Concurrent resource access in cinder is a problem that has caused resource corruption when simultaneous resources are mutated on by multiple cinder entrypoints (api and manager for example). In Icehouse there has been some addition & usage of locks around manager functions to queue up those requests when a resource is being simultaneous worked on by multiple functions (this stops one of those operations from concurrently mutating the underlying resource). Sadly this is more of a sledgehammer approach and hides the symptoms of the problem and makes it non-obvious when debugging what other requests are queued up behind the lock (or why dead-locking is occurring, if and when it does).
To help alleviate and hopefully solve this problem we will try to attack some of these issues in a different manner, integrating a allowed state transition table into the create_volume workflow and doing strategic state transitions and aborting/erroring out when these state transitions are not allowed. In the future this will help create a concrete set of well defined states and transitions for other workflows as well (and will make it clear while looking at code and during debugging which transitions are allowed at the same time and what transitions are actively occurring).
A high-level description of the problem:
More detailed description:
Since the scope of this problem is bigger (it applies to all/most operations that act on resources) we have to start somewhere so we will start by working through how this will look for the create_volume workflow. It does raise a larger question of how can this change be done in a piecemeal fashion since the other operations will still be lock dependent, and mixing state transitions and lock acquisition techniques will likely not end in a correct solution. We will have to explore how to do this in a way that is piecemeal but also does not destabilize cinder more.
Instead of acquiring local filesystem locks in the manager processes refactor the concept of a lock to instead be a set of allowed and disallowed state transitions (which is in concept similar to the internal mechanism that a lock uses anyway).
Lets take an abbreviated example of how this could work:
When a volume is requested to be created, a database record is created for this volume, in this database record there exists a field called status (in fact there exists multiple of these statuses fields, in the future these should maybe be removed?) that is used to report back to the user data about the status of the create volume request as it moves through the various components in cinder (api, scheduler, and manager).
This status itself has a expected transition diagram and itself is a starting point in determining the larger states transitions that a cinder volume create request goes through (and is allowed to go through). Instead of overriding this status field this proposal proposes to augment the data storage layer in cinder with a new resource_states table. It may be represented by something other than a table depending on where this data is stored (if zookeeper was used it would be represented as a resource tree), the only constraint that we must enforce is that we can atomically fetch and update the given state of a resource in a single atomic operation.
A potential schema could look like the following:
|7c92ee46-7a2e-4183-99c5-909f3d46a90e 7c92ee46-7a2e-4183-99c5-909f3d46a90e 7c92ee46-7a2e-4183-99c5-909f3d46a90e 7c92ee46-7a2e-4183-99c5-909f3d46a90e||CREATING_DB SCHEDULING CREATING_VOL NULL/None||2014-05-22T15 2014-05-23T15 2014-05-23T15 2014-05-24T15|
This table structure will then be used (with NULL states to delimit when a request has fulfilled its set of allowed state transitions) to determine at the API level (before a request has been accepted) what a resource is currently being used for and the API server can then attempt to initiate a transition to a desired state (for example, DETACHING) and depending on if this transition is allowed (by looking at the last known state) it may fail or succeed at performing this transition. If it succeeds it continues with the rest of the workflow for the desired operation (subsequent transitions will also be made in the rest of the workflow, as needed, with the final transition being a transition to NULL/None, to denote that the operation has completed). If the transition is disallowed/fails the API request will be denied and the operation will not be allowed to make forward progress (in the future this model can be relaxed to allow for simultaneous state transitions for operations where this makes sense).
To accomplish this, in the create_volume operation there exists the usage of taskflow, which has helped decompose the workflows that volume creation goes through (it also makes it possible to resume from a prior state if the process crashes). This decomposition makes it obvious (or more obvious) where the transitions should occur and what the transitions are. The proposed path is to add in new nodes into the workflow that will perform & validate these state transitions (attempting to mutate the above resource state table) at a granular-enough level to be useful & meaningful (the transition table also can be useful for operators and developers attempting to determine what is happening inside cinder). When this is combined with notifications from taskflow about its own internal states (via notifications) the ability to decipher what is going on internally to cinder becomes very easy & provides invaluable information to users, developers and operators using & operating cinder.
One possibility for avoiding the above resource_states table is to use a DLM and use a similar approach that is being used with file locks in cinder. The usage of it would be similar to the usage of file locks, although there are scenarios at RPC boundaries where it would still require state transition validation. For example when a lock is released and an async RPC call is made there becomes the possibility for other async RPC calls to also be active at the same time and there would require a state transition and lock system to be used when the receivers of those RPC calls accept and perform the requested RPC operation.
Another possible solution that does not require state transitions is to not use async RPC calls but instead use sync RPC calls, and the sender would only release the DLM lock it owns after it has received confirmation that the receiver has started to process (or accepted the request). The receiver would then acquire the lock during this period when it accepts the request, ensuring that correct lock hand-off happens between the send and receiver. This would require a sensitive and hard to get correct lock hand-off code path & process (this path would need to be tested heavily to ensure correctness).
IMHO both of these alternative methods are too fragile and do not make the state transition process and diagram obvious to developers, operators, and users. This lack of information impedes cinder adoption, and makes it more difficult to recovery from (and understand) inevitable failures and operational issues.
I would also like include a note to what the scope of this specification does not encompass.
These are larger cross-project consistency issues and will need to be solved at a higher level across the projects. It should be noted that once a project itself has a consistent set of states and transitions it becomes much easier to make cross-project consistency possible (without internal consistency cross-project resource usage might as well be discouraged/avoided).
See the above proposed table.
We must be careful to retain the existing API so that nova which is dependent on cinders currently visible states continues to work. This just means that we need to have a exposed mapping that nova is compatible with; while we have an internal mapping which is much more detailed and consistent.
Maybe in the future.
None currently, the state transition information could also be sent out to the notification system if this is desirable in the future to do so.
End users should now expect more errors (or try again later) responses when performing operations concurrently on the same set of resources. Previously some of these operations may or may not have succeeded.
A new table will be created in sqlalchemy and a new model will be created for this new schema. This table will be high read and write traffic (since all operations that occur in cinder will write data to it) so it might be recommended to alter the table type to a more friendly format that performs better for this tables limited usage. Since this table is relatively simple it should also be possible in the future (when correctness is achieved) to switch this table to some other backend that can optimize itself for small read/writes with little history (history is not as useful, except for operators and developers who wish to interrogate what has happened to a resource in the past).
Developers would likely get a lot of the benefit of this information to start since it will help them understand the states a workflow goes through (at the cinder level), combining this with the event stream that taskflow emits creates a lot of useful runtime information that can be used while running cinder or while developing cinder (where to add new state transitions in becomes more obvious when the state transitions that occur are well defined and understood).
J/3 into K (this is likely not a short-term specification).
Since this change affects how cinder operates at a low level, it will require a good amount of testing to verify that concurrent operations are disallowed. Currently tempest may not be the best way to test these concurrent operations since to my knowledge it does not run in parallel (and only when it runs in a controlled parallel process can u find these concurrency issues). So the way to test these concurrency issues needs to be determined (is rally the way to go here, using its concurrent scenarios to probe that this feature works?).
There may be new documentation required to explain why operations that were allowed to occur concurrently are no longer allowed to occur concurrently since this new state transition will be more strict as to what can and what can not occur at the same time.
It will also become possible to start to form documents like taskflow states that show exactly what the internals of cinder are doing and what the allowed state transitions (aka the cinder reference operation states) are.