateFromTask(TASK_RUNNING));
Future<TaskStatus> status;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&status))
.WillRepeatedly(Return()); // Ignore retried update due to update framework.
driver.launchTasks(offers.get()[0].id(), {task});
// Wait until TASK_RUNNING of the task is received.
AWAIT_READY(status);
EXPECT_EQ(TASK_RUNNING, status->state());
Future<ReregisterSlaveMessage> reregisterSlave =
FUTURE_PROTOBUF(ReregisterSlaveMessage(), _, _);
// Simulate a spurious master change event (e.g., due to ZooKeeper
// expiration) at the slave to force re-registration.
detector.appoint(master.get()->pid);
// Expect to receive the 'ReregisterSlaveMessage' containing the
// active frameworks.
AWAIT_READY(reregisterSlave);
EXPECT_EQ(1, reregisterSlave->frameworks().size());
EXPECT_CALL(exec, shutdown(_))
.Times(AtMost(1));
driver.stop();
driver.join();
}
// This test verifies that persistent volumes are unmounted properly
// after a checkpointed framework disappears and the slave restarts.
//
// TODO(jieyu): Even though the command task specifies a new
// filesystem root, the executor (command executor) itself does not
// change filesystem root (uses the host filesystem). We need to add a
// test to test the scenario that the executor itself changes rootfs.
TEST_F(LinuxFilesystemIsolatorMesosTest,
ROOT_RecoverOrphanedPersistentVolume)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
string registry = path::join(sandbox.get(), "registry");
AWAIT_READY(DockerArchive::create(registry, "test_image"));
slave::Flags flags = CreateSlaveFlags();
flags.resources = "cpus:2;mem:1024;disk(role1):1024";
flags.isolation = "filesystem/linux,docker/runtime";
flags.docker_registry = registry;
flags.docker_store_dir = path::join(sandbox.get(), "store");
flags.image_providers = "docker";
Fetcher fetcher(flags);
Try<MesosContainerizer*> create =
MesosContainerizer::create(flags, true, &fetcher);
ASSERT_SOME(create);
Owned<Containerizer> containerizer(create.get());
Owned<MasterDetector> detector = master.get()->createDetector();
Try<Owned<cluster::Slave>> slave = StartSlave(
detector.get(),
containerizer.get(),
flags);
ASSERT_SOME(slave);
MockScheduler sched;
FrameworkInfo frameworkInfo = DEFAULT_FRAMEWORK_INFO;
frameworkInfo.set_roles(0, "role1");
frameworkInfo.set_checkpoint(true);
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);
ASSERT_FALSE(offers->empty());
Offer offer = offers.get()[0];
string dir1 = path::join(sandbox.get(), "dir1");
ASSERT_SOME(os::mkdir(dir1));
Resource persistentVolume = createPersistentVolume(
Megabytes(64),
"role1",
"id1",
"path1",
None(),
None(),
frameworkInfo.principal());
// Create a task that does nothing for a long time.
TaskInfo task = createTask(
offer.slave_id(),
Resources::parse("cpus:1;mem:512").get() + persistentVolume,
"sleep 1000");
task.mutable_container()->CopyFrom(createContainerInfo(
"test_image",
{createVolumeHostPath("/tmp", dir1, Volume::RW)}));
Future<TaskStatus> statusStarting;
Future<TaskStatus> statusRunning;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&statusStarting))
.WillOnce(FutureArg<1>(&statusRunning))
.WillRepeatedly(DoDefault());
Future<Nothing> ack =
FUTURE_DISPATCH(_, &Slave::_statusUpdateAcknowledgement);
// Create the persistent volumes and launch task via `acceptOffers`.
driver.acceptOffers(
{offer.id()},
{CREATE(persistentVolume), LAUNCH({task})});
AWAIT_READY(statusStarting);
EXPECT_EQ(TASK_STARTING, statusStarting->state());
AWAIT_READY(statusRunning);
EXPECT_EQ(TASK_RUNNING, statusRunning->state());
// Wait for the ACK to be checkpointed.
AWAIT_READY(ack);
Future<hashset<ContainerID>> containers = containerizer->containers();
AWAIT_READY(containers);
ASSERT_EQ(1u, containers->size());
ContainerID containerId = *containers->begin();
// Restart the slave.
slave.get()->terminate();
// Wipe the slave meta directory so that the slave will treat the
// above running task as an orphan.
ASSERT_SOME(os::rmdir(slave::paths::getMetaRootDir(flags.work_dir)));
Future<Nothing> _recover = FUTURE_DISPATCH(_, &Slave::_recover);
// Recreate the containerizer using the same helper as above.
containerizer.reset();
create = MesosContainerizer::create(flags, true, &fetcher);
ASSERT_SOME(create);
containerizer.reset(create.get());
slave = StartSlave(detector.get(), containerizer.get(), flags);
ASSERT_SOME(slave);
// Wait until slave recovery is complete.
AWAIT_READY(_recover);
// Wait until the orphan containers are cleaned up.
AWAIT_READY(containerizer->wait(containerId));
Try<fs::MountInfoTable> table = fs::MountInfoTable::read();
ASSERT_SOME(table);
// All mount targets should be under this directory.
string directory = slave::paths::getSandboxRootDir(flags.work_dir);
// Verify that the orphaned container's persistent volume and
// the rootfs are unmounted.
foreach (const fs::MountInfoTable::Entry& entry, table->entries) {
EXPECT_FALSE(strings::contains(entry.target, directory))
<< "Target was not unmounted: " << entry.target;
}
driver.stop();
driver.join();
}
EXPECT_TRUE(result.contains(expected));
}
driver1.stop();
driver1.join();
Future<TaskStatus> status2;
EXPECT_CALL(sched2, statusUpdate(&driver2, _))
.WillOnce(FutureArg<1>(&status2));
// Trigger explicit reconciliation.
driver2.reconcileTasks({status.get()});
AWAIT_READY(status2);
EXPECT_EQ(TASK_RUNNING, status2->state());
EXPECT_CALL(exec, shutdown(_))
.Times(AtMost(1));
driver2.stop();
driver2.join();
}
// Checks that resources of an agent will be offered to non-hierarchical roles
// before/after being upgraded to support HIERARCHICAL_ROLE. We first strip
// HIERARCHICAL_ROLE capability in `registerSlaveMessage` to spoof an old agent.
// Then we simulate a master detection event to trigger agent re-registration.
// Before/after 'upgrade', resources of the agent should be offered to
// non-hierarchical roles.
// 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 authorization based endpoint filtering
// works correctly on the /state endpoint.
// Both default users are allowed to view high level frameworks, but only
// one is allowed to view the tasks.
// After launching a single task per each framework, one for role "superhero"
// and the other for role "muggle", this test verifies that each of two
// default users can view resource allocations and resource reservations for
// corresponding allowed roles only.
TYPED_TEST(SlaveAuthorizerTest, FilterStateEndpoint)
{
ACLs acls;
const string roleSuperhero = "superhero";
const string roleMuggle = "muggle";
{
// Default principal can see all frameworks.
mesos::ACL::ViewFramework* acl = acls.add_view_frameworks();
acl->mutable_principals()->add_values(DEFAULT_CREDENTIAL.principal());
acl->mutable_users()->set_type(ACL::Entity::ANY);
}
{
// Second default principal can see all frameworks.
mesos::ACL::ViewFramework* acl = acls.add_view_frameworks();
acl->mutable_principals()->add_values(DEFAULT_CREDENTIAL_2.principal());
acl->mutable_users()->set_type(ACL::Entity::ANY);
}
{
// No other principal can see frameworks running under any user.
ACL::ViewFramework* acl = acls.add_view_frameworks();
acl->mutable_principals()->set_type(ACL::Entity::ANY);
acl->mutable_users()->set_type(ACL::Entity::NONE);
}
{
// Default principal can see all executors.
mesos::ACL::ViewExecutor* acl = acls.add_view_executors();
acl->mutable_principals()->add_values(DEFAULT_CREDENTIAL.principal());
acl->mutable_users()->set_type(ACL::Entity::ANY);
}
{
// No other principal can see executors running under any user.
ACL::ViewExecutor* acl = acls.add_view_executors();
acl->mutable_principals()->set_type(ACL::Entity::ANY);
acl->mutable_users()->set_type(ACL::Entity::NONE);
}
{
// Default principal can see all tasks.
mesos::ACL::ViewTask* acl = acls.add_view_tasks();
acl->mutable_principals()->add_values(DEFAULT_CREDENTIAL.principal());
acl->mutable_users()->set_type(ACL::Entity::ANY);
}
{
// No other principal can see tasks running under any user.
ACL::ViewTask* acl = acls.add_view_tasks();
acl->mutable_principals()->set_type(ACL::Entity::ANY);
acl->mutable_users()->set_type(ACL::Entity::NONE);
}
{
// Default principal can view "superhero" role only.
ACL::ViewRole* acl = acls.add_view_roles();
acl->mutable_principals()->add_values(DEFAULT_CREDENTIAL.principal());
acl->mutable_roles()->add_values(roleSuperhero);
acl = acls.add_view_roles();
acl->mutable_principals()->add_values(DEFAULT_CREDENTIAL.principal());
acl->mutable_roles()->set_type(mesos::ACL::Entity::NONE);
}
{
// Second default principal can view "muggle" role only.
ACL::ViewRole* acl = acls.add_view_roles();
acl->mutable_principals()->add_values(DEFAULT_CREDENTIAL_2.principal());
acl->mutable_roles()->add_values(roleMuggle);
acl = acls.add_view_roles();
acl->mutable_principals()->add_values(DEFAULT_CREDENTIAL_2.principal());
acl->mutable_roles()->set_type(mesos::ACL::Entity::NONE);
}
// Create an `Authorizer` with the ACLs.
Try<Authorizer*> create = TypeParam::create(parameterize(acls));
ASSERT_SOME(create);
Owned<Authorizer> authorizer(create.get());
Try<Owned<cluster::Master>> master = this->StartMaster(authorizer.get());
ASSERT_SOME(master);
// Register framework with user "bar" and role "superhero".
FrameworkInfo frameworkSuperhero = DEFAULT_FRAMEWORK_INFO;
frameworkSuperhero.set_name("framework-" + roleSuperhero);
frameworkSuperhero.set_roles(0, roleSuperhero);
frameworkSuperhero.set_user("bar");
// Create an executor with user "bar".
ExecutorInfo executorSuperhero =
createExecutorInfo("test-executor-" + roleSuperhero, "sleep 2");
executorSuperhero.mutable_command()->set_user("bar");
MockExecutor execSuperhero(executorSuperhero.executor_id());
// Register framework with user "foo" and role "muggle".
FrameworkInfo frameworkMuggle = DEFAULT_FRAMEWORK_INFO;
frameworkMuggle.set_name("framework-" + roleMuggle);
frameworkMuggle.set_principal(DEFAULT_CREDENTIAL_2.principal());
frameworkMuggle.set_roles(0, roleMuggle);
frameworkMuggle.set_user("foo");
// Create an executor with user "foo".
ExecutorInfo executorMuggle =
createExecutorInfo("test-executor-" + roleMuggle, "sleep 2");
executorMuggle.mutable_command()->set_user("foo");
MockExecutor execMuggle(executorMuggle.executor_id());
TestContainerizer containerizer(
{{executorSuperhero.executor_id(), &execSuperhero},
{executorMuggle.executor_id(), &execMuggle}});
slave::Flags flags = this->CreateSlaveFlags();
// Statically reserve resources for each role.
flags.resources = "cpus(" + roleSuperhero + "):2;" + "cpus(" + roleMuggle +
"):3;mem(" + roleSuperhero + "):512;" + "mem(" + roleMuggle + "):1024;";
Owned<MasterDetector> detector = master.get()->createDetector();
Try<Owned<cluster::Slave>> slave = this->StartSlave(
detector.get(), &containerizer, authorizer.get(), flags);
ASSERT_SOME(slave);
MockScheduler schedSuperhero;
MesosSchedulerDriver driverSuperhero(
&schedSuperhero,
frameworkSuperhero,
master.get()->pid,
DEFAULT_CREDENTIAL);
EXPECT_CALL(execSuperhero, registered(_, _, _, _))
.Times(AtMost(1));
Future<FrameworkID> frameworkIdSuperhero;
EXPECT_CALL(schedSuperhero, registered(&driverSuperhero, _, _))
.WillOnce(FutureArg<1>(&frameworkIdSuperhero));
Future<vector<Offer>> offersSuperhero;
EXPECT_CALL(schedSuperhero, resourceOffers(&driverSuperhero, _))
.WillOnce(FutureArg<1>(&offersSuperhero))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driverSuperhero.start();
AWAIT_READY(frameworkIdSuperhero);
AWAIT_READY(offersSuperhero);
ASSERT_FALSE(offersSuperhero->empty());
// Define a task which will run on executorSuperhero of frameworkSuperhero.
TaskInfo taskSuperhero;
taskSuperhero.set_name("test-" + roleSuperhero);
taskSuperhero.mutable_task_id()->set_value("1");
taskSuperhero.mutable_slave_id()->MergeFrom(
offersSuperhero.get()[0].slave_id());
taskSuperhero.mutable_resources()->MergeFrom(
offersSuperhero.get()[0].resources());
taskSuperhero.mutable_executor()->MergeFrom(executorSuperhero);
EXPECT_CALL(execSuperhero, launchTask(_, _))
.WillOnce(SendStatusUpdateFromTask(TASK_RUNNING))
.WillRepeatedly(Return());
Future<TaskStatus> statusSuperhero;
EXPECT_CALL(schedSuperhero, statusUpdate(&driverSuperhero, _))
.WillOnce(FutureArg<1>(&statusSuperhero));
driverSuperhero.launchTasks(offersSuperhero.get()[0].id(), {taskSuperhero});
AWAIT_READY(statusSuperhero);
EXPECT_EQ(TASK_RUNNING, statusSuperhero->state());
MockScheduler schedMuggle;
MesosSchedulerDriver driverMuggle(
&schedMuggle,
frameworkMuggle,
master.get()->pid,
DEFAULT_CREDENTIAL_2);
EXPECT_CALL(execMuggle, registered(_, _, _, _))
.Times(AtMost(1));
Future<FrameworkID> frameworkIdMuggle;
EXPECT_CALL(schedMuggle, registered(&driverMuggle, _, _))
.WillOnce(FutureArg<1>(&frameworkIdMuggle));
Future<vector<Offer>> offersMuggle;
EXPECT_CALL(schedMuggle, resourceOffers(&driverMuggle, _))
.WillOnce(FutureArg<1>(&offersMuggle))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driverMuggle.start();
AWAIT_READY(frameworkIdMuggle);
AWAIT_READY(offersMuggle);
ASSERT_FALSE(offersMuggle->empty());
// Define a task which will run on executorMuggle of frameworkMuggle.
TaskInfo taskMuggle;
taskMuggle.set_name("test-" + roleMuggle);
taskMuggle.mutable_task_id()->set_value("2");
taskMuggle.mutable_slave_id()->MergeFrom(
offersMuggle.get()[0].slave_id());
taskMuggle.mutable_resources()->MergeFrom(
offersMuggle.get()[0].resources());
taskMuggle.mutable_executor()->MergeFrom(executorMuggle);
EXPECT_CALL(execMuggle, launchTask(_, _))
.WillOnce(SendStatusUpdateFromTask(TASK_RUNNING))
.WillRepeatedly(Return());
Future<TaskStatus> statusMuggle;
EXPECT_CALL(schedMuggle, statusUpdate(&driverMuggle, _))
.WillOnce(FutureArg<1>(&statusMuggle));
driverMuggle.launchTasks(offersMuggle.get()[0].id(), {taskMuggle});
AWAIT_READY(statusMuggle);
ASSERT_EQ(TASK_RUNNING, statusMuggle->state());
// Retrieve endpoint with the user allowed to view the frameworks.
// The default user allowed to view role "superhero" only.
{
Future<Response> response = http::get(
slave.get()->pid,
"state",
None(),
createBasicAuthHeaders(DEFAULT_CREDENTIAL));
AWAIT_EXPECT_RESPONSE_STATUS_EQ(OK().status, response);
Try<JSON::Object> parse = JSON::parse<JSON::Object>(response->body);
ASSERT_SOME(parse);
JSON::Object state = parse.get();
ASSERT_TRUE(state.values["frameworks"].is<JSON::Array>());
JSON::Array frameworks = state.values["frameworks"].as<JSON::Array>();
EXPECT_EQ(2u, frameworks.values.size());
foreach (const JSON::Value& value, frameworks.values) {
JSON::Object framework = value.as<JSON::Object>();
EXPECT_FALSE(framework.values.empty());
ASSERT_TRUE(framework.values["executors"].is<JSON::Array>());
JSON::Array executors = framework.values["executors"].as<JSON::Array>();
EXPECT_EQ(1u, executors.values.size());
JSON::Object executor = executors.values.front().as<JSON::Object>();
EXPECT_EQ(1u, executor.values["tasks"].as<JSON::Array>().values.size());
}
ASSERT_TRUE(state.values["reserved_resources"].is<JSON::Object>());
JSON::Object reserved_resources =
state.values["reserved_resources"].as<JSON::Object>();
EXPECT_TRUE(reserved_resources.values[roleSuperhero].is<JSON::Object>());
EXPECT_FALSE(reserved_resources.values[roleMuggle].is<JSON::Object>());
ASSERT_TRUE(
state.values["reserved_resources_allocated"].is<JSON::Object>());
JSON::Object reserved_resources_allocated =
state.values["reserved_resources_allocated"].as<JSON::Object>();
EXPECT_TRUE(
reserved_resources_allocated.values[roleSuperhero].is<JSON::Object>());
EXPECT_FALSE(
reserved_resources_allocated.values[roleMuggle].is<JSON::Object>());
ASSERT_TRUE(state.values["reserved_resources_full"].is<JSON::Object>());
JSON::Object reserved_resources_full =
state.values["reserved_resources_full"].as<JSON::Object>();
EXPECT_TRUE(
reserved_resources_full.values[roleSuperhero].is<JSON::Array>());
EXPECT_FALSE(
reserved_resources_full.values[roleMuggle].is<JSON::Array>());
}
// Retrieve endpoint with the user allowed to view the frameworks,
// but not the executors.
// The second default user allowed to view role "muggle" only.
{
Future<Response> response = http::get(
slave.get()->pid,
"state",
None(),
createBasicAuthHeaders(DEFAULT_CREDENTIAL_2));
AWAIT_EXPECT_RESPONSE_STATUS_EQ(OK().status, response);
Try<JSON::Object> parse = JSON::parse<JSON::Object>(response->body);
ASSERT_SOME(parse);
JSON::Object state = parse.get();
ASSERT_TRUE(state.values["frameworks"].is<JSON::Array>());
JSON::Array frameworks = state.values["frameworks"].as<JSON::Array>();
EXPECT_EQ(2u, frameworks.values.size());
foreach (const JSON::Value& value, frameworks.values) {
JSON::Object framework = value.as<JSON::Object>();
EXPECT_FALSE(framework.values.empty());
EXPECT_TRUE(
framework.values["executors"].as<JSON::Array>().values.empty());
}
ASSERT_TRUE(state.values["reserved_resources"].is<JSON::Object>());
JSON::Object reserved_resources =
state.values["reserved_resources"].as<JSON::Object>();
EXPECT_TRUE(reserved_resources.values[roleMuggle].is<JSON::Object>());
EXPECT_FALSE(reserved_resources.values[roleSuperhero].is<JSON::Object>());
ASSERT_TRUE(
state.values["reserved_resources_allocated"].is<JSON::Object>());
JSON::Object reserved_resources_allocated =
state.values["reserved_resources_allocated"].as<JSON::Object>();
EXPECT_TRUE(
reserved_resources_allocated.values[roleMuggle].is<JSON::Object>());
EXPECT_FALSE(
reserved_resources_allocated.values[roleSuperhero].is<JSON::Object>());
ASSERT_TRUE(state.values["reserved_resources_full"].is<JSON::Object>());
JSON::Object reserved_resources_full =
state.values["reserved_resources_full"].as<JSON::Object>();
EXPECT_TRUE(
reserved_resources_full.values[roleMuggle].is<JSON::Array>());
EXPECT_FALSE(
reserved_resources_full.values[roleSuperhero].is<JSON::Array>());
}
EXPECT_CALL(execSuperhero, shutdown(_))
.Times(AtMost(1));
EXPECT_CALL(execMuggle, shutdown(_))
.Times(AtMost(1));
driverSuperhero.stop();
driverSuperhero.join();
driverMuggle.stop();
driverMuggle.join();
}
// This is the same logic as DiskUsageExceedsQuota except we turn off disk quota
// enforcement, so exceeding the quota should be allowed.
TEST_F(ROOT_XFS_QuotaTest, DiskUsageExceedsQuotaNoEnforce)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
Owned<MasterDetector> detector = master.get()->createDetector();
slave::Flags flags = CreateSlaveFlags();
flags.enforce_container_disk_quota = false;
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_FALSE(offers->empty());
const Offer& offer = offers.get()[0];
// Create a task which requests 1MB disk, but actually uses more
// than 2MB disk.
TaskInfo task = createTask(
offer.slave_id(),
Resources::parse("cpus:1;mem:128;disk:1").get(),
"dd if=/dev/zero of=file bs=1048576 count=2");
Future<TaskStatus> startingStatus;
Future<TaskStatus> runningStatus;
Future<TaskStatus> finishedStatus;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&startingStatus))
.WillOnce(FutureArg<1>(&runningStatus))
.WillOnce(FutureArg<1>(&finishedStatus));
driver.launchTasks(offer.id(), {task});
AWAIT_READY(startingStatus);
EXPECT_EQ(task.task_id(), startingStatus->task_id());
EXPECT_EQ(TASK_STARTING, startingStatus->state());
AWAIT_READY(runningStatus);
EXPECT_EQ(task.task_id(), runningStatus->task_id());
EXPECT_EQ(TASK_RUNNING, runningStatus->state());
// We expect the task to succeed even though it exceeded
// the disk quota.
AWAIT_READY(finishedStatus);
EXPECT_EQ(task.task_id(), finishedStatus->task_id());
EXPECT_EQ(TASK_FINISHED, finishedStatus->state());
driver.stop();
driver.join();
}
// This test verifies that a task group is launched on the agent if the executor
// provides a valid authentication token specifying its own ContainerID.
TEST_F(ExecutorAuthorizationTest, RunTaskGroup)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
// Start an agent with permissive ACLs so that a task can be launched.
ACLs acls;
acls.set_permissive(true);
slave::Flags flags = CreateSlaveFlags();
flags.acls = acls;
Owned<MasterDetector> detector = master.get()->createDetector();
Try<Owned<cluster::Slave>> slave = StartSlave(detector.get(), flags);
ASSERT_SOME(slave);
FrameworkInfo frameworkInfo = DEFAULT_FRAMEWORK_INFO;
MockScheduler sched;
MesosSchedulerDriver driver(
&sched, frameworkInfo, 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()); // Ignore subsequent offers.
driver.start();
AWAIT_READY(frameworkId);
AWAIT_READY(offers);
ASSERT_FALSE(offers->empty());
Offer offer = offers.get()[0];
TaskInfo task = createTask(
offer.slave_id(),
Resources::parse("cpus:0.5;mem:32").get(),
"sleep 1000");
Future<TaskStatus> status;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&status));
Resources executorResources =
allocatedResources(Resources::parse("cpus:0.1;mem:32;disk:32").get(), "*");
ExecutorInfo executor;
executor.mutable_executor_id()->set_value("default");
executor.set_type(ExecutorInfo::DEFAULT);
executor.mutable_framework_id()->CopyFrom(frameworkId.get());
executor.mutable_resources()->CopyFrom(executorResources);
TaskGroupInfo taskGroup;
taskGroup.add_tasks()->CopyFrom(task);
driver.acceptOffers({offer.id()}, {LAUNCH_GROUP(executor, taskGroup)});
AWAIT_READY(status);
ASSERT_EQ(task.task_id(), status->task_id());
EXPECT_EQ(TASK_STARTING, status->state());
driver.stop();
driver.join();
}
// This is the same logic as ResourceStatistics, except the task should
// be allowed to exceed the disk quota, and usage statistics should report
// that the quota was exceeded.
TEST_F(ROOT_XFS_QuotaTest, ResourceStatisticsNoEnforce)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
slave::Flags flags = CreateSlaveFlags();
flags.enforce_container_disk_quota = false;
Fetcher fetcher(flags);
Owned<MasterDetector> detector = master.get()->createDetector();
Try<MesosContainerizer*> _containerizer =
MesosContainerizer::create(flags, true, &fetcher);
ASSERT_SOME(_containerizer);
Owned<MesosContainerizer> containerizer(_containerizer.get());
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);
ASSERT_FALSE(offers->empty());
Offer offer = offers.get()[0];
// Create a task that uses 4MB of 3MB disk and fails if it can't
// write the full amount.
TaskInfo task = createTask(
offer.slave_id(),
Resources::parse("cpus:1;mem:128;disk:3").get(),
"dd if=/dev/zero of=file bs=1048576 count=4 && sleep 1000");
Future<TaskStatus> startingStatus;
Future<TaskStatus> runningStatus;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&startingStatus))
.WillOnce(FutureArg<1>(&runningStatus))
.WillRepeatedly(Return()); // Ignore subsequent updates.
driver.launchTasks(offers.get()[0].id(), {task});
AWAIT_READY(startingStatus);
EXPECT_EQ(task.task_id(), startingStatus->task_id());
EXPECT_EQ(TASK_STARTING, startingStatus->state());
AWAIT_READY(runningStatus);
EXPECT_EQ(task.task_id(), runningStatus->task_id());
EXPECT_EQ(TASK_RUNNING, runningStatus->state());
Future<hashset<ContainerID>> containers = containerizer.get()->containers();
AWAIT_READY(containers);
ASSERT_EQ(1u, containers->size());
ContainerID containerId = *(containers->begin());
Duration diskTimeout = Seconds(5);
Timeout timeout = Timeout::in(diskTimeout);
while (true) {
Future<ResourceStatistics> usage = containerizer.get()->usage(containerId);
AWAIT_READY(usage);
ASSERT_TRUE(usage->has_disk_limit_bytes());
EXPECT_EQ(Megabytes(3), Bytes(usage->disk_limit_bytes()));
if (usage->has_disk_used_bytes()) {
if (usage->disk_used_bytes() >= Megabytes(4).bytes()) {
break;
}
}
// The stopping condition for this test is that the isolator is
// able to report that we wrote the full amount of data without
// being constrained by the task disk limit.
EXPECT_LE(usage->disk_used_bytes(), Megabytes(4).bytes());
ASSERT_FALSE(timeout.expired())
<< "Used " << Bytes(usage->disk_used_bytes())
<< " of expected " << Megabytes(4)
<< " within the " << diskTimeout << " timeout";
os::sleep(Milliseconds(100));
}
driver.stop();
driver.join();
}
// This test verifies that the volume usage accounting for sandboxes
// with bind-mounted volumes (while linux filesystem isolator is used)
// works correctly by creating a file within the volume the size of
// which exceeds the sandbox quota.
TEST_F(LinuxFilesystemIsolatorMesosTest,
ROOT_VolumeUsageExceedsSandboxQuota)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
string registry = path::join(sandbox.get(), "registry");
AWAIT_READY(DockerArchive::create(registry, "test_image"));
slave::Flags flags = CreateSlaveFlags();
flags.resources = "cpus:2;mem:128;disk(role1):128";
flags.isolation = "disk/du,filesystem/linux,docker/runtime";
flags.docker_registry = registry;
flags.docker_store_dir = path::join(sandbox.get(), "store");
flags.image_providers = "docker";
// NOTE: We can't pause the clock because we need the reaper to reap
// the 'du' subprocess.
flags.container_disk_watch_interval = Milliseconds(1);
flags.enforce_container_disk_quota = true;
Owned<MasterDetector> detector = master.get()->createDetector();
Try<Owned<cluster::Slave>> slave = StartSlave(detector.get(), flags);
ASSERT_SOME(slave);
MockScheduler sched;
FrameworkInfo frameworkInfo = DEFAULT_FRAMEWORK_INFO;
frameworkInfo.set_roles(0, "role1");
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);
ASSERT_FALSE(offers->empty());
// We request a sandbox (1MB) that is smaller than the persistent
// volume (4MB) and attempt to create a file in that volume that is
// twice the size of the sanbox (2MB).
Resources volume = createPersistentVolume(
Megabytes(4),
"role1",
"id1",
"volume_path",
None(),
None(),
frameworkInfo.principal());
Resources taskResources =
Resources::parse("cpus:1;mem:64;disk(role1):1").get() + volume;
// We sleep to give quota enforcement (du) a chance to kick in.
TaskInfo task = createTask(
offers.get()[0].slave_id(),
taskResources,
"dd if=/dev/zero of=volume_path/file bs=1048576 count=2 && sleep 1");
Future<TaskStatus> statusStarting;
Future<TaskStatus> statusRunning;
Future<TaskStatus> statusFinished;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&statusStarting))
.WillOnce(FutureArg<1>(&statusRunning))
.WillOnce(FutureArg<1>(&statusFinished));
driver.acceptOffers(
{offers.get()[0].id()},
{CREATE(volume),
LAUNCH({task})});
AWAIT_READY(statusStarting);
EXPECT_EQ(task.task_id(), statusStarting->task_id());
EXPECT_EQ(TASK_STARTING, statusStarting->state());
AWAIT_READY(statusRunning);
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();
}
// In this test, the framework is checkpointed so we expect the executor to
// persist across the slave restart and to have the same resource usage before
// and after.
TEST_F(ROOT_XFS_QuotaTest, CheckpointRecovery)
{
slave::Flags flags = CreateSlaveFlags();
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
Owned<MasterDetector> detector = master.get()->createDetector();
Try<Owned<cluster::Slave>> slave =
StartSlave(detector.get(), CreateSlaveFlags());
ASSERT_SOME(slave);
FrameworkInfo frameworkInfo = DEFAULT_FRAMEWORK_INFO;
frameworkInfo.set_checkpoint(true);
MockScheduler sched;
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];
TaskInfo task = createTask(
offer.slave_id(),
Resources::parse("cpus:1;mem:128;disk:1").get(),
"dd if=/dev/zero of=file bs=1048576 count=1; sleep 1000");
Future<TaskStatus> startingStatus;
Future<TaskStatus> runningStatus;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&startingStatus))
.WillOnce(FutureArg<1>(&runningStatus));
driver.launchTasks(offer.id(), {task});
AWAIT_READY(startingStatus);
EXPECT_EQ(task.task_id(), startingStatus->task_id());
EXPECT_EQ(TASK_STARTING, startingStatus->state());
AWAIT_READY(startingStatus);
EXPECT_EQ(task.task_id(), runningStatus->task_id());
EXPECT_EQ(TASK_RUNNING, runningStatus->state());
Future<ResourceUsage> usage1 =
process::dispatch(slave.get()->pid, &Slave::usage);
AWAIT_READY(usage1);
// We should have 1 executor using resources.
ASSERT_EQ(1, usage1->executors().size());
// Restart the slave.
slave.get()->terminate();
Future<SlaveReregisteredMessage> slaveReregisteredMessage =
FUTURE_PROTOBUF(SlaveReregisteredMessage(), _, _);
slave = StartSlave(detector.get(), flags);
ASSERT_SOME(slave);
// Wait for the slave to re-register.
AWAIT_READY(slaveReregisteredMessage);
Future<ResourceUsage> usage2 =
process::dispatch(slave.get()->pid, &Slave::usage);
AWAIT_READY(usage2);
// We should have still have 1 executor using resources.
ASSERT_EQ(1, usage1->executors().size());
Try<std::list<string>> sandboxes = os::glob(path::join(
slave::paths::getSandboxRootDir(mountPoint.get()),
"*",
"frameworks",
"*",
"executors",
"*",
"runs",
"*"));
ASSERT_SOME(sandboxes);
// One sandbox and one symlink.
ASSERT_EQ(2u, sandboxes->size());
// Scan the remaining sandboxes. We ought to still have project IDs
// assigned to them all.
foreach (const string& sandbox, sandboxes.get()) {
// Skip the "latest" symlink.
if (os::stat::islink(sandbox)) {
continue;
}
EXPECT_SOME(xfs::getProjectId(sandbox));
}
driver.stop();
driver.join();
}
// Tests that the task fails when it attempts to write to a persistent volume
// mounted as read-only. Note that although we use a shared persistent volume,
// the behavior is the same for non-shared persistent volumes.
TEST_F(LinuxFilesystemIsolatorMesosTest,
ROOT_WriteAccessSharedPersistentVolumeReadOnlyMode)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
string registry = path::join(sandbox.get(), "registry");
AWAIT_READY(DockerArchive::create(registry, "test_image"));
slave::Flags flags = CreateSlaveFlags();
flags.resources = "cpus:2;mem:128;disk(role1):128";
flags.isolation = "filesystem/linux,docker/runtime";
flags.docker_registry = registry;
flags.docker_store_dir = path::join(sandbox.get(), "store");
flags.image_providers = "docker";
Owned<MasterDetector> detector = master.get()->createDetector();
Try<Owned<cluster::Slave>> slave = StartSlave(detector.get(), flags);
ASSERT_SOME(slave);
MockScheduler sched;
FrameworkInfo frameworkInfo = DEFAULT_FRAMEWORK_INFO;
frameworkInfo.set_roles(0, "role1");
frameworkInfo.add_capabilities()->set_type(
FrameworkInfo::Capability::SHARED_RESOURCES);
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);
ASSERT_FALSE(offers->empty());
// We create a shared volume which shall be used by the task to
// write to that volume.
Resource volume = createPersistentVolume(
Megabytes(4),
"role1",
"id1",
"volume_path",
None(),
None(),
frameworkInfo.principal(),
true); // Shared volume.
// The task uses the shared volume as read-only.
Resource roVolume = volume;
roVolume.mutable_disk()->mutable_volume()->set_mode(Volume::RO);
Resources taskResources =
Resources::parse("cpus:1;mem:64;disk(role1):1").get() + roVolume;
TaskInfo task = createTask(
offers.get()[0].slave_id(),
taskResources,
"echo hello > volume_path/file");
// The task fails to write to the volume since the task's resources
// intends to use the volume as read-only.
Future<TaskStatus> statusStarting;
Future<TaskStatus> statusRunning;
Future<TaskStatus> statusFailed;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&statusStarting))
.WillOnce(FutureArg<1>(&statusRunning))
.WillOnce(FutureArg<1>(&statusFailed));
driver.acceptOffers(
{offers.get()[0].id()},
{CREATE(volume),
LAUNCH({task})});
AWAIT_READY(statusStarting);
EXPECT_EQ(task.task_id(), statusStarting->task_id());
EXPECT_EQ(TASK_STARTING, statusStarting->state());
AWAIT_READY(statusRunning);
EXPECT_EQ(task.task_id(), statusRunning->task_id());
EXPECT_EQ(TASK_RUNNING, statusRunning->state());
AWAIT_READY(statusFailed);
EXPECT_EQ(task.task_id(), statusFailed->task_id());
EXPECT_EQ(TASK_FAILED, statusFailed->state());
driver.stop();
driver.join();
}
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";
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;
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);
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;
Future<TaskStatus> killed;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&starting))
.WillOnce(FutureArg<1>(&running))
.WillOnce(FutureArg<1>(&killed))
.WillRepeatedly(Return()); // Ignore subsequent updates.
driver.launchTasks(offer.id(), {task});
AWAIT_READY(starting);
EXPECT_EQ(task.task_id(), starting->task_id());
EXPECT_EQ(TASK_STARTING, starting->state());
AWAIT_READY(running);
EXPECT_EQ(task.task_id(), running->task_id());
EXPECT_EQ(TASK_RUNNING, running->state());
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();
// 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();
}
// This test verifies that the framework can launch a command task
// that specifies a container image.
TEST_F(LinuxFilesystemIsolatorMesosTest,
ROOT_ChangeRootFilesystemCommandExecutor)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
string registry = path::join(sandbox.get(), "registry");
AWAIT_READY(DockerArchive::create(registry, "test_image"));
slave::Flags flags = CreateSlaveFlags();
flags.isolation = "filesystem/linux,docker/runtime";
flags.docker_registry = registry;
flags.docker_store_dir = path::join(sandbox.get(), "store");
flags.image_providers = "docker";
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_FALSE(offers->empty());
const Offer& offer = offers.get()[0];
TaskInfo task = createTask(
offer.slave_id(),
offer.resources(),
"test -d " + flags.sandbox_directory);
task.mutable_container()->CopyFrom(createContainerInfo("test_image"));
driver.launchTasks(offer.id(), {task});
Future<TaskStatus> statusStarting;
Future<TaskStatus> statusRunning;
Future<TaskStatus> statusFinished;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&statusStarting))
.WillOnce(FutureArg<1>(&statusRunning))
.WillOnce(FutureArg<1>(&statusFinished));
AWAIT_READY(statusStarting);
EXPECT_EQ(TASK_STARTING, statusStarting->state());
AWAIT_READY(statusRunning);
EXPECT_EQ(TASK_RUNNING, statusRunning->state());
AWAIT_READY(statusFinished);
EXPECT_EQ(TASK_FINISHED, statusFinished->state());
driver.stop();
driver.join();
}
// 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();
}
// This test launches a command task which has checkpoint enabled, and
// agent is terminated when the task is running, after agent is restarted,
// kill the task and then verify we can receive TASK_KILLED for the task.
TEST_F(CniIsolatorTest, ROOT_SlaveRecovery)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
slave::Flags flags = CreateSlaveFlags();
flags.isolation = "network/cni";
flags.network_cni_plugins_dir = cniPluginDir;
flags.network_cni_config_dir = cniConfigDir;
Owned<MasterDetector> detector = master.get()->createDetector();
Try<Owned<cluster::Slave>> slave = StartSlave(detector.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(&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];
CommandInfo command;
command.set_value("sleep 1000");
TaskInfo task = createTask(
offer.slave_id(),
Resources::parse("cpus:1;mem:128").get(),
command);
ContainerInfo* container = task.mutable_container();
container->set_type(ContainerInfo::MESOS);
// Make sure the container join the mock CNI network.
container->add_network_infos()->set_name("__MESOS_TEST__");
Future<TaskStatus> statusRunning;
Future<TaskStatus> statusKilled;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&statusRunning))
.WillOnce(FutureArg<1>(&statusKilled));
EXPECT_CALL(sched, offerRescinded(&driver, _))
.Times(AtMost(1));
Future<Nothing> ack =
FUTURE_DISPATCH(_, &Slave::_statusUpdateAcknowledgement);
driver.launchTasks(offer.id(), {task});
AWAIT_READY(statusRunning);
EXPECT_EQ(task.task_id(), statusRunning->task_id());
EXPECT_EQ(TASK_RUNNING, statusRunning->state());
// Wait for the ACK to be checkpointed.
AWAIT_READY(ack);
// Stop the slave after TASK_RUNNING is received.
slave.get()->terminate();
// Restart the slave.
slave = StartSlave(detector.get(), flags);
ASSERT_SOME(slave);
// Kill the task.
driver.killTask(task.task_id());
AWAIT_READY(statusKilled);
EXPECT_EQ(task.task_id(), statusKilled->task_id());
EXPECT_EQ(TASK_KILLED, statusKilled->state());
driver.stop();
driver.join();
}