// TODO(vinod): Using a paused clock in the tests means that
// future.await() may wait forever. This makes debugging hard.
TEST_F(FrameworksManagerTestFixture, ResurrectExpiringFramework)
{
// This is the crucial test.
// Resurrect an existing framework that is being removed,is being removed,
// which should cause the remove to be unsuccessful.
// Add a dummy framework.
FrameworkID id;
id.set_value("id");
FrameworkInfo info;
info.set_name("test name");
info.set_user("test user");
// Add the framework.
process::dispatch(manager, &FrameworksManager::add, id, info);
Clock::pause();
// Remove after 2 secs.
Future<Result<bool> > future1 =
process::dispatch(manager, &FrameworksManager::remove, id, seconds(2.0));
// Resurrect in the meanwhile.
Future<Result<bool> > future2 =
process::dispatch(manager, &FrameworksManager::resurrect, id);
ASSERT_TRUE(future2.await(2.0));
EXPECT_TRUE(future2.get().get());
Clock::update(Clock::now(manager) + 2.0);
ASSERT_TRUE(future1.await(2.0));
EXPECT_FALSE(future1.get().get());
Clock::resume();
}
// This test verifies that a framework registration with authorized
// role is successful.
TEST_F(MasterAuthorizationTest, AuthorizedRole)
{
// Setup ACLs so that the framework can receive offers for role
// "foo".
ACLs acls;
mesos::ACL::RegisterFramework* acl = acls.add_register_frameworks();
acl->mutable_principals()->add_values(DEFAULT_FRAMEWORK_INFO.principal());
acl->mutable_roles()->add_values("foo");
master::Flags flags = CreateMasterFlags();
flags.roles = "foo";
flags.acls = acls;
Try<PID<Master> > master = StartMaster(flags);
ASSERT_SOME(master);
FrameworkInfo frameworkInfo; // Bug in gcc 4.1.*, must assign on next line.
frameworkInfo = DEFAULT_FRAMEWORK_INFO;
frameworkInfo.set_role("foo");
MockScheduler sched;
MesosSchedulerDriver driver(
&sched, frameworkInfo, master.get(), DEFAULT_CREDENTIAL);
Future<Nothing> registered;
EXPECT_CALL(sched, registered(&driver, _, _))
.WillOnce(FutureSatisfy(®istered));
driver.start();
AWAIT_READY(registered);
driver.stop();
driver.join();
Shutdown();
}
int main(int argc, char** argv)
{
if (argc != 2) {
cerr << "Usage: " << argv[0] << " <master>" << endl;
return -1;
}
DockerNoExecutorScheduler scheduler;
FrameworkInfo framework;
framework.set_user(""); // Have Mesos fill in the current user.
framework.set_name("Docker No Executor Framework (C++)");
framework.set_checkpoint(true);
MesosSchedulerDriver* driver;
if (os::getenv("MESOS_AUTHENTICATE_FRAMEWORKS").isSome()) {
cout << "Enabling authentication for the framework" << endl;
Option<string> value = os::getenv("DEFAULT_PRINCIPAL");
if (value.isNone()) {
EXIT(EXIT_FAILURE)
<< "Expecting authentication principal in the environment";
}
Credential credential;
credential.set_principal(value.get());
framework.set_principal(value.get());
value = os::getenv("DEFAULT_SECRET");
if (value.isNone()) {
EXIT(EXIT_FAILURE)
<< "Expecting authentication secret in the environment";
}
credential.set_secret(value.get());
driver = new MesosSchedulerDriver(
&scheduler, framework, argv[1], credential);
} else {
framework.set_principal("no-executor-framework-cpp");
driver = new MesosSchedulerDriver(
&scheduler, framework, argv[1]);
}
int status = driver->run() == DRIVER_STOPPED ? 0 : 1;
// Ensure that the driver process terminates.
driver->stop();
delete driver;
return status;
}
jobject convert(JNIEnv* env, const FrameworkInfo& frameworkInfo)
{
string data;
frameworkInfo.SerializeToString(&data);
// byte[] data = ..;
jbyteArray jdata = env->NewByteArray(data.size());
env->SetByteArrayRegion(jdata, 0, data.size(), (jbyte*) data.data());
// FrameworkInfo frameworkInfo = FrameworkInfo.parseFrom(data);
jclass clazz = FindMesosClass(env, "org/apache/mesos/Protos$FrameworkInfo");
jmethodID parseFrom =
env->GetStaticMethodID(clazz, "parseFrom",
"([B)Lorg/apache/mesos/Protos$FrameworkInfo;");
jobject jframeworkInfo = env->CallStaticObjectMethod(clazz, parseFrom, jdata);
return jframeworkInfo;
}
// 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();
}
TEST_F(StatusUpdateManagerTest, CheckpointStatusUpdate)
{
Try<PID<Master> > master = StartMaster();
ASSERT_SOME(master);
MockExecutor exec(DEFAULT_EXECUTOR_ID);
slave::Flags flags = CreateSlaveFlags();
flags.checkpoint = true;
Try<PID<Slave> > slave = StartSlave(&exec, flags);
ASSERT_SOME(slave);
FrameworkInfo frameworkInfo; // Bug in gcc 4.1.*, must assign on next line.
frameworkInfo = DEFAULT_FRAMEWORK_INFO;
frameworkInfo.set_checkpoint(true); // Enable checkpointing.
MockScheduler sched;
MesosSchedulerDriver driver(
&sched, frameworkInfo, master.get(), DEFAULT_CREDENTIAL);
EXPECT_CALL(sched, registered(_, _, _))
.Times(1);
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());
EXPECT_CALL(exec, registered(_, _, _, _))
.Times(1);
EXPECT_CALL(exec, launchTask(_, _))
.WillOnce(SendStatusUpdateFromTask(TASK_RUNNING));
Future<TaskStatus> status;
EXPECT_CALL(sched, statusUpdate(_, _))
.WillOnce(FutureArg<1>(&status));
Future<Nothing> _statusUpdateAcknowledgement =
FUTURE_DISPATCH(slave.get(), &Slave::_statusUpdateAcknowledgement);
driver.launchTasks(offers.get()[0].id(), createTasks(offers.get()[0]));
AWAIT_READY(status);
EXPECT_EQ(TASK_RUNNING, status.get().state());
AWAIT_READY(_statusUpdateAcknowledgement);
// Ensure that both the status update and its acknowledgement are
// correctly checkpointed.
Try<list<string> > found = os::find(flags.work_dir, TASK_UPDATES_FILE);
ASSERT_SOME(found);
ASSERT_EQ(1u, found.get().size());
Try<int> fd = os::open(found.get().front(), O_RDONLY);
ASSERT_SOME(fd);
int updates = 0;
int acks = 0;
string uuid;
Result<StatusUpdateRecord> record = None();
while (true) {
record = ::protobuf::read<StatusUpdateRecord>(fd.get());
ASSERT_FALSE(record.isError());
if (record.isNone()) { // Reached EOF.
break;
}
if (record.get().type() == StatusUpdateRecord::UPDATE) {
EXPECT_EQ(TASK_RUNNING, record.get().update().status().state());
uuid = record.get().update().uuid();
updates++;
} else {
EXPECT_EQ(uuid, record.get().uuid());
acks++;
}
}
ASSERT_EQ(1, updates);
ASSERT_EQ(1, acks);
close(fd.get());
EXPECT_CALL(exec, shutdown(_))
.Times(AtMost(1));
driver.stop();
driver.join();
Shutdown();
}
// Ensures that the driver can handle the SUBSCRIBED event
// after a master failover.
TEST_F(SchedulerDriverEventTest, SubscribedMasterFailover)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
FrameworkInfo frameworkInfo = DEFAULT_FRAMEWORK_INFO;
frameworkInfo.set_failover_timeout(Weeks(2).secs());
// Make sure the initial registration calls 'registered'.
MockScheduler sched;
StandaloneMasterDetector detector(master.get()->pid);
TestingMesosSchedulerDriver driver(&sched, &detector, frameworkInfo);
// Intercept the registration message, send a SUBSCRIBED instead.
Future<Message> frameworkRegisteredMessage =
DROP_MESSAGE(Eq(FrameworkRegisteredMessage().GetTypeName()), _, _);
// Ensure that there will be no (re-)registration retries
// from the scheduler driver.
Clock::pause();
driver.start();
AWAIT_READY(frameworkRegisteredMessage);
UPID frameworkPid = frameworkRegisteredMessage.get().to;
FrameworkRegisteredMessage message;
ASSERT_TRUE(message.ParseFromString(frameworkRegisteredMessage.get().body));
FrameworkID frameworkId = message.framework_id();
frameworkInfo.mutable_id()->CopyFrom(frameworkId);
Event event;
event.set_type(Event::SUBSCRIBED);
event.mutable_subscribed()->mutable_framework_id()->CopyFrom(frameworkId);
Future<Nothing> registered;
EXPECT_CALL(sched, registered(&driver, frameworkId, _))
.WillOnce(FutureSatisfy(®istered));
process::post(master.get()->pid, frameworkPid, event);
AWAIT_READY(registered);
EXPECT_CALL(sched, disconnected(&driver));
// Fail over the master and expect a 'reregistered' call.
// Note that the master sends a registered message for
// this case (see MESOS-786).
master->reset();
master = StartMaster();
ASSERT_SOME(master);
frameworkRegisteredMessage =
DROP_MESSAGE(Eq(FrameworkRegisteredMessage().GetTypeName()), _, _);
detector.appoint(master.get()->pid);
AWAIT_READY(frameworkRegisteredMessage);
Future<Nothing> reregistered;
EXPECT_CALL(sched, reregistered(&driver, _))
.WillOnce(FutureSatisfy(&reregistered));
process::post(master.get()->pid, frameworkPid, event);
AWAIT_READY(reregistered);
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();
}
// This test verifies that the slave and status update manager
// properly handle duplicate status updates, when the second
// update with the same UUID is received before the ACK for the
// first update. The proper behavior here is for the status update
// manager to drop the duplicate update.
TEST_F(StatusUpdateManagerTest, DuplicateUpdateBeforeAck)
{
Try<PID<Master> > master = StartMaster();
ASSERT_SOME(master);
MockExecutor exec(DEFAULT_EXECUTOR_ID);
slave::Flags flags = CreateSlaveFlags();
flags.checkpoint = true;
Try<PID<Slave> > slave = StartSlave(&exec, flags);
ASSERT_SOME(slave);
FrameworkInfo frameworkInfo; // Bug in gcc 4.1.*, must assign on next line.
frameworkInfo = DEFAULT_FRAMEWORK_INFO;
frameworkInfo.set_checkpoint(true); // Enable checkpointing.
MockScheduler sched;
MesosSchedulerDriver driver(
&sched, frameworkInfo, master.get(), DEFAULT_CREDENTIAL);
FrameworkID frameworkId;
EXPECT_CALL(sched, registered(_, _, _))
.WillOnce(SaveArg<1>(&frameworkId));
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());
ExecutorDriver* execDriver;
EXPECT_CALL(exec, registered(_, _, _, _))
.WillOnce(SaveArg<0>(&execDriver));
EXPECT_CALL(exec, launchTask(_, _))
.WillOnce(SendStatusUpdateFromTask(TASK_RUNNING));
// Capture the first status update message.
Future<StatusUpdateMessage> statusUpdateMessage =
FUTURE_PROTOBUF(StatusUpdateMessage(), _, _);
Future<TaskStatus> status;
EXPECT_CALL(sched, statusUpdate(_, _))
.WillOnce(FutureArg<1>(&status));
// Drop the first ACK from the scheduler to the slave.
Future<StatusUpdateAcknowledgementMessage> statusUpdateAcknowledgementMessage =
DROP_PROTOBUF(StatusUpdateAcknowledgementMessage(), _, slave.get());
Clock::pause();
driver.launchTasks(offers.get()[0].id(), createTasks(offers.get()[0]));
AWAIT_READY(statusUpdateMessage);
AWAIT_READY(status);
EXPECT_EQ(TASK_RUNNING, status.get().state());
AWAIT_READY(statusUpdateAcknowledgementMessage);
Future<Nothing> _statusUpdate =
FUTURE_DISPATCH(slave.get(), &Slave::_statusUpdate);
// Now resend the TASK_RUNNING update.
process::post(slave.get(), statusUpdateMessage.get());
// At this point the status update manager has handled
// the duplicate status update.
AWAIT_READY(_statusUpdate);
// After we advance the clock, the status update manager should
// retry the TASK_RUNING update and the scheduler should receive
// and acknowledge it.
Future<TaskStatus> update;
EXPECT_CALL(sched, statusUpdate(_, _))
.WillOnce(FutureArg<1>(&update));
Clock::advance(slave::STATUS_UPDATE_RETRY_INTERVAL);
Clock::settle();
// Ensure the scheduler receives TASK_FINISHED.
AWAIT_READY(update);
EXPECT_EQ(TASK_RUNNING, update.get().state());
EXPECT_CALL(exec, shutdown(_))
.Times(AtMost(1));
Clock::resume();
driver.stop();
driver.join();
Shutdown();
}
// 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_FOR(_statusUpdateAcknowledgement, 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.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_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();
}
int main(int argc, char** argv)
{
if (argc != 3) {
std::cerr << "Usage: " << argv[0]
<< " <master> <balloon limit in MB>" << std::endl;
return -1;
}
// Verify the balloon limit.
Try<size_t> limit = numify<size_t>(argv[2]);
if (limit.isError()) {
std::cerr << "Balloon limit is not a valid number" << std::endl;
return -1;
}
if (limit.get() < EXECUTOR_MEMORY_MB) {
std::cerr << "Please use a balloon limit bigger than "
<< EXECUTOR_MEMORY_MB << " MB" << std::endl;
}
// Find this executable's directory to locate executor.
std::string path = os::realpath(::dirname(argv[0])).get();
std::string uri = path + "/balloon-executor";
if (getenv("MESOS_BUILD_DIR")) {
uri = std::string(::getenv("MESOS_BUILD_DIR")) + "/src/balloon-executor";
}
ExecutorInfo executor;
executor.mutable_executor_id()->set_value("default");
executor.mutable_command()->set_value(uri);
executor.set_name("Balloon Executor");
executor.set_source("balloon_test");
Resource* mem = executor.add_resources();
mem->set_name("mem");
mem->set_type(Value::SCALAR);
mem->mutable_scalar()->set_value(EXECUTOR_MEMORY_MB);
BalloonScheduler scheduler(executor, limit.get());
FrameworkInfo framework;
framework.set_user(""); // Have Mesos fill in the current user.
framework.set_name("Balloon Framework (C++)");
// TODO(vinod): Make checkpointing the default when it is default
// on the slave.
if (os::hasenv("MESOS_CHECKPOINT")) {
cout << "Enabling checkpoint for the framework" << endl;
framework.set_checkpoint(true);
}
MesosSchedulerDriver* driver;
if (os::hasenv("MESOS_AUTHENTICATE")) {
cout << "Enabling authentication for the framework" << endl;
if (!os::hasenv("DEFAULT_PRINCIPAL")) {
EXIT(1) << "Expecting authentication principal in the environment";
}
if (!os::hasenv("DEFAULT_SECRET")) {
EXIT(1) << "Expecting authentication secret in the environment";
}
Credential credential;
credential.set_principal(getenv("DEFAULT_PRINCIPAL"));
credential.set_secret(getenv("DEFAULT_SECRET"));
framework.set_principal(getenv("DEFAULT_PRINCIPAL"));
driver = new MesosSchedulerDriver(
&scheduler, framework, argv[1], credential);
} else {
framework.set_principal("balloon-framework-cpp");
driver = new MesosSchedulerDriver(
&scheduler, framework, argv[1]);
}
int status = driver->run() == DRIVER_STOPPED ? 0 : 1;
// Ensure that the driver process terminates.
driver->stop();
delete driver;
return status;
}
// This test verifies that disk quota isolator recovers properly after
// the slave restarts.
TEST_F(DiskQuotaTest, SlaveRecovery)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
slave::Flags flags = CreateSlaveFlags();
flags.isolation = "posix/cpu,posix/mem,disk/du";
flags.container_disk_watch_interval = Milliseconds(1);
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());
const Offer& offer = offers.get()[0];
// Create a task that uses 2MB disk.
TaskInfo task = createTask(
offer.slave_id(),
Resources::parse("cpus:1;mem:128;disk:3").get(),
"dd if=/dev/zero of=file bs=1048576 count=2 && sleep 1000");
Future<TaskStatus> status;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&status))
.WillRepeatedly(Return()); // Ignore subsequent updates.
driver.launchTasks(offer.id(), {task});
AWAIT_READY(status);
EXPECT_EQ(task.task_id(), status->task_id());
EXPECT_EQ(TASK_RUNNING, status->state());
Future<hashset<ContainerID>> containers = containerizer->containers();
AWAIT_READY(containers);
ASSERT_EQ(1u, containers->size());
const ContainerID& containerId = *(containers->begin());
// Stop the slave.
slave.get()->terminate();
Future<ReregisterExecutorMessage> reregisterExecutorMessage =
FUTURE_PROTOBUF(ReregisterExecutorMessage(), _, _);
Future<Nothing> _recover = FUTURE_DISPATCH(_, &Slave::_recover);
_containerizer = MesosContainerizer::create(flags, true, &fetcher);
ASSERT_SOME(_containerizer);
containerizer.reset(_containerizer.get());
detector = master.get()->createDetector();
slave = StartSlave(detector.get(), containerizer.get(), flags);
ASSERT_SOME(slave);
Clock::pause();
AWAIT_READY(_recover);
// Wait for slave to schedule reregister timeout.
Clock::settle();
// Ensure the executor re-registers before completing recovery.
AWAIT_READY(reregisterExecutorMessage);
// Ensure the slave considers itself recovered.
Clock::advance(flags.executor_reregistration_timeout);
// NOTE: We resume the clock because we need the reaper to reap the
// 'du' subprocess.
Clock::resume();
// Wait until disk usage can be retrieved.
Duration elapsed = Duration::zero();
while (true) {
Future<ResourceStatistics> usage = containerizer->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()) {
EXPECT_LE(usage->disk_used_bytes(), usage->disk_limit_bytes());
// NOTE: This is to capture the regression in MESOS-2452. The data
// stored in the executor meta directory should be less than 64K.
if (usage->disk_used_bytes() > Kilobytes(64).bytes()) {
break;
}
}
ASSERT_LT(elapsed, Seconds(15));
os::sleep(Milliseconds(1));
elapsed += Milliseconds(1);
}
driver.stop();
driver.join();
}
TEST_F(StatusUpdateManagerTest, RetryStatusUpdate)
{
Try<PID<Master> > master = StartMaster();
ASSERT_SOME(master);
MockExecutor exec(DEFAULT_EXECUTOR_ID);
slave::Flags flags = CreateSlaveFlags();
flags.checkpoint = true;
Try<PID<Slave> > slave = StartSlave(&exec, flags);
ASSERT_SOME(slave);
FrameworkInfo frameworkInfo; // Bug in gcc 4.1.*, must assign on next line.
frameworkInfo = DEFAULT_FRAMEWORK_INFO;
frameworkInfo.set_checkpoint(true); // Enable checkpointing.
MockScheduler sched;
MesosSchedulerDriver driver(
&sched, frameworkInfo, master.get(), DEFAULT_CREDENTIAL);
EXPECT_CALL(sched, registered(_, _, _))
.Times(1);
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());
EXPECT_CALL(exec, registered(_, _, _, _))
.Times(1);
EXPECT_CALL(exec, launchTask(_, _))
.WillOnce(SendStatusUpdateFromTask(TASK_RUNNING));
Future<StatusUpdateMessage> statusUpdateMessage =
DROP_PROTOBUF(StatusUpdateMessage(), master.get(), _);
Clock::pause();
driver.launchTasks(offers.get()[0].id(), createTasks(offers.get()[0]));
AWAIT_READY(statusUpdateMessage);
Future<TaskStatus> status;
EXPECT_CALL(sched, statusUpdate(_, _))
.WillOnce(FutureArg<1>(&status));
Clock::advance(slave::STATUS_UPDATE_RETRY_INTERVAL);
AWAIT_READY(status);
EXPECT_EQ(TASK_RUNNING, status.get().state());
Clock::resume();
EXPECT_CALL(exec, shutdown(_))
.Times(AtMost(1));
driver.stop();
driver.join();
Shutdown();
}
// This test verifies that the container will be killed if the volume
// usage exceeds its quota.
TEST_F(DiskQuotaTest, VolumeUsageExceedsQuota)
{
FrameworkInfo frameworkInfo = DEFAULT_FRAMEWORK_INFO;
frameworkInfo.set_role("role1");
master::Flags masterFlags = CreateMasterFlags();
Try<Owned<cluster::Master>> master = StartMaster(masterFlags);
ASSERT_SOME(master);
slave::Flags slaveFlags = CreateSlaveFlags();
slaveFlags.isolation = "posix/cpu,posix/mem,disk/du";
// NOTE: We can't pause the clock because we need the reaper to reap
// the 'du' subprocess.
slaveFlags.container_disk_watch_interval = Milliseconds(1);
slaveFlags.enforce_container_disk_quota = true;
slaveFlags.resources = "cpus:2;mem:128;disk(role1):128";
Try<Resources> initialResources =
Resources::parse(slaveFlags.resources.get());
ASSERT_SOME(initialResources);
Owned<MasterDetector> detector = master.get()->createDetector();
Try<Owned<cluster::Slave>> slave = StartSlave(detector.get(), slaveFlags);
ASSERT_SOME(slave);
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());
const Offer& offer = offers.get()[0];
// Create a task that requests a 1 MB persistent volume but attempts
// to use 2MB.
Resources volume = createPersistentVolume(
Megabytes(1),
"role1",
"id1",
"volume_path",
None(),
None(),
frameworkInfo.principal());
// We intentionally request a sandbox that is much bugger (16MB) than
// the file the task writes (2MB) to the persistent volume (1MB). This
// makes sure that the quota is indeed enforced on the persistent volume.
Resources taskResources =
Resources::parse("cpus:1;mem:64;disk(role1):16").get() + volume;
TaskInfo task = createTask(
offer.slave_id(),
taskResources,
"dd if=/dev/zero of=volume_path/file bs=1048576 count=2 && sleep 1000");
Future<TaskStatus> status1;
Future<TaskStatus> status2;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&status1))
.WillOnce(FutureArg<1>(&status2));
// Create the volume and launch the task.
driver.acceptOffers(
{offer.id()},
{CREATE(volume),
LAUNCH({task})});
AWAIT_READY(status1);
EXPECT_EQ(task.task_id(), status1->task_id());
EXPECT_EQ(TASK_RUNNING, status1->state());
AWAIT_READY(status2);
EXPECT_EQ(task.task_id(), status1->task_id());
EXPECT_EQ(TASK_FAILED, status2->state());
driver.stop();
driver.join();
}
TEST_F(DiskQuotaTest, ResourceStatistics)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
slave::Flags flags = CreateSlaveFlags();
flags.isolation = "posix/cpu,posix/mem,disk/du";
flags.resources = strings::format("disk(%s):10", DEFAULT_TEST_ROLE).get();
// NOTE: We can't pause the clock because we need the reaper to reap
// the 'du' subprocess.
flags.container_disk_watch_interval = Milliseconds(1);
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);
FrameworkInfo frameworkInfo = DEFAULT_FRAMEWORK_INFO;
frameworkInfo.set_role(DEFAULT_TEST_ROLE);
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());
const Offer& offer = offers.get()[0];
Resource volume = createPersistentVolume(
Megabytes(4),
DEFAULT_TEST_ROLE,
"id1",
"path1",
None(),
None(),
DEFAULT_CREDENTIAL.principal());
Resources taskResources = Resources::parse("cpus:1;mem:128").get();
taskResources += createDiskResource(
"3",
DEFAULT_TEST_ROLE,
None(),
None());
taskResources += volume;
// Create a task that uses 2MB disk.
TaskInfo task = createTask(
offer.slave_id(),
taskResources,
"dd if=/dev/zero of=file bs=1048576 count=2 && "
"dd if=/dev/zero of=path1/file bs=1048576 count=2 && "
"sleep 1000");
Future<TaskStatus> status1;
Future<TaskStatus> status2;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&status1))
.WillOnce(FutureArg<1>(&status2))
.WillRepeatedly(Return()); // Ignore subsequent updates.
driver.acceptOffers(
{offer.id()},
{CREATE(volume),
LAUNCH({task})});
AWAIT_READY(status1);
EXPECT_EQ(task.task_id(), status1->task_id());
EXPECT_EQ(TASK_RUNNING, status1->state());
Future<hashset<ContainerID>> containers = containerizer->containers();
AWAIT_READY(containers);
ASSERT_EQ(1u, containers->size());
const ContainerID& containerId = *(containers->begin());
// Wait until disk usage can be retrieved.
Duration elapsed = Duration::zero();
while (true) {
Future<ResourceStatistics> usage = containerizer->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()) {
EXPECT_LE(usage->disk_used_bytes(), usage->disk_limit_bytes());
}
ASSERT_EQ(2u, usage->disk_statistics().size());
bool done = true;
foreach (const DiskStatistics& statistics, usage->disk_statistics()) {
ASSERT_TRUE(statistics.has_limit_bytes());
EXPECT_EQ(
statistics.has_persistence() ? Megabytes(4) : Megabytes(3),
statistics.limit_bytes());
if (!statistics.has_used_bytes()) {
done = false;
} else {
EXPECT_GT(
statistics.has_persistence() ? Megabytes(4) : Megabytes(3),
statistics.used_bytes());
}
}
if (done) {
break;
}
ASSERT_LT(elapsed, Seconds(5));
os::sleep(Milliseconds(1));
elapsed += Milliseconds(1);
}
driver.killTask(task.task_id());
AWAIT_READY(status2);
EXPECT_EQ(task.task_id(), status2->task_id());
EXPECT_EQ(TASK_KILLED, status2->state());
driver.stop();
driver.join();
}
int main(int argc, char** argv)
{
string master = MASTER;
string k3binary;
string paramfile;
int total_peers;
YAML::Node k3vars;
// Parse Command Line options
string path = os::realpath(dirname(argv[0])).get();
namespace po = boost::program_options;
vector <string> optionalVars;
po::options_description desc("K3 Run options");
desc.add_options()
("program", po::value<string>(&k3binary)->required(), "K3 executable program filename")
("numpeers", po::value<int>(&total_peers)->required(), "# of K3 Peers to launch")
("help,h", "Print help message")
("param,p", po::value<string>(¶mfile)->required(), "YAML Formatted input file");
po::positional_options_description positionalOptions;
positionalOptions.add("program", 1);
positionalOptions.add("numpeers", 1);
po::variables_map vm;
try {
po::store(po::command_line_parser(argc, argv).options(desc)
.positional(positionalOptions).run(), vm);
if (vm.count("help") || vm.empty()) {
cout << "K3 Distributed program framework backed by Mesos cluser" << endl;
cout << desc << endl;
return 0;
}
po::notify(vm);
k3vars = YAML::LoadFile(paramfile);
}
catch (boost::program_options::required_option& e) {
cerr << " ERROR: " << e.what() << endl << endl;
cout << desc << endl;
return 1;
}
catch (boost::program_options::error& e) {
cerr << " ERROR: " << e.what() << endl << endl;
cout << desc << endl;
return 1;
}
KDScheduler scheduler(k3binary, total_peers, k3vars, path);
FrameworkInfo framework;
framework.set_user(""); // Have Mesos fill in the current user.
framework.set_name(k3binary + "-" + stringify(total_peers));
framework.mutable_id()->set_value(k3binary);
if (os::hasenv("MESOS_CHECKPOINT")) {
cout << "Enabling checkpoint for the framework" << endl;
framework.set_checkpoint(true);
}
MesosSchedulerDriver* driver;
if (os::hasenv("MESOS_AUTHENTICATE")) {
cout << "Enabling authentication for the framework" << endl;
if (!os::hasenv("DEFAULT_PRINCIPAL")) {
EXIT(1) << "Expecting authentication principal in the environment";
}
if (!os::hasenv("DEFAULT_SECRET")) {
EXIT(1) << "Expecting authentication secret in the environment";
}
Credential credential;
credential.set_principal(getenv("DEFAULT_PRINCIPAL"));
credential.set_secret(getenv("DEFAULT_SECRET"));
framework.set_principal(getenv("DEFAULT_PRINCIPAL"));
driver = new MesosSchedulerDriver(
&scheduler, framework, master, credential);
} else {
framework.set_principal("k3-docker-no-executor-framework-cpp");
driver = new MesosSchedulerDriver(&scheduler, framework, master);
}
int status = driver->run() == DRIVER_STOPPED ? 0 : 1;
// Ensure that the driver process terminates.
driver->stop();
delete driver;
return status;
}
// This test verifies that, after a master failover, reconciliation of an
// operation that is still pending on an agent results in `OPERATION_PENDING`.
TEST_P(OperationReconciliationTest, AgentPendingOperationAfterMasterFailover)
{
Clock::pause();
mesos::internal::master::Flags masterFlags = CreateMasterFlags();
Try<Owned<cluster::Master>> master = StartMaster(masterFlags);
ASSERT_SOME(master);
Future<UpdateSlaveMessage> updateSlaveMessage =
FUTURE_PROTOBUF(UpdateSlaveMessage(), _, _);
auto detector = std::make_shared<StandaloneMasterDetector>(master.get()->pid);
mesos::internal::slave::Flags slaveFlags = CreateSlaveFlags();
Try<Owned<cluster::Slave>> slave = StartSlave(detector.get(), slaveFlags);
ASSERT_SOME(slave);
// Advance the clock to trigger agent registration.
Clock::advance(slaveFlags.registration_backoff_factor);
// Wait for the agent to register.
AWAIT_READY(updateSlaveMessage);
// Start and register a resource provider.
ResourceProviderInfo resourceProviderInfo;
resourceProviderInfo.set_type("org.apache.mesos.rp.test");
resourceProviderInfo.set_name("test");
Resource disk = createDiskResource(
"200", "*", None(), None(), createDiskSourceRaw(None(), "profile"));
Owned<MockResourceProvider> resourceProvider(
new MockResourceProvider(
resourceProviderInfo,
Resources(disk)));
// We override the mock resource provider's default action, so the operation
// will stay in `OPERATION_PENDING`.
Future<resource_provider::Event::ApplyOperation> applyOperation;
EXPECT_CALL(*resourceProvider, applyOperation(_))
.WillOnce(FutureArg<0>(&applyOperation));
Owned<EndpointDetector> endpointDetector(
mesos::internal::tests::resource_provider::createEndpointDetector(
slave.get()->pid));
updateSlaveMessage = FUTURE_PROTOBUF(UpdateSlaveMessage(), _, _);
// NOTE: We need to resume the clock so that the resource provider can
// fully register.
Clock::resume();
ContentType contentType = GetParam();
resourceProvider->start(endpointDetector, contentType);
// Wait until the agent's resources have been updated to include the
// resource provider resources.
AWAIT_READY(updateSlaveMessage);
ASSERT_TRUE(updateSlaveMessage->has_resource_providers());
ASSERT_EQ(1, updateSlaveMessage->resource_providers().providers_size());
Clock::pause();
// Start a v1 framework.
auto scheduler = std::make_shared<MockHTTPScheduler>();
FrameworkInfo frameworkInfo = DEFAULT_FRAMEWORK_INFO;
frameworkInfo.set_roles(0, DEFAULT_TEST_ROLE);
EXPECT_CALL(*scheduler, connected(_))
.WillOnce(scheduler::SendSubscribe(frameworkInfo));
Future<scheduler::Event::Subscribed> subscribed;
EXPECT_CALL(*scheduler, subscribed(_, _))
.WillOnce(FutureArg<1>(&subscribed));
// Ignore heartbeats.
EXPECT_CALL(*scheduler, heartbeat(_))
.WillRepeatedly(Return());
// Decline offers that do not contain wanted resources.
EXPECT_CALL(*scheduler, offers(_, _))
.WillRepeatedly(scheduler::DeclineOffers());
Future<scheduler::Event::Offers> offers;
auto isRaw = [](const Resource& r) {
return r.has_disk() &&
r.disk().has_source() &&
r.disk().source().type() == Resource::DiskInfo::Source::RAW;
};
EXPECT_CALL(*scheduler, offers(_, scheduler::OffersHaveAnyResource(
std::bind(isRaw, lambda::_1))))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(scheduler::DeclineOffers()); // Decline successive offers.
scheduler::TestMesos mesos(
master.get()->pid, contentType, scheduler, detector);
AWAIT_READY(subscribed);
FrameworkID frameworkId(subscribed->framework_id());
// NOTE: If the framework has not declined an unwanted offer yet when
// the master updates the agent with the RAW disk resource, the new
// allocation triggered by this update won't generate an allocatable
// offer due to no CPU and memory resources. So here we first settle
// the clock to ensure that the unwanted offer has been declined, then
// advance the clock to trigger another allocation.
Clock::settle();
Clock::advance(masterFlags.allocation_interval);
AWAIT_READY(offers);
ASSERT_FALSE(offers->offers().empty());
const Offer& offer = offers->offers(0);
const AgentID& agentId = offer.agent_id();
Option<Resource> source;
Option<ResourceProviderID> resourceProviderId;
foreach (const Resource& resource, offer.resources()) {
if (isRaw(resource)) {
source = resource;
ASSERT_TRUE(resource.has_provider_id());
resourceProviderId = resource.provider_id();
break;
}
}
ASSERT_SOME(source);
ASSERT_SOME(resourceProviderId);
OperationID operationId;
operationId.set_value("operation");
mesos.send(createCallAccept(
frameworkId,
offer,
{CREATE_DISK(
source.get(),
Resource::DiskInfo::Source::MOUNT,
None(),
operationId)}));
AWAIT_READY(applyOperation);
// Simulate master failover.
EXPECT_CALL(*scheduler, disconnected(_));
detector->appoint(None());
master->reset();
master = StartMaster();
ASSERT_SOME(master);
// Settle the clock to ensure the master finishes recovering the registry.
Clock::settle();
Future<SlaveReregisteredMessage> slaveReregistered = FUTURE_PROTOBUF(
SlaveReregisteredMessage(), master.get()->pid, slave.get()->pid);
updateSlaveMessage = FUTURE_PROTOBUF(UpdateSlaveMessage(), _, _);
EXPECT_CALL(*scheduler, connected(_))
.WillOnce(scheduler::SendSubscribe(frameworkInfo, frameworkId));
Future<scheduler::Event::Subscribed> frameworkResubscribed;
EXPECT_CALL(*scheduler, subscribed(_, _))
.WillOnce(FutureArg<1>(&frameworkResubscribed));
// Simulate a new master detected event to the agent and the scheduler.
detector->appoint(master.get()->pid);
// Advance the clock, so that the agent re-registers.
Clock::advance(slaveFlags.registration_backoff_factor);
// Resume the clock to avoid deadlocks related to agent registration.
// See MESOS-8828.
Clock::resume();
// Wait for the framework and agent to re-register.
AWAIT_READY(slaveReregistered);
AWAIT_READY(updateSlaveMessage);
AWAIT_READY(frameworkResubscribed);
Clock::pause();
// Test explicit reconciliation
{
scheduler::Call::ReconcileOperations::Operation operation;
operation.mutable_operation_id()->CopyFrom(operationId);
operation.mutable_agent_id()->CopyFrom(agentId);
const Future<scheduler::APIResult> result =
mesos.call({createCallReconcileOperations(frameworkId, {operation})});
AWAIT_READY(result);
// The master should respond with '200 OK' and with a `scheduler::Response`.
ASSERT_EQ(process::http::Status::OK, result->status_code());
ASSERT_TRUE(result->has_response());
const scheduler::Response response = result->response();
ASSERT_EQ(scheduler::Response::RECONCILE_OPERATIONS, response.type());
ASSERT_TRUE(response.has_reconcile_operations());
const scheduler::Response::ReconcileOperations& reconcile =
response.reconcile_operations();
ASSERT_EQ(1, reconcile.operation_statuses_size());
const OperationStatus& operationStatus = reconcile.operation_statuses(0);
EXPECT_EQ(operationId, operationStatus.operation_id());
EXPECT_EQ(OPERATION_PENDING, operationStatus.state());
EXPECT_FALSE(operationStatus.has_uuid());
}
// Test implicit reconciliation
{
const Future<scheduler::APIResult> result =
mesos.call({createCallReconcileOperations(frameworkId, {})});
AWAIT_READY(result);
// The master should respond with '200 OK' and with a `scheduler::Response`.
ASSERT_EQ(process::http::Status::OK, result->status_code());
ASSERT_TRUE(result->has_response());
const scheduler::Response response = result->response();
ASSERT_EQ(scheduler::Response::RECONCILE_OPERATIONS, response.type());
ASSERT_TRUE(response.has_reconcile_operations());
const scheduler::Response::ReconcileOperations& reconcile =
response.reconcile_operations();
ASSERT_EQ(1, reconcile.operation_statuses_size());
const OperationStatus& operationStatus = reconcile.operation_statuses(0);
EXPECT_EQ(operationId, operationStatus.operation_id());
EXPECT_EQ(OPERATION_PENDING, operationStatus.state());
EXPECT_FALSE(operationStatus.has_uuid());
}
}
// 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();
}
int main(int argc, char** argv)
{
if (argc != 2) {
cerr << "Usage: " << argv[0] << " <master>" << endl;
return -1;
}
// Find this executable's directory to locate executor.
string path = os::realpath(dirname(argv[0])).get();
string uri = path + "/long-lived-executor";
if (getenv("MESOS_BUILD_DIR")) {
uri = string(getenv("MESOS_BUILD_DIR")) + "/src/long-lived-executor";
}
ExecutorInfo executor;
executor.mutable_executor_id()->set_value("default");
executor.mutable_command()->set_value(uri);
executor.set_name("Long Lived Executor (C++)");
executor.set_source("cpp_long_lived_framework");
LongLivedScheduler scheduler(executor);
FrameworkInfo framework;
framework.set_user(""); // Have Mesos fill in the current user.
framework.set_name("Long Lived Framework (C++)");
// TODO(vinod): Make checkpointing the default when it is default
// on the slave.
if (os::hasenv("MESOS_CHECKPOINT")) {
cout << "Enabling checkpoint for the framework" << endl;
framework.set_checkpoint(true);
}
MesosSchedulerDriver* driver;
if (os::hasenv("MESOS_AUTHENTICATE")) {
cout << "Enabling authentication for the framework" << endl;
if (!os::hasenv("DEFAULT_PRINCIPAL")) {
EXIT(1) << "Expecting authentication principal in the environment";
}
if (!os::hasenv("DEFAULT_SECRET")) {
EXIT(1) << "Expecting authentication secret in the environment";
}
Credential credential;
credential.set_principal(getenv("DEFAULT_PRINCIPAL"));
credential.set_secret(getenv("DEFAULT_SECRET"));
framework.set_principal(getenv("DEFAULT_PRINCIPAL"));
driver = new MesosSchedulerDriver(
&scheduler, framework, argv[1], credential);
} else {
framework.set_principal("long-lived-framework-cpp");
driver = new MesosSchedulerDriver(
&scheduler, framework, argv[1]);
}
int status = driver->run() == DRIVER_STOPPED ? 0 : 1;
// Ensure that the driver process terminates.
driver->stop();
delete driver;
return status;
}
inline bool operator==(const FrameworkInfo& left, const FrameworkInfo& right)
{
return (left.name() == right.name()) && (left.user() == right.user());
}
// 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();
}
// 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();
}
// Ensures that the driver can handle the SUBSCRIBED event
// after a scheduler failover.
TEST_F(SchedulerDriverEventTest, SubscribedSchedulerFailover)
{
Try<PID<Master>> master = StartMaster();
ASSERT_SOME(master);
FrameworkInfo frameworkInfo = DEFAULT_FRAMEWORK_INFO;
frameworkInfo.set_failover_timeout(Weeks(2).secs());
// Make sure the initial registration calls 'registered'.
MockScheduler sched;
MesosSchedulerDriver driver(
&sched, frameworkInfo, master.get(), DEFAULT_CREDENTIAL);
// Intercept the registration message, send a SUBSCRIBED instead.
Future<Message> frameworkRegisteredMessage =
DROP_MESSAGE(Eq(FrameworkRegisteredMessage().GetTypeName()), _, _);
// Ensure that there will be no (re-)registration retries
// from the scheduler driver.
Clock::pause();
driver.start();
AWAIT_READY(frameworkRegisteredMessage);
UPID frameworkPid = frameworkRegisteredMessage.get().to;
FrameworkRegisteredMessage message;
ASSERT_TRUE(message.ParseFromString(frameworkRegisteredMessage.get().body));
FrameworkID frameworkId = message.framework_id();
frameworkInfo.mutable_id()->CopyFrom(frameworkId);
Event event;
event.set_type(Event::SUBSCRIBED);
event.mutable_subscribed()->mutable_framework_id()->CopyFrom(frameworkId);
Future<Nothing> registered;
EXPECT_CALL(sched, registered(&driver, frameworkId, _))
.WillOnce(FutureSatisfy(®istered));
process::post(master.get(), frameworkPid, event);
AWAIT_READY(registered);
// Fail over the scheduler and expect a 'registered' call.
driver.stop(true);
MockScheduler sched2;
MesosSchedulerDriver driver2(
&sched2, frameworkInfo, master.get(), DEFAULT_CREDENTIAL);
frameworkRegisteredMessage =
DROP_MESSAGE(Eq(FrameworkRegisteredMessage().GetTypeName()), _, _);
driver2.start();
AWAIT_READY(frameworkRegisteredMessage);
UPID frameworkPid2 = frameworkRegisteredMessage.get().to;
Future<Nothing> registered2;
EXPECT_CALL(sched2, registered(&driver2, frameworkId, _))
.WillOnce(FutureSatisfy(®istered2));
process::post(master.get(), frameworkPid2, event);
AWAIT_READY(registered2);
}
int main(int argc, char** argv)
{
if (argc != 2) {
cerr << "Usage: " << argv[0] << " <master>" << endl;
return -1;
}
// Find this executable's directory to locate executor.
string uri;
Option<string> value = os::getenv("MESOS_BUILD_DIR");
if (value.isSome()) {
uri = path::join(value.get(), "src", "long-lived-executor");
} else {
uri = path::join(
os::realpath(Path(argv[0]).dirname()).get(),
"long-lived-executor");
}
ExecutorInfo executor;
executor.mutable_executor_id()->set_value("default");
executor.mutable_command()->set_value(uri);
executor.set_name("Long Lived Executor (C++)");
executor.set_source("cpp_long_lived_framework");
LongLivedScheduler scheduler(executor);
FrameworkInfo framework;
framework.set_user(""); // Have Mesos fill in the current user.
framework.set_name("Long Lived Framework (C++)");
value = os::getenv("MESOS_CHECKPOINT");
if (value.isSome()) {
framework.set_checkpoint(
numify<bool>(value.get()).get());
}
MesosSchedulerDriver* driver;
if (os::getenv("MESOS_AUTHENTICATE").isSome()) {
cout << "Enabling authentication for the framework" << endl;
value = os::getenv("DEFAULT_PRINCIPAL");
if (value.isNone()) {
EXIT(1) << "Expecting authentication principal in the environment";
}
Credential credential;
credential.set_principal(value.get());
framework.set_principal(value.get());
value = os::getenv("DEFAULT_SECRET");
if (value.isNone()) {
EXIT(1) << "Expecting authentication secret in the environment";
}
credential.set_secret(value.get());
driver = new MesosSchedulerDriver(
&scheduler, framework, argv[1], credential);
} else {
framework.set_principal("long-lived-framework-cpp");
driver = new MesosSchedulerDriver(
&scheduler, framework, argv[1]);
}
int status = driver->run() == DRIVER_STOPPED ? 0 : 1;
// Ensure that the driver process terminates.
driver->stop();
delete driver;
return status;
}
// This test verifies that status update manager ignores
// duplicate ACK for an earlier update when it is waiting
// for an ACK for a later update. This could happen when the
// duplicate ACK is for a retried update.
TEST_F(StatusUpdateManagerTest, IgnoreDuplicateStatusUpdateAck)
{
Try<PID<Master> > master = StartMaster();
ASSERT_SOME(master);
MockExecutor exec(DEFAULT_EXECUTOR_ID);
slave::Flags flags = CreateSlaveFlags();
flags.checkpoint = true;
Try<PID<Slave> > slave = StartSlave(&exec, flags);
ASSERT_SOME(slave);
FrameworkInfo frameworkInfo; // Bug in gcc 4.1.*, must assign on next line.
frameworkInfo = DEFAULT_FRAMEWORK_INFO;
frameworkInfo.set_checkpoint(true); // Enable checkpointing.
MockScheduler sched;
MesosSchedulerDriver driver(
&sched, frameworkInfo, master.get(), DEFAULT_CREDENTIAL);
FrameworkID frameworkId;
EXPECT_CALL(sched, registered(_, _, _))
.WillOnce(SaveArg<1>(&frameworkId));
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());
ExecutorDriver* execDriver;
EXPECT_CALL(exec, registered(_, _, _, _))
.WillOnce(SaveArg<0>(&execDriver));
EXPECT_CALL(exec, launchTask(_, _))
.WillOnce(SendStatusUpdateFromTask(TASK_RUNNING));
// Drop the first update, so that status update manager
// resends the update.
Future<StatusUpdateMessage> statusUpdateMessage =
DROP_PROTOBUF(StatusUpdateMessage(), master.get(), _);
Clock::pause();
driver.launchTasks(offers.get()[0].id(), createTasks(offers.get()[0]));
AWAIT_READY(statusUpdateMessage);
StatusUpdate update = statusUpdateMessage.get().update();
Future<TaskStatus> status;
EXPECT_CALL(sched, statusUpdate(_, _))
.WillOnce(FutureArg<1>(&status));
// This is the ACK for the retried update.
Future<Nothing> ack =
FUTURE_DISPATCH(_, &Slave::_statusUpdateAcknowledgement);
Clock::advance(slave::STATUS_UPDATE_RETRY_INTERVAL);
AWAIT_READY(status);
EXPECT_EQ(TASK_RUNNING, status.get().state());
AWAIT_READY(ack);
// Now send TASK_FINISHED update so that the status update manager
// is waiting for its ACK, which it never gets because we drop the
// update.
DROP_PROTOBUFS(StatusUpdateMessage(), master.get(), _);
Future<Nothing> update2 = FUTURE_DISPATCH(_, &Slave::_statusUpdate);
TaskStatus status2 = status.get();
status2.set_state(TASK_FINISHED);
execDriver->sendStatusUpdate(status2);
AWAIT_READY(update2);
// This is to catch the duplicate ack for TASK_RUNNING.
Future<Nothing> duplicateAck =
FUTURE_DISPATCH(_, &Slave::_statusUpdateAcknowledgement);
// Now send a duplicate ACK for the TASK_RUNNING update.
process::dispatch(
slave.get(),
&Slave::statusUpdateAcknowledgement,
update.slave_id(),
frameworkId,
update.status().task_id(),
update.uuid());
AWAIT_READY(duplicateAck);
Clock::resume();
EXPECT_CALL(exec, shutdown(_))
.Times(AtMost(1));
driver.stop();
driver.join();
Shutdown();
}
int main(int argc, char** argv)
{
Flags flags;
Try<flags::Warnings> load = flags.load("MESOS_EXAMPLE_", argc, argv);
if (load.isError()) {
std::cerr << flags.usage(load.error()) << std::endl;
return EXIT_FAILURE;
}
if (flags.help) {
std::cout << flags.usage() << std::endl;
return EXIT_SUCCESS;
}
mesos::internal::logging::initialize(argv[0], false);
// Log any flag warnings (after logging is initialized).
foreach (const flags::Warning& warning, load->warnings) {
LOG(WARNING) << warning.message;
}
if (flags.qps <= 0.0) {
EXIT(EXIT_FAILURE) << "Flag '--qps' needs to be greater than zero";
}
LoadGeneratorScheduler scheduler(flags.qps, flags.duration);
FrameworkInfo framework;
framework.set_user(""); // Have Mesos fill in the current user.
framework.set_principal(flags.principal);
framework.set_name(FRAMEWORK_NAME);
framework.set_checkpoint(flags.checkpoint);
framework.add_roles(flags.role);
framework.add_capabilities()->set_type(
FrameworkInfo::Capability::RESERVATION_REFINEMENT);
framework.set_checkpoint(flags.checkpoint);
if (flags.master == "local") {
// Configure master.
os::setenv("MESOS_ROLES", flags.role);
os::setenv("MESOS_AUTHENTICATE_FRAMEWORKS", stringify(flags.authenticate));
ACLs acls;
ACL::RegisterFramework* acl = acls.add_register_frameworks();
acl->mutable_principals()->set_type(ACL::Entity::ANY);
acl->mutable_roles()->add_values("*");
os::setenv("MESOS_ACLS", stringify(JSON::protobuf(acls)));
}
MesosSchedulerDriver* driver;
if (flags.authenticate) {
LOG(INFO) << "Enabling authentication for the framework";
Credential credential;
credential.set_principal(flags.principal);
if (flags.secret.isSome()) {
credential.set_secret(flags.secret.get());
}
driver = new MesosSchedulerDriver(
&scheduler,
framework,
flags.master,
credential);
} else {
driver = new MesosSchedulerDriver(
&scheduler,
framework,
flags.master);
}
int status = driver->run() == DRIVER_STOPPED ? EXIT_SUCCESS : EXIT_FAILURE;
// Ensure that the driver process terminates.
driver->stop();
delete driver;
return status;
}
// This test ensures that the command executor sends TASK_KILLING
// to frameworks that support the capability.
// TODO(hausdorff): Enable test. The executor tests use the replicated log
// by default. This is not currently supported on Windows, so they will all
// fail until that changes.
TEST_P_TEMP_DISABLED_ON_WINDOWS(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();
}
// This test verifies that status update manager ignores
// unexpected ACK for an earlier update when it is waiting
// for an ACK for another update. We do this by dropping ACKs
// for the original update and sending a random ACK to the slave.
TEST_F(StatusUpdateManagerTest, IgnoreUnexpectedStatusUpdateAck)
{
Try<PID<Master> > master = StartMaster();
ASSERT_SOME(master);
MockExecutor exec(DEFAULT_EXECUTOR_ID);
slave::Flags flags = CreateSlaveFlags();
flags.checkpoint = true;
Try<PID<Slave> > slave = StartSlave(&exec, flags);
ASSERT_SOME(slave);
FrameworkInfo frameworkInfo; // Bug in gcc 4.1.*, must assign on next line.
frameworkInfo = DEFAULT_FRAMEWORK_INFO;
frameworkInfo.set_checkpoint(true); // Enable checkpointing.
MockScheduler sched;
MesosSchedulerDriver driver(
&sched, frameworkInfo, master.get(), DEFAULT_CREDENTIAL);
FrameworkID frameworkId;
EXPECT_CALL(sched, registered(_, _, _))
.WillOnce(SaveArg<1>(&frameworkId));
Future<vector<Offer> > offers;
EXPECT_CALL(sched, resourceOffers(_, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
Future<TaskStatus> status;
EXPECT_CALL(sched, statusUpdate(_, _))
.WillOnce(FutureArg<1>(&status));
driver.start();
AWAIT_READY(offers);
EXPECT_NE(0u, offers.get().size());
ExecutorDriver* execDriver;
EXPECT_CALL(exec, registered(_, _, _, _))
.WillOnce(SaveArg<0>(&execDriver));
EXPECT_CALL(exec, launchTask(_, _))
.WillOnce(SendStatusUpdateFromTask(TASK_RUNNING));
Future<StatusUpdateMessage> statusUpdateMessage =
FUTURE_PROTOBUF(StatusUpdateMessage(), master.get(), _);
// Drop the ACKs, so that status update manager
// retries the update.
DROP_PROTOBUFS(StatusUpdateAcknowledgementMessage(), _, _);
driver.launchTasks(offers.get()[0].id(), createTasks(offers.get()[0]));
AWAIT_READY(statusUpdateMessage);
StatusUpdate update = statusUpdateMessage.get().update();
AWAIT_READY(status);
EXPECT_EQ(TASK_RUNNING, status.get().state());
Future<Nothing> unexpectedAck =
FUTURE_DISPATCH(_, &Slave::_statusUpdateAcknowledgement);
// Now send an ACK with a random UUID.
process::dispatch(
slave.get(),
&Slave::statusUpdateAcknowledgement,
update.slave_id(),
frameworkId,
update.status().task_id(),
UUID::random().toBytes());
AWAIT_READY(unexpectedAck);
EXPECT_CALL(exec, shutdown(_))
.Times(AtMost(1));
driver.stop();
driver.join();
Shutdown();
}
int main(int argc, char** argv) {
string execpath, master, remotecmd;
uint64_t mem = 64;
double cpus = -1.0, cores = -1.0;
shift;
while(true) {
const string s = argc > 0 ? argv[0] : "--help";
if(argc > 1 && s == "--exec-uri") {
execpath = argv[1];
shift; shift;
}
else if(argc > 1 && s == "--master") {
master = argv[1];
shift; shift;
}
else if(argc > 1 && s == "--cpus") {
cpus = (double)stof(argv[1]);
assert(cpus > 0.0);
shift; shift;
}
else if(argc > 1 && s == "--mem") {
mem = Bytes::parse(argv[1]).get().megabytes();
shift; shift;
}
else if(argc > 1 && s == "--cores") {
cores = (double)stof(argv[1]);
assert(cores> -1.0);
shift; shift;
}
else if(argc > 1 && s == "--remote-cmd") {
remotecmd = argv[1];
shift; shift;
}
else { break; }
}
if(master.length() == 0 || execpath.length() == 0 || cpus == -1) {
printf("Usage: chapel-scheduler --master <ip:port|required> --exec-uri <uri-to-the-location-of-a.out_real|required> --cpus <number-of-nodes|%N> --cores <number-of-cores to use|optional default is '2'> --mem <amt-of-ram-per-node|optional default is '64MB'> --remote-cmd %C\n");
exit(1);
}
const string& hostname = getHostname();
const uid_t uid = geteuid();
const struct passwd* pw = getpwuid(uid); // don't free this pointer! check the docs of this function to confirm!
FrameworkInfo framework;
framework.set_user(pw->pw_name);
framework.set_name("Chapel Framework");
//framework.set_role(role);
framework.set_principal("cpp");
ChapelScheduler scheduler(cpus, mem, cores, hostname, execpath, remotecmd, pw->pw_name);
// set up signal handler for clean shutdown
//
struct sigaction action;
action.sa_handler = SIGINTHandler;
sigemptyset(&action.sa_mask);
action.sa_flags = 0;
sigaction(SIGINT, &action, NULL);
sigaction(SIGHUP, &action, NULL);
sigaction(SIGQUIT, &action, NULL);
sigaction(SIGCHLD, &action, NULL);
sigaction(SIGSTOP, &action, NULL);
sigaction(SIGABRT, &action, NULL);
schedulerDriver = new MesosSchedulerDriver(&scheduler, framework, master);
const uint8_t status = schedulerDriver->run() == DRIVER_STOPPED ? 0 : 1;
if(status) {
cout << "Driver Stopped!" << endl;
scheduler.terminateAllTasks(schedulerDriver);
schedulerDriver.stop();
shutdown();
if(schedulerDriver != NULL) {
delete schedulerDriver;
schedulerDriver = NULL;
}
}
return status;
}
// This test verifies that the slave and status update manager
// properly handle duplicate terminal status updates, when the
// second update is received after the ACK for the first update.
// The proper behavior here is for the status update manager to
// forward the duplicate update to the scheduler.
TEST_F(StatusUpdateManagerTest, DuplicateTerminalUpdateAfterAck)
{
Try<PID<Master> > master = StartMaster();
ASSERT_SOME(master);
MockExecutor exec(DEFAULT_EXECUTOR_ID);
slave::Flags flags = CreateSlaveFlags();
flags.checkpoint = true;
Try<PID<Slave> > slave = StartSlave(&exec, flags);
ASSERT_SOME(slave);
FrameworkInfo frameworkInfo; // Bug in gcc 4.1.*, must assign on next line.
frameworkInfo = DEFAULT_FRAMEWORK_INFO;
frameworkInfo.set_checkpoint(true); // Enable checkpointing.
MockScheduler sched;
MesosSchedulerDriver driver(
&sched, frameworkInfo, master.get(), DEFAULT_CREDENTIAL);
FrameworkID frameworkId;
EXPECT_CALL(sched, registered(_, _, _))
.WillOnce(SaveArg<1>(&frameworkId));
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());
ExecutorDriver* execDriver;
EXPECT_CALL(exec, registered(_, _, _, _))
.WillOnce(SaveArg<0>(&execDriver));
// Send a terminal update right away.
EXPECT_CALL(exec, launchTask(_, _))
.WillOnce(SendStatusUpdateFromTask(TASK_FINISHED));
Future<TaskStatus> status;
EXPECT_CALL(sched, statusUpdate(_, _))
.WillOnce(FutureArg<1>(&status));
Future<Nothing> _statusUpdateAcknowledgement =
FUTURE_DISPATCH(slave.get(), &Slave::_statusUpdateAcknowledgement);
driver.launchTasks(offers.get()[0].id(), createTasks(offers.get()[0]));
AWAIT_READY(status);
EXPECT_EQ(TASK_FINISHED, status.get().state());
AWAIT_READY(_statusUpdateAcknowledgement);
Future<TaskStatus> update;
EXPECT_CALL(sched, statusUpdate(_, _))
.WillOnce(FutureArg<1>(&update));
Future<Nothing> _statusUpdateAcknowledgement2 =
FUTURE_DISPATCH(slave.get(), &Slave::_statusUpdateAcknowledgement);
Clock::pause();
// Now send a TASK_KILLED update for the same task.
TaskStatus status2 = status.get();
status2.set_state(TASK_KILLED);
execDriver->sendStatusUpdate(status2);
// Ensure the scheduler receives TASK_KILLED.
AWAIT_READY(update);
EXPECT_EQ(TASK_KILLED, update.get().state());
// Ensure the slave properly handles the ACK.
// Clock::settle() ensures that the slave successfully
// executes Slave::_statusUpdateAcknowledgement().
AWAIT_READY(_statusUpdateAcknowledgement2);
Clock::settle();
Clock::resume();
EXPECT_CALL(exec, shutdown(_))
.Times(AtMost(1));
driver.stop();
driver.join();
Shutdown();
}
