Serialization

Serialization

A Cinder driver is stateless on itself, but it still requires the right data to work, and that’s why the cinder-volume service takes care of storing the state in the DB. This means that cinderlib will have to simulate the DB for the drivers, as some operations actually return additional data that needs to be kept and provided in any future operation.

Originally cinderlib stored all the required metadata in RAM, and passed the responsibility of persisting this information to the user of the library.

Library users would create or modify resources using cinderlib, and then would have to serialize the resources and manage the storage of this information. This allowed referencing those resources after exiting the application and in case of a crash.

Now we support Metadata Persistence plugins, but there are still cases were we’ll want to serialize the data:

  • When logging or debugging resources.

  • When using a metadata plugin that stores the data in memory.

  • Over the wire transmission of the connection information to attach a volume on a remote nodattach a volume on a remote node.

We have multiple methods to satisfy these needs, to serialize the data (json, jsons, dump, dumps), to deserialize it (load), and to convert to a user friendly object (to_dict).

To JSON

We can get a JSON representation of any cinderlib object - Backend, Volume, Snapshot, and Connection - using their following properties:

  • json: Returns a JSON representation of the current object information as a Python dictionary. Lazy loadable objects that have not been loaded will not be present in the resulting dictionary.

  • jsons: Returns a string with the JSON representation. It’s the equivalent of converting to a string the dictionary from the json property.

  • dump: Identical to the json property with the exception that it ensures all lazy loadable attributes have been loaded. If an attribute had already been loaded its contents will not be refreshed.

  • dumps: Returns a string with the JSON representation of the fully loaded object. It’s the equivalent of converting to a string the dictionary from the dump property.

Besides these resource specific properties, we also have their equivalent methods at the library level that will operate on all the Backends present in the application.

Attention

On the objects, these are properties (volume.dumps), but on the library, these are methods (cinderlib.dumps()).

Note

We don’t have to worry about circular references, such as a Volume with a Snapshot that has a reference to its source Volume, since cinderlib is prepared to handle them.

To demonstrate the serialization in cinderlib we can look at an easy way to save all the Backends’ resources information from an application that uses cinderlib with the metadata stored in memory:

with open('cinderlib.txt', 'w') as f:
    f.write(cinderlib.dumps())

In a similar way we can also store a single Backend or a single Volume:

vol = lvm.create_volume(size=1)

with open('lvm.txt', 'w') as f:
    f.write(lvm.dumps)

with open('vol.txt', 'w') as f:
    f.write(vol.dumps)

We must remember that dump and dumps triggers loading of properties that are not already loaded. Any lazy loadable property that was already loaded will not be updated. A good way to ensure we are using the latest data is to trigger a refresh on the backends before doing the dump or dumps.

for backend in cinderlib.Backend.backends:
    backend.refresh()

with open('cinderlib.txt', 'w') as f:
    f.write(cinderlib.dumps())

When serializing cinderlib resources we’ll get all the data currently present. This means that when serializing a volume that is attached and has snapshots we’ll get them all serialized.

There are some cases where we don’t want this, such as when implementing a persistence metadata plugin. We should use the to_json and to_jsons methods for such cases, as they will return a simplified serialization of the resource containing only the data from the resource itself.

From JSON

Just like we had the json, jsons, dump, and dumps methods in all the cinderlib objects to serialize data, we also have the load method to deserialize this data back and recreate a cinderlib internal representation from JSON, be it stored in a Python string or a Python dictionary.

The load method is present in Backend, Volume, Snapshot, and Connection classes as well as in the library itself. The resource specific load class method is the exact counterpart of the serialization methods, and it will deserialize the specific resource from the class its being called from.

The library’s load method is capable of loading anything we have serialized. Not only can it load the full list of Backends with their resources, but it can also load individual resources. This makes it the recommended way to deserialize any data in cinderlib. By default, serialization and the metadata storage are disconnected, so loading serialized data will not ensure that the data is present in the persistence storage. We can ensure that deserialized data is present in the persistence storage passing save=True to the loading method.

Considering the files we created in the earlier examples we can easily load our whole configuration with:

# We must have initialized the Backends before reaching this point

with open('cinderlib.txt', 'r') as f:
    data = f.read()
backends = cinderlib.load(data, save=True)

And for a specific backend or an individual volume:

# We must have initialized the Backends before reaching this point

with open('lvm.txt', 'r') as f:
    data = f.read()
lvm = cinderlib.load(data, save=True)

with open('vol.txt', 'r') as f:
    data = f.read()
vol = cinderlib.load(data)

This is the preferred way to deserialize objects, but we could also use the specific object’s load method.

# We must have initialized the Backends before reaching this point

with open('lvm.txt', 'r') as f:
    data = f.read()
lvm = cinderlib.Backend.load(data)

with open('vol.txt', 'r') as f:
    data = f.read()
vol = cinderlib.Volume.load(data)

To dict

Serialization properties and methos presented earlier are meant to store all the data and allow reuse of that data when using drivers of different releases. So it will include all required information to be backward compatible when moving from release N Cinder drivers to release N+1 drivers.

There will be times when we’ll just want to have a nice dictionary representation of a resource, be it to log it, to display it while debugging, or to send it from our controller application to the node where we are going to be doing the attachment. For these specific cases all resources, except the Backend have a to_dict method (not property this time) that will only return the relevant data from the resources.

Backend configuration

When cinderlib serializes any object it also stores the Backend this object belongs to. For security reasons it only stores the identifier of the backend by default, which is the volume_backend_name. Since we are only storing a reference to the Backend, this means that when we are going through the deserialization process the Backend the object belonged to must already be present in cinderlib.

This should be OK for most cinderlib usages, since it’s common practice to store the storage backend connection information (credentials, addresses, etc.) in a different location than the data; but there may be situations (for example while testing) where we’ll want to store everything in the same file, not only the cinderlib representation of all the storage resources but also the Backend configuration required to access the storage array.

To enable the serialization of the whole driver configuration we have to specify output_all_backend_info=True on the cinderlib initialization resulting in a self contained file with all the information required to manage the resources.

This means that with this configuration option we won’t need to configure the Backends prior to loading the serialized JSON data, we can just load the data and cinderlib will automatically setup the Backends.

Creative Commons Attribution 3.0 License

Except where otherwise noted, this document is licensed under Creative Commons Attribution 3.0 License. See all OpenStack Legal Documents.