Amphora Driver Interface

Amphora Driver Interface

https://blueprints.launchpad.net/octavia/+spec/amphora-driver-interface

This blueprint describes how a driver will interface with the controller. It will describe the base class and other classes required. It will not describe the REST interface needed to talk to an amphora nor how health information or statistics are gathered from the amphora.

Problem description

The controller needs to talk through a driver to the amphora to allow for custom APIs and custom rendering of configuration data for different amphora implementations.

The controller will heavily utilize taskflow [2] to accomplish its goals so it is highly encouraged for drivers to use taskflow to organize their work, too.

Proposed change

Establish a base class to model the desire functionality:

class AmphoraLoadBalancerDriver(object):

    def update(self, listener, vip):
        """updates the amphora with a new configuration

        for the listener on the vip.
        """
        raise NotImplementedError

    def stop(self, listener, vip):
        """stops the listener on the vip."""
        return None

    def start(self, listener, vip):
        """starts the listener on the vip."""
        return None

    def delete(self, listener, vip):
        """deletes the listener on the vip."""
        raise NotImplementedError

    def get_info(self, amphora):
        """Get detailed information about an amphora

        returns information about the amphora, e.g. {"Rest Interface":
        "1.0", "Amphorae": "1.0", "packages":{"ha proxy":"1.5"},
        "network-interfaces": {"eth0":{"ip":...}} some information might
        come from querying the amphora
        """
        raise NotImplementedError

    def get_diagnostics(self, amphora):
        """OPTIONAL - Run diagnostics

        run some expensive self tests to determine if the amphora and the
        lbs are healthy the idea is that those tests are triggered more
        infrequent than the heartbeat
        """
        raise NotImplementedError

    def finalize_amphora(self, amphora):
        """OPTIONAL - called once an amphora has been build but before

        any listeners are configured. This is a hook for drivers who need
        to do additional work before am amphora becomes ready to accept
        listeners. Please keep in mind that amphora might be kept in am
        offline pool after this call.
        """
        pass

    def post_network_plug(self, amphora, port):
        """OPTIONAL - called after adding a compute instance to a network.

        This will perform any necessary actions to allow for connectivity
        for that network on that instance.

        port is an instance of octavia.network.data_models.Port.  It
        contains information about the port, subnet, and network that
        was just plugged.
        """

    def post_vip_plug(self, load_balancer, amphorae_network_config):
        """OPTIONAL - called after plug_vip method of the network driver.

        This is to do any additional work needed on the amphorae to plug
        the vip, such as bring up interfaces.

        amphorae_network_config is a dictionary of objects that include
        network specific information about each amphora's connections.
        """

    def start_health_check(self, health_mixin):
        """start check health

        :param health_mixin: health mixin object
        :type amphora: object

        Start listener process and  calls HealthMixin to update
        databases information.
        """
        pass

    def stop_health_check(self):
        """stop check health

        Stop listener process and  calls HealthMixin to update
        databases information.
        """
        pass

The referenced listener is a listener object and vip a vip as described in our model. The model is detached from the DB so the driver can’t write to the DB. Because our initial goal is to render a whole config no special methods for adding nodes, health monitors, etc. are supported at this juncture. This might be added in later versions.

No method for obtaining logs has been added. This will be done in a future blueprint.

Exception Model

The driver is expected to raise the following well defined exceptions

  • NotImplementedError - this functionality is not implemented/not supported

  • AmphoraDriverError - a super class for all other exceptions and the catch

    all if no specific exception can be found

    • NotFoundError - this amphora couldn’t be found/ was deleted by nova

    • InfoException - gathering information about this amphora failed

    • NetworkConfigException - gathering network information failed

    • UnauthorizedException - the driver can’t access the amphora

    • TimeOutException - contacting the amphora timed out

    • UnavailableException - the amphora is temporary unavailable

    • SuspendFaied - this load balancer couldn’t be suspended

    • EnableFailed - this load balancer couldn’t be enabled

    • DeleteFailed - this load balancer couldn’t be deleted

    • ProvisioningErrors - those are errors which happen during provisioning

      • ListenerProvisioningError - could not provision Listener
      • LoadBalancerProvisoningError - could not provision LoadBalancer
      • HealthMonitorProvisioningError - could not provision HealthMonitor
      • NodeProvisioningError - could not provision Node

Health and Stat Mixin

It has been suggested to gather health and statistic information via UDP packets emitted from the amphora. This requires each driver to spin up a thread to listen on a UDP port and then hand the information to the controller as a mixin to make sense of it.

Here is the mixin definition:

class HealthMixIn(object):
    def update_health(health):
        #map: {"amphora-status":HEALTHY, loadbalancers: {"loadbalancer-id": {"loadbalancer-status": HEALTHY,
        # "listeners":{"listener-id":{"listener-status":HEALTHY, "nodes":{"node-id":HEALTHY, ...}}, ...}, ...}}
        # only items whose health has changed need to be submitted
        # awesome update code
        pass

class StatsMixIn(object):
    def update_stats(stats):
        #uses map {"loadbalancer-id":{"listener-id": {"bytes-in": 123, "bytes_out":123, "active_connections":123,
        # "total_connections", 123}, ...}
        # elements are named to keep it extensible for future versions
        #awesome update code and code to send to ceilometer
        pass

Things a good driver should do:

  • Non blocking IO - throw an appropriate exception instead to wait forever; use timeouts on sockets
  • We might employ a circuit breaker to insulate driver problems from controller problems [1]
  • Use appropriate logging
  • Use the preferred threading model

This will be demonstrated in the Noop-driver code.

Alternatives

Require all amphora to implement a common REST interface and use that as the integration point.

Data model impact

None

REST API impact

None

Security impact

None

Notifications impact

None - since initial version

Other end user impact

None

Performance Impact

Minimal

Other deployer impact

Deployers need to make sure to bundle the compatible versions of amphora, driver, controller –

Developer impact

Need to write towards this clean interface.

Implementation

Assignee(s)

German Eichberger

Work Items

  • Write abstract interface
  • Write Noop driver
  • Write tests

Dependencies

None

Testing

  • Unit tests with tox and Noop-Driver
  • tempest tests with Noop-Driver

Documentation Impact

None - we won’t document the interface for 0.5. If that changes we need to write an interface documentation so 3rd party drivers know what we expect.

Creative Commons Attribution 3.0 License

Except where otherwise noted, this document is licensed under Creative Commons Attribution 3.0 License. See all OpenStack Legal Documents.