Dependency-driven multi-step resource deallocation¶
Murano components may allocate various kinds of resources: virtual machines, networks, volumes etc. When these components get removed from the deployment appropriate resources have to be deallocated. Current implementation of this process has some significant limitations and flaws.
This specification aims to address these issues and provide a design for the better resource deallocation / garbage collection system for MuranoPL.
In Murano the deallocation of resources is managed by a garbage collection system (GC). Its present implementation is based on the execution of special methods called .destroy() which may be defined in each MuranoPL class and are intended to contain the custom code to deallocate resources allocated by the objects of those classes. These methods get executed when their objects leave object graph, however the exact order of these executions is currently undefined.
There are two different scenarios when the objects may leave the object graph thus causing their .destroy methods to be called:
- “Offline” changes of the Object Graph, i.e. the changes introduced in the serialized version of the object graph via the API. These changes are detected by comparing the incoming object graph (the one passed from the API for deployment) with the “snapshot” of current environment made after the previous deployment has been completed. If some object exists in the “snapshot” but is missing in the input graph it is considered to be removed. Such objects are deserialized from the snapshot and their .destroy methods are called in the order from deepest nested objects towards topmost ones.
- “Runtime” changes. Some objects may be removed from the object graph during deployment: they may be unreferenced or assigned to Runtime properties or local variable only. As a result after the deployment completes these objects are not serialized neither into the output version of the object graph nor into its “snapshot”. The next deployments will know nothing about their existence so the objects will be lost forever. To recover the resources allocated by such unreferenced objects murano analyzes its ObjectStore after the execution is complete. Each object which is present in the store but is not present in the output version of the object graph is considered to be “orphan” and thus its .destroy method is called. The order of these calls is currently undefined: the objects get destroyed based on their position in the object store, which is hardly predictable.
Such a design is insufficient for production grade applications which often require the following scenarios:
If some object is going to be deleted another object (either owning or just referencing it) may need to execute some actions before or after the object is deleted.
When a group of nested or interconnected objects is about to be deleted the order in which their destructors should be executed may be different in different cases.
Sometimes the actions being executed during the destruction of an object may depend on the fact whether some other object is about to be deleted or not.
Consider an Application which consists of a Network component and several VM components. All the components are owned by the Application but there are no ownership relationships between them. When the Application is going to be deleted (i.e. its whole subgraph leaves the environment) all of its components are about to be removed as well, and so their ``.destroy`` methods will be called. Since Network does not own the VM’s the order of these calls is undefined. Due to various implementation details it may be impossible to remove the Network before the VMs which are connected to it (e.g. in case when the VM has a mandatory requirement to be always connected to at least one network). In this case the ``.destroy`` of a Network component should be always called after all the VMs have been destroyed.
Another issue is that murano uses Objects and ObjectsCopy objects to transfer data between deployments. When destruction dependencies are implemented, the handler will make changes (if any) to objects in ObjectsCopy. Therefore, these changes are not applied during the next deployment and this should be addressed.
Also, sometimes it can be useful to deallocate resources used by the unreferenced objects even before the end of deployment on demand from the application code.
Several improvements have to be made in Murano engine to address the problems described above.
There should be a way to establish directional destruction dependencies between Murano objects. If object Foo establishes such a dependency on the object Bar then:
- Foo will be notified when Bar is about to be destroyed. These notifications are covered in details in “Multi-step destruction” section below.
- If both Foo and Bar are going to be destroyed in the same garbage collection execution, Bar will be destroyed before Foo.
These dependencies are not related to object ownership relationship or property-based cross-references: the owner may have a destruction dependency on its nested object or vise versa; the objects referencing each other may have some destruction dependency established. Even entirely unrelated objects may have a destruction dependency between them.
Since the destruction dependencies are directional there is a theoretical possibility of a circular dependency to exist. In case if two or more objects form such a circle they will still be notified about pending destruction of their dependencies, however the order of this notifications - and the destruction itself - is undefined in this case.
Destruction dependencies are going to be used during all kinds of garbage collection: pre-deployment (“offline”), on-demand (during deployment) and post-deployment.
Instead of just iterating through all the objects going to be destructed and calling their .destroy methods Murano should perform a multi-step garbage collection according to the following algorithm:
Detect all the objects going to be destroyed. It can be done by using Python GC to track and collect objects, as described in the Garbage collection executions section.
Sort the list of detected objects using the following comparator: for any two objects A and B in the list:
IF (A owns B) THEN A>B ELSE IF (A has-a-destruction-dependency-on B) AND (NOT B has-a-destruction-dependency-on A) THEN A>B ELSE A==B
where has-a-destruction-dependency-on means that the left operand object has a destruction dependency (probably transitive) on right operand object, owns means that the left operand object owns (probably transitively) the right operand object.
The objects which are considered to be equal by the algorithm above can be destroyed in parallel.
The result of the sorting is the dictionary with indexes as keys and lists of objects with equal destruction priority as values.
Sort the keys of the dictionary in the reversed order of destruction priority and for each key start parallel notification about scheduled destruction of the objects in the corresponding group. During notification, handlers of the objects that have a destruction dependency on some “sentenced” object will be invoked.
Sort the keys of the dictionary in the direct order of destruction priority and for each index in the dictionary start parallel destruction of all objects in the corresponding group. Destruction of individual object means calling the .destroy() method of the object if present and changing the object’s status to “Destroyed” (see below).
As an environment does not have owners, it will always be in the last group of destruction. There is no guarantee that some other objects (for example, heat stack) will be alive at the time of its destruction. Thus io.murano.Environment class should not have the .destroy() method.
When an object is being processed by a garbage collector, it means that there are no live references to it from the objects of the environment. However there may be cases when the code which handles either the pre-destroy notification (step 4 above) or the actual .destroy method re-establishes the references to the object being destructed, and thus the object remains in the object graph after the GC is completed. Since the resources may be deallocated at this time the regular usage of the object is not possible, however if it is assigned to a property of some another object in the graph it may not always be possible to just nullify that property since it may cause a contract violation.
To resolve such collisions it is proposed to explicitly mark such destroyed objects as “destroyed”. It means setting object’s destroyed attribute to True and removing the self-reference from it. MuranoPL executor will not allow to execute any methods on such objects, however their properties remain accessible (i.e. readable) so any runtime information associated with them may be recovered. Destroyed objects will be serialized with the rest of object graph but the json-representation of the object will have a special flag in their class header (the ”?” section) to indicate their special status. When deserialized from json such objects will retain their “destroyed” status, so the method execution will still be impossible even in subsequent deployments.
When “destroyed” objects are unreferenced from the object graph, their properties get nullified and they get destroyed automatically by Python GC.
Garbage collection executions¶
The multi-step object destruction described above should take place in three different scenarios:
- (currently existing) Before the deployment, destroys objects which were present in the object graph after the previous deployment was finished but were not found in the incoming object graph of a new deployment, i.e. the ones explicitly removed using the API.
- (currently existing) After the deployment, destroys the objects existing in the Object Store but not being a part of the persistable object graph of current environment, i.e. having no references to them from the persistable (In, Out, InOut) properties of the environment or its transitive children).
- (proposed) During the deployment, explicitly initiated from MuranoPL code. Destroys objects which are not part of the complete object graph, i.e. having no references to them from any properties of the environment (including runtime and private properties) AND not being referenced by local variables in any frame of all the green threads of current deployment.
To implement scenario 3, a new algorithm is needed. As mentioned in the Multi-step destruction section, Python garbage collector can be used for that. MuranoPL can make use of the way Python’s gc.collect works.
Python library gc allows running on-demand garbage collection through the gc.collect() method and provides access to unreachable objects that the collector found but cannot free through collection gc.garbage .
To make use of this, there should be an ability to:
- Make object store have weak links only and prevent hard links in other DSL places so that only links between objects remain.
- Prevent murano objects that should have been destroyed by Python GC from being destroyed.
- Get the list of such objects and destroy them in correct order and notify subscribers about destruction.
The prevention can be done by adding __del__() magic method to the MuranoObject class and creating cyclic reference in the object to itself. When gc.collect() is done, all unreachable objects can be examined and murano objects owned by current executor can be distinguished among them.
The difference in Python 3.4 and higher is that objects with a __del__() method don’t end up in gc.garbage anymore and this list should be empty most of the time . So logic of adding object to GC candidates can be added directly to __del__().
It means that in Python versions 3.4 and higher, murano objects will be added to planned destruction from __del__() call caused by gc.collect(), and in versions prior to 3.4, presence of __del__() along with cyclic reference to itself will provide adding the object to gc.garbage list, and it can be added to candidates for destruction from there.
This logic can be used for garbage collection in all three scenarios mentioned above.
The resulting list of GC candidates is then destroyed as described in the Multi-step destruction section above.
With this approach, the comparison of Objects and ObjectsCopy is not needed anymore. Garbage collector works with the same objects on each deployment, so all changes are saved properly.
A new python-backed Murano class called GC should be added to the core library. It should have the following static methods:
collect() - initiates garbage collection of unreferenced objects of current deployment (see p.3 in “Garbage collection executions” section above).
isDestroyed(object) - checks if the object was already destroyed during some GC session and thus its methods cannot be called.
isDoomed(object) - can be used within the .destroy() method to test if another object is also going to be destroyed.
subscribeDestruction(publisher, subscriber, handler=null) - establishes a destruction dependency from the subscriber to the object passed as publisher. Method may be called several times, in this case only a single destruction dependency will be established, however the same amount of calls of unsubscribeDestruction will be required to remove it.
handler argument is optional. If passed it should be the name of an instance method defined by the caller class to handle notification of publisher‘s destruction (see “Multi-step destruction” section above: this handler is executed for p. 3.1)
The following arguments will be passed to the handler method:
- object - a target object which is going to be destroyed. It is not recommended to persist the reference to this object anywhere. This will not prevent the object from being garbage collected but the object will be moved to the “destroyed” state which is almost always bad. The option to do so is considered to be advanced feature which should not be done unless it is absolutely necessary.
unsubscribeDestruction(publisher, subscriber, handler=null) - removes the destruction dependency from the subscriber to the object passed as publisher. Method may be called several times without any side-effects. If subscribeDestruction was called more than once the same (or more) amount of calls to unsubscribeDestruction is needed to remove the dependency.
handler argument is optional and must correspond the handler passed during subscription if it was provided back then.
Application developers may try to implement their own event-based notification logic to notify about pending and completed object destructions. However it will solve only part of the problem: notifications will work properly, but they will not affect the order in which the objects are destroyed, so the workflows will be too complicated. Also this alternative will not have the advanced features proposed in this spec, such as ability to check if some object is going to be destroyed.
Data model impact¶
REST API impact¶
The proposed change is completely backwards compatible: without explicit destruction dependencies objects will be collected based on their ownership relationships, i.e. as it is done in the current implementation.
The packages containing classes which explicitly call the methods of GC should have package format of at least 1.4 to prevent their execution on older versions of Murano which do not have this feature.
Other end user impact¶
Developers will get the new MuranoPL-based API to manage resource deallocation lifecycle. If they do not want to use it they don’t need to do anything.
Murano-dashboard / Horizon impact¶
- Primary assignee:
- Other contributors:
- Stan Lagun <istalker2> starodubcevna
- Implement a system to define and use destruction dependencies in runtime.
- Introduce changes to MuranoObject class to keep track of “destroyed” object status.
- Modify the serializer / deserializer to properly persist the value of the “destroyed object” flag.
- Implement collecting unreferenced murano objects utilizing Python gc library
- Implement sorting algorithms to arrange objects-to-be-destroyed based on criteria defined in p.2 of “Multi-step destruction” section above.
- Implement multi-step destruction workflow.
- Implement GC class to bind all the above.
- Create test-runner-based tests to cover all the test scenarios.
- Document the new features.
The development of this feature will enable Application Development Framework  to address resource deallocation problems during application uninstall.
Tests should be written for test-runner to cover various scenarios of resource deallocation.
Runtime garbage collection¶
There should be test cases covering that:
- objects assigned to persistent (Input, Output, InputOutput) properties (both locally-declared and inherited) of objects reachable from the current roots are NOT garbage collected;
- objects assigned to transient (Runtime and undeclared) properties (both locally-declared and inherited) of objects reachable from the current roots are NOT garbage collected; target properties should be both locally-declared and inherited;
- objects assigned to static properties of various classes are NOT garbage collected;
- objects passed to python-backed objects and unreferenced in MuranoPL are NOT garbage collected unless their MuranoObjectInterface proxies are unreferenced / GC’ed in python;
- objects assigned to local variables of the current execution frame (i.e. variables of the current method and all the caller methods in call stack) including method arguments are NOT garbage collected;
- single unreferenced objects ARE garbage collected;
- graphs of interconnected objects having no references from non-collected objects ARE garbage collected;
- objects passed to python-backed objects and unreferenced in both MuranoPL and python ARE garbage collected;
- garbage collector correctly processes stack-frame objects from green-threads other than the one it is executed from
Destruction dependency resolution order¶
There should be test cases covering that:
- if some child object has a destruction dependency on its parent, the parent gets destroyed before the child;
- if some parent object has a destruction dependency on its child, the child gets destroyed before the parent;
- if some objects not being the part of some ownership hierarchy have some destruction dependency, the dependency-object is destroyed before the dependent one;
- if some objects have circular destruction dependency they are all destroyed (the order is not enforced by the test);
Given the base scenario of object A having a destruction dependency on object B and B being GC’ed, there should be tests covering that:
- the right order of events occurs (B scheduled for destruction -> A is notified about planned B’s destruction -> B’s .destroy() method is called -> B gets destroyed);
- A may prevent B’s destruction by establishing a reference on B in the handler;
- A may establish more than 1 destruction dependency on B and still be notified just once;
- A may remove the destruction dependency and not get notified on B’s destruction;
- If A established N destruction dependencies and then removed them M times, (N>M) then notifications are still delivered;
- If A established N destruction dependencies and then removed them M times, (N<=M) then notifications are not delivered;
- B may establish a destruction dependency on itself thus subscribing to appropriate notifications;
- isDoomed and isDestroyed methods return appropriate values when called by A for B in appropriate event handlers.
Developers documentation should be updated to describe the new GC class and its methods, as well as the design guidelines for application developers to follow to utilize the new capability.