When Aurora reads a configuration file and finds a
Job definition, it:
Jobinto its constituent
Tasks to the scheduler.
PENDINGstate, starting each
Task’s life cycle.
Please note, a couple of task states described below are missing from this state diagram.
Task is in the
PENDING state, the scheduler constantly
searches for machines satisfying that
Task’s resource request
requirements (RAM, disk space, CPU time) while maintaining configuration
constraints such as “a
Task must run on machines dedicated to a
particular role” or attribute limit constraints such as “at most 2
Tasks from the same
Job may run on each rack”. When the scheduler
finds a suitable match, it assigns the
Task to a machine and puts the
Task into the
ASSIGNED state, the scheduler sends an RPC to the agent
Task configuration, which the agent uses to spawn
an executor responsible for the
Task’s lifecycle. When the scheduler
receives an acknowledgment that the machine has accepted the
Task goes into
STARTING state initializes a
Task sandbox. When the sandbox is fully
initialized, Thermos begins to invoke
Processes. Also, the agent
machine sends an update to the scheduler that the
RUNNING state, only after the task satisfies the liveness requirements.
See Health Checking for more details
for how to configure health checks.
There are various ways that an active
Task can transition into a terminal
state. By definition, it can never leave this state. However, depending on
nature of the termination and the originating
max_task_failures), a replacement
Task might be
Task can terminate without direct user interaction. For
example, it may be a finite computation that finishes, even something as
echo hello world., or it could be an exceptional condition in
a long-lived service. If the
Task is successful (its underlying
processes have succeeded with exit status
0 or finished without
reaching failure limits) it moves into
FINISHED state. If it finished
after reaching a set of failure limits, it goes into
TASK which is subject to rescheduling will be temporarily
THROTTLED, if it is considered to be flapping. A task is flapping, if its
previous invocation was terminated after less than 5 minutes (scheduler
default). The time penalty a task has to remain in the
before it is eligible for rescheduling, increases with each consecutive
You can terminate a
Task by issuing an
aurora job kill command, which
moves it into
KILLING state. The scheduler then sends the agent a
request to terminate the
Task. If the scheduler receives a successful
response, it moves the Task into
KILLED state and never restarts it.
Task is forced into the
RESTARTING state via the
aurora job restart
command, the scheduler kills the underlying task but in parallel schedules
an identical replacement for it.
In any case, the responsible executor on the agent follows an escalation sequence when killing a running task:
HttpLifecycleConfigis not present, skip to (4).
graceful_shutdown_endpointand wait 5 seconds.
shutdown_endpointand wait 5 seconds.
kill) and wait at most
If the executor notices that all
Processes in a
Task have aborted
during this sequence, it will not proceed with subsequent steps.
Note that graceful shutdown is best-effort, and due to the many
inevitable realities of distributed systems, it may not be performed.
Task stays in a transient task state for too long (such as
STARTING), the scheduler forces it into
LOST state, creating a new
Task in its place that’s sent into
In addition, if the Mesos core tells the scheduler that a agent has
become unhealthy (or outright disappeared), the
Tasks assigned to that
agent go into
LOST state and new
Tasks are created in their place.
PENDING state, there is no guarantee a
Task will be reassigned
to the same machine unless job constraints explicitly force it there.
Sometimes a Task needs to be interrupted, such as when a non-production
Task’s resources are needed by a higher priority production Task. This
type of interruption is called a pre-emption. When this happens in
Aurora, the non-production Task is killed and moved into
PREEMPTING state when both the following are true:
PENDINGproduction task that hasn’t been scheduled due to a lack of resources.
The scheduler UI shows the non-production task was preempted in favor of
the production task. At some point, tasks in
PREEMPTING move to
Note that non-production tasks consuming many resources are likely to be preempted in favor of production tasks.
Cluster operators can set agent into maintenance mode. This will transition
Task running on this agent into
DRAINING and eventually to
Tasks will be restarted on other agents for which no maintenance
has been announced yet.
Due to the many inevitable realities of distributed systems, there might
be a mismatch of perceived and actual cluster state (e.g. a machine returns
netsplit but the scheduler has already marked all its
LOST and rescheduled them).
Aurora regularly runs a state reconciliation process in order to detect
and correct such issues (e.g. by killing the errant
By default, the proper detection of all failure scenarios and inconsistencies
may take up to an hour.
To emphasize this point: there is no uniqueness guarantee for a single
instance of a job in the presence of network partitions. If the
requires that, it should be baked in at the application level using a
distributed coordination service such as Zookeeper.