Provider Validation for L3 Plugins¶
Currently, validation of end user selections is tightly bound to the reference implementation. As a result, alternate providers cannot validate end user configuration selections, until after they have already been persisted.
When an L3 service request is received by Neutron, it invokes the corresponding service plugin, which validates and persists info for the request (typically in one step). Finally, the plugin will invoke the service driver to apply the changes.
It is at the apply stage, when the provider’s service driver has an opportunity to take action on the request. If the provider has additional constraints, it can reject the request, but the resource(s) will already have been persisted.
For example, when creating a VPN IPsec site connection using an IKE policy that selected v2, the connection will pass validation, be persisted, and the service driver called. With the Cisco CSR, which currently does not support v2 via its REST API, the service driver can raise an exception during the apply. However, the connection will have already been created and persisted.
In addition, if a service driver has persistence to perform, it must do that during the apply phase.
This proposal is to break out the validation into a separate function, and provide a way for provider service drivers to extend or override the validation logic, if desired.
Research into the current validation shows that it must be done during the persistence step, to handle the situation where multiple clients may be making changes to the same resources simultaneously. The transaction will enforce consistency.
The approach planned is to create default validation logic in the service driver abstract base class, and during persistance, select the active service driver and call it’s validation method.
The service driver can perform validation before, after, or instead of the default logic, or not provide a validation method and the default validation in the ABC will be performed.
Goal is to roll out the same proposal for all L3 service plugins.
This will be a structural change, and not a behavorial change.
As part of the Juno summit there was a discussion of using the TaskFlow package for managing the processing for validate, persistence, and apply operations.
More research will be needed on TaskFlow, but regardless, the validation logic should be separated from the persistence logic and can be done as preliminary work.
An alternative is to mark the resource status/state as ERROR, but the database will still be inconsistent with the actual state of the requested operation.
Data model impact¶
No changes to the database schema or operations is expected, other that it would now accurately reflect the actual state of the resources.
REST API impact¶
No expected changes to REST APIs.
Other end user impact¶
There is no significant performance impact expected.
Other deployer impact¶
Providers will want to update their service drivers to take advantage of the new ability to validate and reject requests, based on provider capabilities.
- Primary assignee:
- Other contributors:
Given the scope and number of L3 services, additional contributers would be greatly appreciated!
For each L3 service create and update API:
- Refactor the database functionality * Pull out validation logic and move to service driver ABC * Add hook to call validation logic now in service driver
- Add validation method to provider’s service driver, if needed
- Update unit tests based on structural changes
Note: The separation of the validation logic can be done on each L3 service independently (there are no dependencies), and as needed (e.g. only on resources where multiple implementations have different validation needs), if desired.
Since this is a structural change, it is expected that no additional tempest tests will be needed.
Unit tests will need modification, mostly due to method name changes, but may also change due to better isolation of the logic. Tests that were checking both validation and persistence (or were doing persistence to check validation), can be split out into more isolated test cases. This should reduce the amount of overlap in test cases (possibly eliminating some tests), better isolation, and possibly better performance (not having to do several tasks to test some functionality).
None at this time.