Related-Bug: #1604916 Change-Id: Ic5daf0d31207b7c123cb8b4228dae624383a48ab
7.0 KiB
Add node resource_class field
https://bugs.launchpad.net/ironic/+bug/1604916
Nova has a plan for scheduling ironic resources the same way as other nova resources, that involves making each ironic node a "resource provider", which has a "resource class". That resource class is referenced by the nova flavor, in short saying that the flavor requires exactly one thing from that resource class. When running the resource tracker, nova must be able to associate each ironic node with a resource class, which operators must be able to specify (because operators are the people creating the flavors and need to know how to associate the flavor). To allow for this, we add a new "resource_class" field to ironic's node object.
Problem description
Currently, nova tracks ironic resources with a (host, node) tuple. This causes a number of problems within nova's internals, and the nova team is trying to get away from this. At the same time, nova wants to schedule ironic instances the same way as other instances are scheduled. This allows ironic to follow along in the ongoing scheduler changes, which will help reduce (or eventually eliminate) scheduling races, and benefit from other scheduler optimizations coming down the road (like making qualitative decisions more performant).
Proposed change
The current proposal in nova is to make each ironic node a "resource provider", which is associated with a "resource class".[0] A nova flavor may declare that it requires some amount of a given resource class; in ironic's case it would require one and only one. The resource provider record for an ironic node will declare that it provides exactly 1 (or 0, if it is in maintenance or similar) of the resource class for that provider record.
This proposal still allows nodes to be scheduled to based on qualitative properties, such as capabilities or affinity rules. The resource class is simply the first "filter" in this case (though it isn't a traditional scheduler filter).
To do so, nova needs to know which resource class the resource provider record needs to be associated with. As operators manage the flavors that will be linked to these classes, we need a way for the operator to specify what class each node is in, so that the flavor is linked correctly. As such, we add a "resource_class" field to the node record in ironic, which nova will use when creating the resource provider record.
As an example, imagine an ironic deployment has the following nodes:
- node-1:
resource_class: ironic-gold
properties:
cpus: 4
memory_mb: 16384
capabilities:
boot_mode: uefi,bios
- node-2:
resource_class: ironic-silver
properties:
cpus: 2
memory_mb: 8192
The operator might define the flavors as such:
- baremetal-gold
resources:
ironic-gold: 1
extra_specs:
capabilities: boot_mode:bios
- baremetal-gold-uefi
resources:
ironic-gold: 1
extra_specs:
capabilities: boot_mode:uefi
- baremetal-silver
resources:
ironic-silver: 1
Note that the flavor definition is a strawman and may not be the actual keys used when this is implemented.
A nova user booting an instance with either the baremetal-gold or baremetal-gold-uefi flavor would land on node-1, because capabilities can still be passed down to ironic, and the resource_class on the node matches what is required by flavor. The baremetal-silver flavor would match node-2.
Alternatives
Keep doing the (host, node) thing long enough such that the nova team decides they want to remove our driver from nova, and subsequently remove the (host, node) tuple and break our driver horribly.
Data model impact
Adds a "resource_class" field to the nodes table. This will be a VARCHAR(80) because that matches nova's table structure. There will be data migrations provided for this change. The objects API will also take this change, with a corresponding version bump. This will default to NULL.
State Machine Impact
None.
REST API impact
The new "resource_class" field will be exposed in the API, just like any other string field, with the same semantics.
Default policy for this field should be the same as driver, network_interface, etc.
Filtering and sorting on this field will be added.
There will be a microversion bump, as usual.
Client (CLI) impact
The client will be updated to add the field.
"ironic" CLI
The CLI will be updated to add the field.
"openstack baremetal" CLI
The CLI will be updated to add the field.
RPC API impact
Only the objects changes mentioned above.
Driver API impact
None.
Nova driver impact
Immediately, we'll pass the resource_class field back up to the resource tracker, so that nova can put these resources in the placement database in Newton. There will be a small patch that bumps the API version we're using and passes the field back in the resource_dict. This will need a release note dictating the ironicclient version and available API version needed to run this code.
During Ocata, work will be done on the scheduler to use this for scheduling, however that is outside of the ironic driver.
Ramdisk impact
None.
Security impact
None.
Other end user impact
None.
Scalability impact
None.
Performance Impact
None.
Other deployer impact
After the deployment of the Newton version of nova, deployers will need to populate the resource_class field in ironic, and associate the flavors, before deploying the Ocata version of nova.
Developer impact
None.
Implementation
Assignee(s)
jroll
Work Items
- Add the field to the DB.
- Add the field to the objects model.
- Add the field to the API, with filtering and sorting.
- Doc updates for install guide and such.
Dependencies
None.
Testing
Unit tests should suffice here.
Upgrades and Backwards Compatibility
No direct impact, but it's important to note that scheduling in the Ocata version of nova will still fall back to the old method, if resource_class is set to NULL for any node. Operators should have up until the P version of nova to populate this data.
Documentation Impact
Add the new field to the API reference.
We'll also need to update the install guide and any other docs that talk about setting up nova with ironic, to make sure that deployers are setting this field when adding nodes. This will also need to be communicated extremely hard via release notes (and probably ops list emails).