Allocation API

https://storyboard.openstack.org/#!/story/2004341

This spec proposed creating of API for allocation of nodes for deployment.

Problem description

The users of standalone ironic do not have an out-of-box means to find a suitable node to deploy onto. The metalsmith project was created to add this gap short-term, but it is not suitable for consumer code that is not written in Python. A potential consumer is a K8S provider for standalone ironic.

The API user story is as follows:

Given a resource class and, optionally, a list of required traits, return me an available bare metal node and set instance_uuid on it to make it as reserved.

Proposed change

Overview

This RFE proposed a new ReST API endpoint /v1/allocations that will initially allow creating and deleting Allocation resources.

Two implementations of the allocation process are planned:

  1. Via the database, similar to how metalsmith now operates.

  2. Via the Placement service, similar to how nova currently operates.

This spec concentrates on the API design and the first (standalone) case.

Allocation process

An allocation happens as follows:

  1. An API client sends a POST /v1/allocations request, specifying a resource class, and optionally traits and node UUID.

  2. The allocation request is routed to a random available conductor.

  3. The conductor creates an Allocation object in the database with state=allocating and conductor_affinity=<host name>.

  4. A thread is spawned for the remaining allocation process, and the allocation object is returned to the caller.

Allocation: database backend

The following actions are done by the conductor handling the allocation when database is used as backend (the only option in this spec):

  1. Fetch list of nodes from the database with:

    • provision_state=available

    • maintenance=False

    • power_state!=None

      Note

      This is required for compatibility with really old API versions that allow creating nodes directly in the available state.

    • instance_uuid=None

    • resource_class=<requested resource class>

    • uuid in the list of candidate nodes (if provided)

    • requested traits are a superset of node traits

  2. If the list is empty, set allocation’s state to error and last_error to the explanation.

  3. Shuffle the list, so that several processes do not try reserving nodes in the same order.

  4. Acquire a lock on the first node. If locking succeeds, check that the assumptions are still valid about this node, and reserve it by setting its instance_uuid to the uuid of the allocation. In the same database transaction:

    • set allocation’s node_id to the node’s ID,

    • set allocation’s state to active,

    • set node’s allocation_id to the allocation’s ID,

    • add matched traits to node’s instance_info.

    Note

    Since the conductor may not have the hardware type for the selected node, we will update TaskManager to avoid constructing the driver object.

  5. If something fails on the previous step, proceed to the next node. If no more nodes are left, set the allocation’s state to error and last_error to the explanation.

Deallocation: database backend

The deallocation process will in one transation:

  • unset node’s instance_uuid,

  • unset node’s allocation_id,

  • delete the allocation.

The deallocation is triggered either explicitly via API or when a node is torn down (at the same time when node’s instance_uuid and instance_info are purged).

Note

In the future we might consider supporting sticky allocations which survive node’s tear down. This is outside the scope of this spec.

There is one important difference between using just instance_uuid and using the allocation API: instance_uuid can be set and unset for active nodes, while for allocations it will be forbidden. The reason is that with the future Placement backend removing an allocation would mark the node as available in Placement.

HA and take over

  • When a conductor restarts, it fetches allocations with

    • conductor_affinity=<host name>

    • state=allocating

    and starts the allocation procedure for each of them.

  • If the conductor handling an allocation stops without a replacement, the reservation becomes orphaned. All conductors periodically fetch list of allocations belonging to dead conductors and each tries to resume them.

    First, it tries to update the conductor_affinity by doing something like:

    UPDATE allocations SET conductor_affinity=<new host name>
       WHERE id=<allocation ID> AND conductor_affinity=<dead host name>
    

    If the query updated 1 row, we know that the new conductor now manages the allocation. Otherwise we know that another conductor took it over.

  • To avoid rare races with this take over procedure, the normal update will also look like:

    UPDATE allocations SET <new values>
        WHERE id=<allocation ID> AND conductor_affinity=<current host name>
    

Alternatives

  • Make each consumer invent their own allocation procedure or use metalsmith.

  • Write a new service for managing reservations (probably based on metalsmith code base).

  • Make the API blocking and avoid having states for allocations. Such an approach would result in easier API and implementation, but it can be problematic when using an external system, such as Placement, as a backend, since the requests to it make block the RPC thread.

    Additionally, the asynchronous design will make it easier to introduce a bulk allocation API in the future, if we decide so.

Data model impact

Introduce a new database/RPC object Allocation with the following fields:

  • id internal integer ID, not exposed to users.

  • uuid unique UUID of the allocation.

  • name unique name of the allocation, follows the same format as node’s names.

    Note

    This field is useful, for example, for systems using host names, like metalsmith.

  • node_id reserved node ID (can be null) - foreign key to the nodes table.

  • updated_at/created_at standard update/creation date time fields.

  • resource_class mandatory requested resource class.

  • candidate_nodes list of node UUIDs to choose from (can be null).

    Note

    This allows a caller to pre-filter nodes by arbitrary criteria.

  • state allocation state, see State Machine Impact.

  • last_error last error message.

  • conductor_affinity internal field, specifying which conductor currently handles this allocation.

Introduce a helper table allocation_traits mapping an allocation to its requested traits (very similar to node_traits).

Update the nodes table with a new foreign key allocation_id. It will be set to a ID of a corresponding allocation. Unlike instance_uuid, it will only be set when an allocation is created. If allocation_id is not empty, instance_uuid will hold the UUID of the corresponding allocation (the opposite is not necessary true).

State Machine Impact

No impact on the node state machine.

This RFE introduces a simple state machine for Allocation objects, consisting of three states:

  • allocating allocation is in progress (initial state).

  • active allocation active.

  • error allocation failed.

In the initial version only the following paths are possible:

  • from allocating to active on success.

  • from allocating to error on failure.

In the future we may allow moving from error back to allocating to retry the allocation.

REST API impact

  • POST /v1/allocations create an allocation.

    The API accepts a JSON object. The following field is mandatory:

    • resource_class requested node’s resource class.

    The following fields are accepted:

    • uuid to create an allocation with the specific UUID.

    • candidate_nodes to limit the query to one of these nodes.

      Note

      This value is converted from names to UUIDs internally.

    • traits list of requested traits.

    • name allocation name.

    The normal response is HTTP CREATED with the response body being the created allocation representation. An allocation is created in the allocating state.

    Error codes:

    • 400 Bad Request if

      • any node from candidate_nodes cannot be found (by name or UUID).

      • the resource_class value is invalid.

      • traits is provided and is not a list of valid trait strings.

      • name is not an accepted name.

    • 406 Conflict if

      • the provided uuid is not unique or matches instance_uuid of any node.

      • the provided name is not unique.

  • GET /v1/allocations list allocations.

    Parameters:

    • fields list of fields to retrieve for each allocation.

    • state filter allocations by the state.

    • resource_class filter allocations by resource class.

    • node filter allocations by node UUID or name.

    Error codes:

    • 400 Bad Request if

      • state is invalid.

      • resource_class is invalid.

      • node does not exist.

      • any of the requested fields is invalid.

  • GET /v1/allocations/<uuid or name> retrieve an allocation.

    Parameters:

    • fields list of fields to retrieve.

    Error codes:

    • 400 Bad Request if any of the requested fields is invalid.

    • 404 Not Found if the allocation is not found.

  • DELETE /v1/allocations/<uuid or name> remove the allocation and release the node.

    No request or response body. Response code is HTTP 204 No Content.

    Error codes:

    • 404 Not Found if the allocation is not found.

    • 409 Conflict if the corresponding node is active or is in a state where updates are not allowed.

    Note

    This request will succeed only for real allocations. It will not be possible to unset instance_uuid created without an allocation (i.e. by direct PATCH to a node) using this API.

  • GET /v1/nodes/<node UUID or name>/allocation get allocation associated with the node.

    Parameters:

    • fields list of fields to retrieve.

    The response body is the Allocation object representation.

    Error codes:

    • 404 Not Found if

      • the node cannot be found.

      • there is no allocation for the node.

    • 400 Bad Request if

      • node has instance_uuid that does not correspond to any allocation.

      • any of the requested fields is invalid.

  • Update GET /v1/nodes, GET /v1/nodes/detail and GET /v1/nodes/<node UUID or name:

    Expose the new allocation_uuid field (converted from the node’s allocation_id).

  • Update PATCH /v1/nodes/<node UUID or name>:

    If instance_uuid is unset and the current value corresponds to an allocation:

    • if node is active or in a state that disallows updates, and maintenance mode is off, return HTTP 409 Conflict,

    • otherwise delete the allocation.

    If instance_uuid is set, do NOT create an allocation, keep the previous behavior.

    Note

    This is needed to avoid affecting the nova virt driver. This decision may be revisited in future API versions.

    The allocation_uuid field is read-only, an attempt to change it directly will result in HTTP 400 (Bad Request).

  • Update DELETE /v1/nodes/<node UUID or name>:

    If a node is deleted with allocation (possible only in maintenance mode), the allocation is deleted as well.

Client (CLI) impact

“ironic” CLI

None.

“openstack baremetal” CLI

The matching commands will be created:

openstack baremetal allocation create --resource-class <class> \
   [--trait <trait>] [--trait <trait>] [--uuid <uuid>] [--name <name>]
openstack baremetal allocation list [--state <state>] [--fields <fields>]
   [--resource-class <class>] [--node <UUID or name>]
openstack baremetal allocation get <uuid or name> [--fields <fields>]
openstack baremetal allocation delete <uuid or name>

The allocation_uuid field will be exposed.

RPC API impact

Two new RPC calls are introduced:

  • Creating an allocation

    def create_allocation(self, context, allocation):
       """Create an allocation.
    
       Creates the provided allocation in the database, then starts a thread
       to process it.
    
       :param context: context
       :param allocation: allocation object.
       """
    
  • Deleting an allocation

    def destroy_allocation(self, context, allocation):
       """Destroy an allocation.
    
       Removes the allocation from the database and releases the node.
    
       :param context: context
       :param allocation: allocation object.
       """
    

metalsmith impact

The metalsmith project implements a superset of the proposed feature on a client side. After this API is introduced, metalsmith will switch the reserve_node call to using it in the following way:

  • If the request contains a resource_class and, optionally, traits and candidate nodes, the new API will be used.

  • If the request contains anything not supported by the new API, metalsmith will continue client-side node filtering, and will create an allocation with a list of suitable nodes.

Driver API impact

None

Nova driver impact

None

In the future we may use the allocation API in the nova driver, but there are no plans for it now. Currently going through the allocation API will result in an attempt of double allocation in Placement if Placement is used as an allocation backend.

Ramdisk impact

None

Security impact

None

Other end user impact

None

Scalability impact

None

Performance Impact

A new periodic task will run on each conductor to periodically check for stalled reservations belonging to dead conductors. The default period will be 60 seconds. It will be possible to disable it, in which case the allocations may get stuck forever if their assigned conductor dies.

Other deployer impact

None

Developer impact

None

Implementation

Assignee(s)

Primary assignee:

dtantsur

Work Items

  • Add new tables and the Allocation RPC object.

  • Add RPC for allocating/deallocating.

  • Add API for allocations creation and deletion, and API reference.

  • Update conductor starting procedure to check for unfinished allocations.

  • Add a periodic task to check for orphaned unfinished allocations.

Dependencies

None

Testing

  • Unit tests and Tempest API will be provided.

  • The standalone integration tests will be updated to use the new API.

  • We can add support for the new API to bifrost (e.g. via metalsmith), and test it in a bifrost CI job.

Upgrades and Backwards Compatibility

This change is fully backward compatible. Code using instance_uuid for allocations is not affected.

Documentation Impact

API reference will be provided.

References