Strict Two-Level Limits Enforcement Model

This specification describes the behaviors and use-cases of a strict enforcement model for limits associated to resources in a hierarchical project structure.

bp strict-two-level-model

Problem Description

The unified limit specification and implementation, introduced in the Queens release, ignore all details about project structure. It’s only enforcement model is flat. This means limits associated to any part of the tree are not validated against each other.

Proposed Change

This specification goes through the details for a strict two-level hierarchical enforcement model.

Use Cases

• As an operator, I want to be able to set the limit for a top-level project and ensure its usage never exceeds that limit, resulting in strict usage

• As a user responsible for managing limits across projects, I want to be able to set limits across child projects in a way that is flexible enough to allow resources to flow between projects under a top-level project

These use cases were mentioned on the mailing list in an early discussion about unified limits.

Model Behaviors

This model:

• requires project hierarchy never exceeds a depth of two, meaning hierarchies are limited to parent and child relationships

• requires each tree have a single parent, or tree root

• allows parents, or tree roots, to have any number of children

• allows quota overcommit, i.e. the aggregate quota limit (not usage) may exceed the limit of the parent. Overcommit and user-experience related to overcommit is a leading factor for the strict two-level hierarchy.

• does not directly solve sharing data across endpoints, e.g. each nova would not be aware of the other nova’s quota consumption meaning a user could consume the full amount of quota on each endpoint.

This model implements limit validation in keystone that:

• allows the sum of all child limits to exceed the limit of the parent, or tree root

• disallows a child limit from exceeding the parent limit

• assumes registered limits as the default for projects that are not given a project-specific override

This model is consumed by oslo.limit in a way that:

• requires services responsible for resources to implement a usage callback for oslo.limit to use to calculate usage for the project tree

• requires that usage be calculated on every request

The oslo.limit library will enforce the model such that the resource usage sum across the entire tree cannot exceed the resource limit set by the parent.

This model is called a strict-two-level enforcement model. It is strict because the usage of a resource across the entire tree can never exceed the parent limit. It is considered a two-level model because it only assumes to work on project hierarchies of two or less.

Enforcement Diagrams

The following diagrams illustrate the above behaviors, using projects named A, B, C, and D. Assume the resource in question is cores, and the default registered limit for cores is 10. The labels in the diagrams below use shorthand notation for limit and usage as l and u, respectively.

Technically, both B and C can use up to 8 cores each, resulting in:

If A attempts to claim two more cores, the usage check will fail because oslo.limit will fetch the hierarchy from keystone and check the usage of each project in the hierarchy by using the callback provided by the service to see that total usage of A, B, and C is equal to the limit of the tree, set by A.limit.

Despite the usage of the tree being equal to the limit, we can still add children to the tree:

Even though the project can be created, the current usage of cores across the tree prevents D from claiming any cores:

Creating a grandchild of project A is forbidden because it violates the two-level hierarchy constraint. This is a fundamental contraint of this design because it provides a very clear escalation path. When a request fails because the tree limit has been exceeded, a user has all the information they need to provide meaningful context in a support ticket (e.g. their project ID and the parent project ID). An administrator of project A should be able to reshuffle usage accordingly. A system administrator should be able to as well. Providing this information in tree structures with more than a depth of two is much harder, but may be implemented with a separate model.

Granting B the ability to claim more cores can be done by giving B a project-specific override for cores:

Note that regardless of this update, any subsequent requests to claim more cores in the tree will be forbidden since the usage is equal to the limit of the A. If cores are released from A and C, B can claim them:

While C is still under its default allocation of 10 cores, it won’t be able to claim any more cores because the total usage of the tree is equal to the limit of A, thus preventing C from reclaiming the cores it had:

Creating or updating a project with a limit that exceeds the limit of A is forbidden. Even though it is possible for the sum of all limits under A to exceed the limit of A, the total usage is capped at A.limit. Allowing children to have explicit overrides greater than the limit of the parent would result in strange user experience and be misleading since the total usage of the tree would be capped at the limit of the parent:

Finally, let’s still assume the default registered limit for cores is 10, but we’re going to create project A with a limit of 6.

When we create project B, which is a child of project A, the limit API should ensure that project B doesn’t assume the default of 10. Instead, we should obey the parent’s limit since no single child limit should exceed the limit of the parent:

This behavior should be consistent regardless of the number of children added under project A.

Creating limit overrides while creating projects seems counter-productive given the whole purpose of a registered default, but it also seems unlikely to throttle a parent project by specifying it’s default to be less than a registered default. This behavior maintains consistency with the requirement that the sum of all child limits may exceed the parent limit, but the limit of any one child may not.

Proposed Server Changes

Keystone will need to encapsulate this logic into a new enforcement model. Ideally, this enforcement model can be called from within the unified limit API to validate limits before writing them to the backend.

If keystone is configured to use the strict-two-level enforcement model and current project structure within keystone violates the two-level project constraint, keystone should fail to start. On start-up, keystone will scan the database to ensure that all projects don’t exceed two levels of hierarchy and that keystone.conf [DEFAULT] max_project_tree_depth = 2. If either condition fails, keystone will report an appropriate error message and refuse to start.

To aid operators, we can develop a keystone-manage command, to check the hierarchical structure of the projects in the deployment and warn operators if keystone is going to fail to start. This gives operators the ability to check and fix their project hierarchy before they deploy keystone with the new model. This clearly communicates a set project structure to operators at run time.

Proposed Library Changes & Consumption

The oslo.limit library is going to have to know when to enforce usage based on the strict-two-level model. It can ask for the current model by querying the limit API directly:

Request: GET /v3/limits/model

Response

• 200 - OK

• 401 - OK

Response Body

{
    "model": {
        "name": "strict-two-level",
        "description": "Strict usage enforcement for parent/child relationships."
     }
}

The library should expose an object for claims and a context manager so that consuming services can make the following call from within their API business logic:

from oslo_limit import limit
LIMIT_ENFORCER = limit.Enforcer()

 def create_foobar(self, context, foobar):

     claim = limit.ProjectClaim('foobars', context.project_id, quantity=1)
     callback = self.get_resource_usage_for_project
     with limit.Enforcer(claim, callback=callback):
         driver.create_foobar(foobar)

In the above code example, the service builds a ProjectClaim object that describes the resource being consumed and the project. The claim is then passed to an oslo.limit context manager and supplimented with a callback method from the service. The service’s callback method is responsible for calculating resource usage per project. The oslo.limit library can use the project_id from the context object to get the limit information from keystone and calculate usage across the project tree with the callback. The usage check for the project hierarchy will be executed when the context manager is instantiated or executing __enter__. By default, exiting the context manager will verify that the usage was not exceeded by another request, protecting from race conditions across requests. This can be disabled explicity using the following:

from oslo_limit import limit
LIMIT_ENFORCER = limit.Enforcer()

 def create_foobar(self, context, foobar):

     claim = limit.ProjectClaim('foobars', context.project_id, quantity=1)
     callback = self.get_resource_usage_for_project
     with limit.Enforcer(claim, callback=callback, verify=False):
         driver.create_foobar(foobar)

Fetching project hierarchy

The (current) default policy prevents users with a member role on a project from retrieving the entire project hierarchy. The library that needs the hierarchy to calculate usage must call the API as a project administrator or use a service user token. This API is used for both operators and oslo.limit.

Request: GET /limits?show_hierarchy=true

Request filter

• show_hierachy - Whether to show the hierarchy project limit or not.

Response:

A list of the hierarchy project limits.

Response Code:

• 200 - OK

• 404 - Not Found

Response Body:

{
    "limits":[
        {
            "id": "c1403b468a9443dcabf7a388234f3f68",
            "service_id": "e02156d4fa704d02ac11de4ddba81044",
            "region_id": null,
            "resource_name": "ram_mb",
            "resource_limit": 20480,
            "project_id": "fba8184f0b8a454da28a80f54d80b869",
            "limits": [
                {
                    "id": "7842e3ff904b48d89191e9b37c2d29af",
                    "project_id": "f7120b7c7efb4c2c8859441eafaa0c0f",
                    "region_id": null,
                    "resource_limit": 10240,
                    "resource_name": "ram_mb",
                    "service_id": "e02156d4fa704d02ac11de4ddba81044"
                },
                {
                    "id": "d2a6ebbc5b0141178c07951a10ff547c",
                    "project_id": "443aed1062884dd38cd3893089c3f109",
                    "region_id": null,
                    "resource_limit": 5120,
                    "resource_name": "ram_mb",
                    "service_id": "e02156d4fa704d02ac11de4ddba81044"
                },
                {
                    "id": "f8b7f4da96854c4cafe3d985acc5110f",
                    "project_id": "ca7e4b4cd7b849febb34f6cc137089d0",
                    "region_id": null,
                    "resource_limit": 2560,
                    "resource_name": "ram_mb",
                    "service_id": "e02156d4fa704d02ac11de4ddba81044"
                }
            ]
        }
    ]
}

The above is an example response given the following diagram, where the default registered limit for ram_mb is 2560, which applies to D.

Alternatives

Stick with a flat enforcement model, requiring operators to manually implement hierarchical limit knowledge.

Security Impact

None

Notifications Impact

None

Other End User Impact

None

Performance Impact

Usage Caching

Performance of this model is expected to be sub-optimal in comparison to flat enforcement. The main factor contributing to expected performance loss is the calculation of usage for the tree. The oslo.limit library will need to calculate the usage for every project in the tree in order to provide an answer to the service regarding the request.

One possible solution to mitigate performance concerns here would be to calculate usage for projects in parallel.

Limit Caching

Other services will be required to make additional calls to keystone to retrieve limit information in order to do quota enforcement. This will add some overhead to the overall performance of the API call because it requires a round-trip to keystone.

It is also worth noting that both Registered Limits and Project Limits are not expected to change frequently. This means that the limit data could possibly be cached client-side in oslo.limit. However, this introduces concerns about limit invalidation. Consider the following example:

• client-side cache TTL is set to one hour for limit information from keystone

• one minute later, an administrator decreases the limit for cores on a particular project

• two minutes later, a user makes a request to create an instance in the project that should bring them over the limit just set by the administrator

• due to client-side caching, the service considers the project within it’s limit and allows the instance to be created

• the current usage is out of sync with the limit set by the administrator and the service won’t realize this until the TTL expires in another 57 minutes

Client-side caching is going to be a very specific case that needs to be handled carefully because cache invalidation strategies are going to be distributed across services. One possible mitigation would be for client-side caching and keystone to share the same cache instance, making it easier to perform cache invalidation. Conversely, this raises the operational bar for administrators and requires assumptions about underlying infrastructure.

Until we know we can make those types of assumptions or find an alternative solution for cache invalidation, client-side caching should be avoided to prevent situations like what was described above. We should error on the side of accuracy when retrieving limit information.

Other Deployer Impact

None

Developer Impact

The enforcement library oslo.limit should be implemented based on the enforcement model implemented in keystone.

The consuming component (e.g. nova, neutron, cinder, etc..)should add the new way to fetching quota limit from keystone in the future.

Implementation

Assignee(s)

Primary assignee:

• wangxiyuan <wangxiyuan@huawei.com> wxy

• Lance Bragstad <lbragstad@gmail.com> lbragstad

Other contributors:

Work Items

• Add the new API GET /limits/model

• Abstract limit validation into a model object

• Implement a new limit model for strict-two-level

• Implement strict-two-level enforcement in oslo.limit

• Add the new show_hierachy parameter for limits.

• Add keystone client support for limits.

Future work

Limit and Usage Awareness Across Endpoints

oslo.limit and keystone server can be enhanced to utilize etcd (or other shared data store) to represent limit data and cross-endpoint quota-usage. This falls out of scope for this particular specification. It should be noted that the model should be able to consume the data from whatever store is used, not restricted to a local-only datastore.

Optimized Usage Calculation

During the design of this enforcement model, various parties mentioned performance-related concerns when employing this model for trees with many projects. For example, calculating usage for cores across hundreds or thousands of projects. Consider the following tree structure:

Consider that each project not only has the concept of usage and limit, but also something called an aggregate. An aggregate is the sum of a projects usage and all aggregrate counts of its children.

For example, when claiming two cores on C, C.usage=2 and C.aggregate=2. The tree root, A, is also updated in this case where A.aggregate=2. When a subsequent claim is made on B updating its usage to B.usage=2, the usage calculation only needs to check the aggregate usage property of the parent project, or the project tree.

This simplifies the usage calculation process by only having to query the parent, or tree root, for it’s aggregate usage. As opposed to querying each project for it’s usage and sum the result of each aggregate stored for the parent.

The following illustrates a more extreme example:

Let’s assume each project has usage=0 and limit=10. The following might be a possible scenario: When claiming resources on D.usage=4

• SET D.usage=4 AND D.aggregate=4

• SET B.aggregate=4, since B.usage=0 currently

• SET A.aggregate=4, since A.usage=0 currently

• SET C.usage=6 AND C.aggregate=6

• SET A.aggregate=10, since A.usage=0 still

• SET E.usage=2 fails

The last step in the flow fails because the entire tree is already at limit capacity for cores once C finalizes its claim. The advantage is that we only need to ask for A.aggregate in order to calculate usage when E attempts to make its claim, since finalized claims “bubble up” the tree.

Note that this requires services to track usage and aggregate usage, raising the bar for adoption if this is a desired path forward.

Dependencies

• Requires API to expose configured limit model (see bug 1765193)

• Abstract model into it’s own area of keystone to keep limit CRUD isolated from enforcement model

Documentation Impact

• The supported limit model and the new enforcement step must be documented.

References

• Queens Unified Limit specification

• High-level description of unified limits

• Rocky PTG design session etherpad

• Early review containing model context

• Adam’s blog about quota