Reliable quota enforcement¶
Enforce network resource quotas in a reliable and efficient way, without changing the quota API interface.
Quotas enforcement is a relatively simple task. When a request for creating one or more resources is received, the available resource quota is checked, and if the request cannot be satisfied an error is returned to the user. This works perfectly under the assumption that API requests are served serially.
Unfortunately, this assumption is not true when multiple REST API workers are employed. The same applies also if there is a single API worker, but multiple API servers are operating behind a load balancer.
In these conditions, two or more concurrent requests can pass quota enforcement even if there is not enough quota available to accept all of them.
For instance, assume there are 2 REST API workers and the current quota usage is Nmax - 1 where Nmax is the resource quota. Two requests are then sent and dispatched to the workers. Each worker enforces the quota before the other worker completes processing the API request. Both requests are accepted, leading to a overall resource usage of Nmax + 1
Therefore, leveraging multiple workers and bulk creates, a tenant could in theory create up to Nmax * Nworkers resources. While this is unlikely to happen, there is still a good chance that tenants could manage to use more resources then those they’ve been assigned by the cloud provider. This situation is far from ideal both for the cloud provider and the user, as the former might end up in a situation where data center resources are oversubscribed, whereas the latter could be billed more than expected.
For RPC API workers instead, quotas are not enforced at all. This implies that quotas are not enforced for resources created via RPC, usually “internal” ports. While not enforcing quotas for service ports such as DHCP ports might be a desired thing, this is probably not the correct behavior. Indeed these ports will still contribute to the overall usage upon enforcement for REST API requests.
Add a concept of ‘resource reservation’, which is already adopted in other projects such as nova and cinder. Whenever a quota enforcement check is successful, mark the requested amount of resources as ‘reserved’ immediately.
The reservation shall be removed as soon as the corresponding API operation is completed, successfully or not. This will be ideally done upon return from the plugin call. In case of server crashes, an expiration time should be provided to ensure stale reservations do not prevent allocation of resources upon server restarts.
Multiple reservations can exist at anytime for the same tenant and the same resource. The quota check should add all the existing reservations to the current usage count.
Even with reservations, there still is a race condition which can occur if a worker performs the quota check before the others specify their reservations. This can be forced using lock-free algorithms which retry an operation if there are not the conditions to perform it. DB integrity constraints can be used to this aim, without recurring to locking queries or other distributed synchronization constructs.
This can be achieved by introducing a table with the only purpose of keeping track of reservations in progress.
The resource reservation process will therefore look as it follows:
1) ACQUIRE RESERVATION FACILITY Write down in an appropriate DB table, say reservation_locks, that a given worker is attempting to reserve a certain amount of a given resource for some tenant.
—————————————————* | RESOURCE_TYPE | TENANT_ID | LOCKED_BY | EXPIRATION | PK: RESOURCE_TYPE, —————|———–|———–————* TENANT_ID | network | xxx | worker_0 | some time | —————————————————*
The selected primary key ensures that it won’t be possible for two distinct workers to attempt to concurrently reserve the same resource. The expiration time stamp ensures stale locks are ignored in case of server failures.
2) COMMIT TRANSACTION. This will trigger an integrity violation error if another worker attempts to make another reservation for the same tenant and resource type.
In this case, step 1 will be tried again for a few times. There will be an exponential backoff time between attempts.
- DO RESERVATION
Record the reservation in an appropriate table
——————————————–—————-* | RESOURCE_TYPE | TENANT_ID | BOOKING_AMOUNT | EXPIRATION | —————|———–|—————-—————-* | network | xxx | 2 | at some point | ——————————————–—————-*
- RELEASE RESERVATION FACILITY
- COMMIT TRANSACTION
worker 0 worker 1 ———- ———- acquire_resv_facility acquire_resv_facility COMMIT - SUCCESS COMMIT - FAIL (INTEGRITY VIOLATION) do_reservation RETRY release_resv_facility COMMIT - FAIL (INTEGRITY VIOLATION) COMMIT - SUCCESS RETRY - do_reservation - COMMIT - SUCCESS - release_resv_facility - COMMIT - SUCCESS
The table for acquiring the reservation facility acts as a centralized lock, leveraging only primary keys, which are correctly enforced even in active/active clusters such as MySql Galera. The algorithm should however be regarded as non-blocking since workers will always actively retry to perform the reservation; furthermore the backoff mechanism should prevent starvation.
It is also worth noting that the proposed expiration timestamps will prevent stale records from preventing acquiring the reservation facility or create ‘ghost’ resource reservation. The expiration timeout might be configurable, but a default of at least two minutes is advisable considering that the current neutron implementation still suffers of DB deadlocks triggered by eventlet, which with the default MySql setting block threads for about 50 seconds.
Data Model Impact¶
This document proposes two new tables:
Table name: bookings Attributes: resource_type string tenant_id string booking_amount integer expiration datetime
Table name: booking_locks resource_type string tenant_id string locked_by string expiration datetime
REST API Impact¶
No impact on the API interface.
The proposed change should not open any vulnerability which could lead to DoS, leaking tenant information, or allowing attackers to plug into other tenant’s networks.
All the DB queries performed in the implementation of this specification will be scoped by tenant, and built in a way to prevent SQL injection.
Other End User Impact¶
Some additional DB operations will be performed upon quota enforcement.
This change might also serialize some operations which were previously processed in parallel by distinct workers.
While the overall performance impact is expected to be negligible, it will be important to evaluate it upon code review.
There is also an interesting question pertaining resource usage. It is indeed worth exploring whether it is better to count it every time or updating their usage counter whenever resources are created or deleted. To this aim, the cost of SELECT queries vs the cost of adding db hooks on resource create/delete and performing the corresponding UPDATE query should be carefully compared. The gain/loss in terms of performance depends on relative frequency of GET operations vs POST/DELETE operations. Update is definitely more expensive, but SELECT are way more frequent. However, it should be possible to implement this specification leaving resource usage calculation unchanged, and then perhaps come back to it in the future.
Other Deployer Impact¶
As service ports are not anymore counted in resource usage, deployers might expect in theory an increase in port usage. This increase should however be contained if one considers ports are typically used with instances, and the enforcement criteria for instances are not being altered by this specification.
The implementation for this specification will come with adequate developer documentation.
None, I don’t think so... But one can never know.
One alternative worth considering is to us a single table for managing bookings. While the tuple (resource_type, tenant_id, resource_amount) cannot be reliably used as a primary key for serializing bookings among workers, the “locked_by” attribute can be added to this table and the following tuple can constitute the primary key:
(resource_type, tenant_id, locked_by)
When committing the booking the locked_by value would be erased and this will allow other workers to make their booking.
Nevertheless, this solution does not appear to bring benefits in terms of performance and scalability, and also makes the code less readable.
Even if it would be possible to use locking queries (e.g.: SELECT .. FOR UPDATE) statements, this solution has already proved not ideal, and will also miserably fail with some DB backends if active/active replication is employed.
Moreover, an alternative non-blocking algorithms could have been constructed along the lines of the one proposed for Nova . That algorithm has the advantage of performing on average 1 DB transaction instead of 2 for each quota enforcement operation, so it is therefore more efficient. On the other hand, the cost of a retry operation in case of conflict would be slightly higher. While the latter details is not really important, this alternative algorithm cannot be applied to Neutron as it does not have resource usage counters; without them the implementation of this alternative lock free algorithm would be rather difficult.
Another alternative would consist in a separate quota granting authority. This is a possibility, and that is what the ‘Blazar’  resource reservation project advocates for, but might result in an overkill for many OpenStack deployments. Moreover, while in theory it is possible to use Blazar to this aim, the project has actually been conceived to book resources which are meant to be time-shared across tenants. On the other hand the ‘Boson’  project might represent a more viable alternative were quota management and enforcement are delegate to a 3rd application. While this project is very interesting, its not yet in a developments stage such that it can be considered for adoption by Neutron.
Finally, this problem can also be solved by introducing a distributed lock among API workers. memcached or zookeeper could be used with relative ease to implement this sort of distributed coordination. Nevertheless, there is probably no need to resort to distributed coordination if a lock-free algorithm can be devised just leveraging DB integrity.
- Preliminary yak shaving - refactor existing quota module
- Add resource booking logic and use them in quota enforcement
- Remove service ports from quota enforcement
The big dependencies for this change could be the following: 1) removal of self-grown WSGI framework and subsequent switch to pecan 2) review of the plugin interface.
While the above work items pretty define new hooks for performing quota enforcement, they won’t change the logic of the quota enforcement module, which can therefore be implemented orthogonally.
None. This specification does not change the API interface nor any change which might have an impact on the integrated gate.
Appropriate functional tests will be added to validate correct quota enforcement. As a proper verification will require triggering the race condition, some sort of fault injection might be needed. The need and feasibility of this will be evaluated separately.
No further API test is needed.
Document that service ports won’t count anymore in the overall resource usage.
As there is currently no developer documentation for quotas, this is rather easy: “do developer documentation for the quota enforcement module”