void ResourceProviderManagerProcess::subscribe(
const HttpConnection& http,
const Call::Subscribe& subscribe)
{
ResourceProviderInfo resourceProviderInfo =
subscribe.resource_provider_info();
resourceProviderInfo.mutable_id()->CopyFrom(newResourceProviderId());
ResourceProvider resourceProvider(resourceProviderInfo, http);
Event event;
event.set_type(Event::SUBSCRIBED);
event.mutable_subscribed()->mutable_provider_id()->CopyFrom(
resourceProvider.info.id());
if (!resourceProvider.http.send(event)) {
LOG(WARNING) << "Unable to send event to resource provider "
<< stringify(resourceProvider.info.id())
<< ": connection closed";
}
resourceProviders.put(resourceProviderInfo.id(), std::move(resourceProvider));
}
bool operator==(
const ResourceProviderInfo& left,
const ResourceProviderInfo& right)
{
if (left.id() != right.id()) {
return false;
}
if (Attributes(left.attributes()) != Attributes(right.attributes())) {
return false;
}
if (Resources(left.resources()) != Resources(right.resources())) {
return false;
}
return true;
}
// 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());
}
}
bool operator==(
const ResourceProviderInfo& left,
const ResourceProviderInfo& right)
{
// Order of reservations is important.
if (left.default_reservations_size() != right.default_reservations_size()) {
return false;
}
for (int i = 0; i < left.default_reservations_size(); i++) {
if (left.default_reservations(i) != right.default_reservations(i)) {
return false;
}
}
return left.has_id() == right.has_id() &&
(!left.has_id() || left.id() == right.id()) &&
Attributes(left.attributes()) == Attributes(right.attributes()) &&
left.type() == right.type() &&
left.name() == right.name() &&
left.has_storage() == right.has_storage() &&
(!left.has_storage() || left.storage() == right.storage());
}