Health Manager 2.0 (HM-2) is a complete re-write of the original Health Manager.
HM-2 monitors the state of the applications and ensures that started applications are indeed running, their versions and number of instances correct.
Conceptually, this is done by maintaining a Known State of applications and comparing it against the Expected State. When discrepancies are found, actions are initiated to bring the applications to the Expected State, e.g., start/stop commands are issued for missing/extra instances, respectively.
Additionally, Health Manager collects and exposes statistics and health status for individual applications, as well as aggregates for frameworks, runtimes, etc.
The state of each application is represented by an instance of an aptly named class AppState. AppState gets forwarded important state-changing messages (i.e. hearbeats and exit signals), updates its internal state accordingly and then invokes registered event handlers. It is the job of these handlers (housed in the Harmonizer, see below) to enforce complex policies, e.g., whether to restart application, if so, with which priority, etc.
HM-2 comprises the following components:
- Manager
- Harmonizer
- Scheduler
- ExpectedStateProvider
- KnownStateProvider
- Nudger
- Reporter
- Shadower
Provides an entry point, configures, initializes and registers other components.
Expresses the policy of bringing the applications to the Expected State by observing the Known State.
Harmonizer sets up the interactions between other components, and aims to achieve clarity of the intent through delegation:
Known State and Expected State are compared periodically with the use of the Scheduler and Nudger actions are Scheduled to bring the States into harmony.
Encapsulates EventMachine-related functionality such as timer setup and cancellation, quantization of long-running tasks to prevent EM Reactor loop blocking.
Provides the expected state of the application, e.g., whether the application was Started or Stopped, how many instances should be running, etc. This information comes from the Cloud Controller by way of http-based Bulk API, hence the concrete class is BulkBasedExpectedStateProvider
The Known state will be discovered from NatsBasedKnownStateProvider, that will listen to heartbeat and other messages on the NATS bus.
The State of each application is represented by an instant of object
AppState. That object receives updates of the application state,
stores them and notifies registered listeners about events, such as
instances_missing
, etc.
Nudger is the interface for health manager to effect the change on the
world, by dispatching cloudcontrollers.hm.requests
messages
that instruct CCs to start or stop instances. Nudger maintains a
priority queue of these requests, and deques the messages by a
batchful.
Reporter responds to healthmanager.status
and healthmanager.health
requests.
The Shadower component exists primarily to smoothen the transition from the original HealthManager to this new HealthManager 2.0
The Shadower implements shadow-mode for HM-2, where it quietly and passively observes the state of the world and comes up with harmonization decisions, but rather than publishing harmonization messages on the NATS bus, it keeps track of them. It also listens for harmonization messages (most likely coming from the original HealthManager) and compares those messages with the ones HM-2 has produced. Ideally, these two sets of messages should perfectly overlap. When they don't, shadower issues warnings.
When in shadow-mode, all outgoing messages are sent to the Shadower, rather than being published to NATS.
Conceptually, harmonization happens in two ways:
- by reacting to messages (such as
droplet.exited
); - by periodically scanning the world and enumerating applications, looking for anomalies.
There are three distinct scenarios possible when droplet.exited
signal arrives:
-
application is stopped; means the application was stopped explicitly, no action required;
-
DEA evacuation; the DEA is being evacuated and all instances from that DEA need to be restarted somewhere else. HM-2 initiates that restarting;
-
application instance crashed; That instance needs to be restarted unless it crashed multiple times in short period of time, in which case it is declared
flapping
. See more on this below.
An instance of application is declared flapping
if it crashed more
than flapping_death
times within flapping_timeout
seconds. There
are several possible reasons for flapping:
- app is completely broken and simply does not start;
- app has a bug that results in a crash every once in a while;
- app has a dependency on the external world or a CF-provisioned service, and that dependency is unavailable, perhaps temporarily, resulting in app repeatedly crashing.
Handling flapping apps is hard. We'd like to:
- make the best effort to restart an app, when it makes sense;
- provide the crashlogs for crashing instances;
- cut down on the overhead associated with restarting an app, particularly relating to moving application bits to DEA and storing it there.
- avoid IO spikes due to massive simultaneous restarts
In order to accomodate these conflicting requirements, the following policy for flapping instances (FI) adopted:
- initially the FI is restarted with a delay defined by
min_restart_delay
config value; - for each subsequent crash, the delay is doubled, but not to exceed
max_restart_delay
config value; - a random noise is added to the value of delay, its maximum absolute value defined by
delay_time_noise
config value; - if the number of crashes for a given FI exceeds
giveup_crash_number
, give up restarting attempts. This behavior can be turned off.
DEAs peridically send out heartbeat messages on NATS bus. These heartbeats contain DEA identifying information, as well as information on application instances running on respective DEAs.
The heartbeats are used to establish "missing" and "extra" indices. Missing indices are then commanded to start, extra indices are commanded to stop.
AppState object tracks heartbeats for each instance of each version.
An instance is "missing" if a live version of this instance has not
received a heartbeat in the last droplet_lost
seconds.
However, an instance_missing event is only triggered if the AppState
was not reset recently, and if check_for_missing_instances
method
has been invoked.
To file a bug against Cloud Foundry Open Source and its components, sign up and use our bug tracking system: http://cloudfoundry.atlassian.net