You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@mesos.apache.org by ab...@apache.org on 2019/08/26 13:03:27 UTC
[mesos] branch 1.8.x updated (f3aa802 -> adc958f)
This is an automated email from the ASF dual-hosted git repository.
abudnik pushed a change to branch 1.8.x
in repository https://gitbox.apache.org/repos/asf/mesos.git.
from f3aa802 Added MESOS-9836 to the 1.8.2 CHANGELOG.
new 14abb82 Added missing `return` statement in `Slave::statusUpdate`.
new 4bbb037 Fixed out-of-order processing of terminal status updates in agent.
new adc958f Added MESOS-9887 to the 1.8.2 CHANGELOG.
The 3 revisions listed above as "new" are entirely new to this
repository and will be described in separate emails. The revisions
listed as "add" were already present in the repository and have only
been added to this reference.
Summary of changes:
CHANGELOG | 1 +
src/slave/slave.cpp | 64 ++++++++++++++++++++++++++++++++++++++++++++++++++---
src/slave/slave.hpp | 6 +++++
3 files changed, 68 insertions(+), 3 deletions(-)
[mesos] 02/03: Fixed out-of-order processing of terminal status
updates in agent.
Posted by ab...@apache.org.
This is an automated email from the ASF dual-hosted git repository.
abudnik pushed a commit to branch 1.8.x
in repository https://gitbox.apache.org/repos/asf/mesos.git
commit 4bbb0376cd584a4160a2c5c2f0ac4f3ecaa5e622
Author: Andrei Budnik <ab...@apache.org>
AuthorDate: Tue Aug 20 19:24:44 2019 +0200
Fixed out-of-order processing of terminal status updates in agent.
Previously, Mesos agent could send TASK_FAILED status update on
executor termination while processing of TASK_FINISHED status update
was in progress. Processing of task status updates involves sending
requests to the containerizer, which might finish processing of these
requests out-of-order, e.g. `MesosContainerizer::status`. Also,
the agent does not overwrite status of the terminal status update once
it's stored in the `terminatedTasks`. Hence, there was a race condition
between two terminal status updates.
Note that V1 Executors are not affected by this problem because they
wait for an acknowledgement of the terminal status update by the agent
before terminating.
This patch introduces a new data structure `pendingStatusUpdates`,
which holds a list of status updates that are being processed. This
data structure allows validating the order of processing of status
updates by the agent.
Review: https://reviews.apache.org/r/71343
---
src/slave/slave.cpp | 62 ++++++++++++++++++++++++++++++++++++++++++++++++++---
src/slave/slave.hpp | 6 ++++++
2 files changed, 65 insertions(+), 3 deletions(-)
diff --git a/src/slave/slave.cpp b/src/slave/slave.cpp
index 50a7d68..8d8cef3 100644
--- a/src/slave/slave.cpp
+++ b/src/slave/slave.cpp
@@ -5727,6 +5727,8 @@ void Slave::statusUpdate(StatusUpdate update, const Option<UPID>& pid)
metrics.valid_status_updates++;
+ executor->addPendingTaskStatus(status);
+
// Before sending update, we need to retrieve the container status
// if the task reached the executor. For tasks that are queued, we
// do not need to send the container status and we must
@@ -5938,6 +5940,17 @@ void Slave::___statusUpdate(
VLOG(1) << "Task status update manager successfully handled status update "
<< update;
+ const TaskStatus& status = update.status();
+
+ Executor* executor = nullptr;
+ Framework* framework = getFramework(update.framework_id());
+ if (framework != nullptr) {
+ executor = framework->getExecutor(status.task_id());
+ if (executor != nullptr) {
+ executor->removePendingTaskStatus(status);
+ }
+ }
+
if (pid == UPID()) {
return;
}
@@ -5945,7 +5958,7 @@ void Slave::___statusUpdate(
StatusUpdateAcknowledgementMessage message;
message.mutable_framework_id()->MergeFrom(update.framework_id());
message.mutable_slave_id()->MergeFrom(update.slave_id());
- message.mutable_task_id()->MergeFrom(update.status().task_id());
+ message.mutable_task_id()->MergeFrom(status.task_id());
message.set_uuid(update.uuid());
// Task status update manager successfully handled the status update.
@@ -5957,14 +5970,12 @@ void Slave::___statusUpdate(
send(pid.get(), message);
} else {
// Acknowledge the HTTP based executor.
- Framework* framework = getFramework(update.framework_id());
if (framework == nullptr) {
LOG(WARNING) << "Ignoring sending acknowledgement for status update "
<< update << " of unknown framework";
return;
}
- Executor* executor = framework->getExecutor(update.status().task_id());
if (executor == nullptr) {
// Refer to the comments in 'statusUpdate()' on when this can
// happen.
@@ -10520,6 +10531,33 @@ void Executor::recoverTask(const TaskState& state, bool recheckpointTask)
}
+void Executor::addPendingTaskStatus(const TaskStatus& status)
+{
+ auto uuid = id::UUID::fromBytes(status.uuid()).get();
+ pendingStatusUpdates[status.task_id()][uuid] = status;
+}
+
+
+void Executor::removePendingTaskStatus(const TaskStatus& status)
+{
+ const TaskID& taskId = status.task_id();
+
+ auto uuid = id::UUID::fromBytes(status.uuid()).get();
+
+ if (!pendingStatusUpdates.contains(taskId) ||
+ !pendingStatusUpdates[taskId].contains(uuid)) {
+ LOG(WARNING) << "Unknown pending status update (uuid: " << uuid << ")";
+ return;
+ }
+
+ pendingStatusUpdates[taskId].erase(uuid);
+
+ if (pendingStatusUpdates[taskId].empty()) {
+ pendingStatusUpdates.erase(taskId);
+ }
+}
+
+
Try<Nothing> Executor::updateTaskState(const TaskStatus& status)
{
bool terminal = protobuf::isTerminalState(status.state());
@@ -10543,6 +10581,24 @@ Try<Nothing> Executor::updateTaskState(const TaskStatus& status)
task = launchedTasks.at(status.task_id());
if (terminal) {
+ if (pendingStatusUpdates.contains(status.task_id())) {
+ auto statusUpdates = pendingStatusUpdates[status.task_id()].values();
+
+ auto firstTerminal = std::find_if(
+ statusUpdates.begin(),
+ statusUpdates.end(),
+ [](const TaskStatus& status) {
+ return protobuf::isTerminalState(status.state());
+ });
+
+ CHECK(firstTerminal != statusUpdates.end());
+
+ if (firstTerminal->uuid() != status.uuid()) {
+ return Error("Unexpected terminal status update after first status"
+ " update " + stringify(firstTerminal->state()));
+ }
+ }
+
launchedTasks.erase(taskId);
}
} else if (terminatedTasks.contains(taskId)) {
diff --git a/src/slave/slave.hpp b/src/slave/slave.hpp
index c740bf7..bb14f5f 100644
--- a/src/slave/slave.hpp
+++ b/src/slave/slave.hpp
@@ -918,6 +918,9 @@ public:
void recoverTask(const state::TaskState& state, bool recheckpointTask);
+ void addPendingTaskStatus(const TaskStatus& status);
+ void removePendingTaskStatus(const TaskStatus& status);
+
Try<Nothing> updateTaskState(const TaskStatus& status);
// Returns true if there are any queued/launched/terminated tasks.
@@ -1062,6 +1065,9 @@ public:
// non-terminal tasks.
Option<mesos::slave::ContainerTermination> pendingTermination;
+ // Task status updates that are being processed by the agent.
+ hashmap<TaskID, LinkedHashMap<id::UUID, TaskStatus>> pendingStatusUpdates;
+
private:
Executor(const Executor&) = delete;
Executor& operator=(const Executor&) = delete;
[mesos] 01/03: Added missing `return` statement in
`Slave::statusUpdate`.
Posted by ab...@apache.org.
This is an automated email from the ASF dual-hosted git repository.
abudnik pushed a commit to branch 1.8.x
in repository https://gitbox.apache.org/repos/asf/mesos.git
commit 14abb82925cdbce746238bc20dc7b8c279a96a67
Author: Andrei Budnik <ab...@apache.org>
AuthorDate: Fri Aug 23 14:36:18 2019 +0200
Added missing `return` statement in `Slave::statusUpdate`.
Previously, if `statusUpdate` was called for a pending task, it would
forward the status update and then continue executing `statusUpdate`,
which then checks if there is an executor that is aware of this task.
Given that a pending task is not known to any executor, it would always
handle it by forwarding status update one more time. This patch adds
missing `return` statement, which fixes the issue.
Review: https://reviews.apache.org/r/71361
---
src/slave/slave.cpp | 2 ++
1 file changed, 2 insertions(+)
diff --git a/src/slave/slave.cpp b/src/slave/slave.cpp
index bf87be0..50a7d68 100644
--- a/src/slave/slave.cpp
+++ b/src/slave/slave.cpp
@@ -5659,6 +5659,8 @@ void Slave::statusUpdate(StatusUpdate update, const Option<UPID>& pid)
taskStatusUpdateManager->update(update, info.id())
.onAny(defer(self(), &Slave::___statusUpdate, lambda::_1, update, pid));
+
+ return;
}
Executor* executor = framework->getExecutor(status.task_id());
[mesos] 03/03: Added MESOS-9887 to the 1.8.2 CHANGELOG.
Posted by ab...@apache.org.
This is an automated email from the ASF dual-hosted git repository.
abudnik pushed a commit to branch 1.8.x
in repository https://gitbox.apache.org/repos/asf/mesos.git
commit adc958f553c3728aab5529de56b0ddc30c0f9b68
Author: Andrei Budnik <ab...@apache.org>
AuthorDate: Mon Aug 26 15:02:40 2019 +0200
Added MESOS-9887 to the 1.8.2 CHANGELOG.
---
CHANGELOG | 1 +
1 file changed, 1 insertion(+)
diff --git a/CHANGELOG b/CHANGELOG
index b3fca25..ff89605 100644
--- a/CHANGELOG
+++ b/CHANGELOG
@@ -6,6 +6,7 @@ Release Notes - Mesos - Version 1.8.2 (WIP)
* [MESOS-9785] - Frameworks recovered from reregistered agents are not reported to master `/api/v1` subscribers.
* [MESOS-9836] - Docker containerizer overwrites `/mesos/slave` cgroups.
* [MESOS-9868] - NetworkInfo from the agent /state endpoint is not correct.
+ * [MESOS-9887] - Race condition between two terminal task status updates for Docker/Command executor.
* [MESOS-9893] - `volume/secret` isolator should cleanup the stored secret from runtime directory when the container is destroyed.
* [MESOS-9925] - Default executor takes a couple of seconds to start and subscribe Mesos agent.