Remove dependence on flavor table and flavor cache

[knikolla]

The information related to which GPU is in use by an instance is available in the pci_requests column of the instance_extra table.

For example:

[
  {
    "count": 1,
    "spec": [
      {
        "product_id": "<REDACTED>",
        "vendor_id": "<REDACTED>",
        "dev_type": "type-PCI"
      }
    ],
    "alias_name": "V100",
    "is_new": false,
    "numa_policy": "legacy",
    "request_id": null,
    "requester_id": null
  }
]

• pci_info[0]["alias_name"] from instance_extra.pci_requests contains the GPU name if the instance has one.
• pci_info[0]["count"] contains the GPU count.
• instances.memory_mb contains the RAM in MB.
• instances.vcpus contains the number of CVPUS.

If no PCI information the instance is assumed as CPU SU.

All flavors are 20GB, so that is hardcoded for the moment, and root disk storage is not to be taken into account for calculations of invoicing.

It is therefore unnecessary to fetch any flavor information and invoicing can be performed directly on the instance information.

I have compared the values of January's billing values before and after the change and they are the same.

With all the information now being fetched from the instance, we don't need to persist flavor information anymore between runs nor fetch flavor information from the nova_api database.

• Removed connection to nova_api
• Removed getting flavor listing from nova_api.flavors
• Removed command line argument to flavor cache
• Removed flavor cache and etc/flavor_cache.json

[knikolla]

Just to make sure I understand: effectively with this change we're no longer billing based on flavors; we're billing on the cpu/memory/etc requests made by each specific instance (which may or may not correspond to an existing flavor).

Assuming I'm correct, that seems reasonable.

Correct. Flavors were an unreliable source of that information as they got deleted often leaving no traces behind as to what they might originally have been - this was especially tricky because GPU SUs were detected based on flavor name. The instance information is a much more reliable source of that same information.