Introduction to Perspective Broker

Introduction

Suppose you find yourself in control of both ends of the wire: you have two programs that need to talk to each other, and you get to use any protocol you want. If you can think of your problem in terms of objects that need to make method calls on each other, then chances are good that you can use Twisted’s Perspective Broker protocol rather than trying to shoehorn your needs into something like HTTP, or implementing yet another RPC mechanism [1] .

The Perspective Broker system (abbreviated “PB” , spawning numerous sandwich-related puns) is based upon a few central concepts:

  • serialization : taking fairly arbitrary objects and types, turning them into a chunk of bytes, sending them over a wire, then reconstituting them on the other end. By keeping careful track of object ids, the serialized objects can contain references to other objects and the remote copy will still be useful.
  • remote method calls : doing something to a local object and causing a method to get run on a distant one. The local object is called a RemoteReference , and you “do something” by running its .callRemote method.

This document will contain several examples that will (hopefully) appear redundant and verbose once you’ve figured out what’s going on. To begin with, much of the code will just be labelled “magic” : don’t worry about how these parts work yet. It will be explained more fully later.

Object Roadmap

To start with, here are the major classes, interfaces, and functions involved in PB, with links to the file where they are defined (all of which are under twisted/, of course). Don’t worry about understanding what they all do yet: it’s easier to figure them out through their interaction than explaining them one at a time.

Other classes that are involved at some point:

  • RemoteReference : spread/pb.py
  • pb.Root : spread/pb.py , actually defined as twisted.spread.flavors.Root in spread/flavors.py
  • pb.Referenceable : spread/pb.py , actually defined as twisted.spread.flavors.Referenceable in spread/flavors.py

Classes and interfaces that get involved when you start to care about authorization and security:

  • Portal : cred/portal.py
  • IRealm : cred/portal.py
  • IPerspective : spread/pb.py , which you will usually be interacting with via pb.Avatar (a basic implementor of the interface).

Subclassing and Implementing

Technically you can subclass anything you want, but technically you could also write a whole new framework, which would just waste a lot of time. Knowing which classes are useful to subclass or which interfaces to implement is one of the bits of knowledge that’s crucial to using PB (and all of Twisted) successfully. Here are some hints to get started:

  • pb.Root , pb.Referenceable : you’ll subclass these to make remotely-referenceable objects (i.e., objects which you can call methods on remotely) using PB. You don’t need to change any of the existing behavior, just inherit all of it and add the remotely-accessible methods that you want to export.
  • pb.Avatar : You’ll be subclassing this when you get into PB programming with authorization. This is an implementor of IPerspective.
  • ICredentialsChecker : Implement this if you want to authenticate your users against some sort of data store: i.e., an LDAP database, an RDBMS, etc. There are already a few implementations of this for various back-ends in twisted.cred.checkers.

Things you can Call Remotely

At this writing, there are three “flavors” of objects that can be accessed remotely through RemoteReference objects. Each of these flavors has a rule for how the callRemote message is transformed into a local method call on the server. In order to use one of these “flavors” , subclass them and name your published methods with the appropriate prefix.

  • twisted.spread.pb.IPerspective implementors

    This is the first interface we deal with. It is a “perspective” onto your PB application. Perspectives are slightly special because they are usually the first object that a given user can access in your application (after they log on). A user should only receive a reference to their own perspective. PB works hard to verify, as best it can, that any method that can be called on a perspective directly is being called on behalf of the user who is represented by that perspective. (Services with unusual requirements for “on behalf of” , such as simulations with the ability to posses another player’s avatar, are accomplished by providing indirected access to another user’s perspective.)

    Perspectives are not usually serialized as remote references, so do not return an IPerspective-implementor directly.

    The way most people will want to implement IPerspective is by subclassing pb.Avatar. Remotely accessible methods on pb.Avatar instances are named with the perspective_ prefix.

  • twisted.spread.pb.Referenceable

    Referenceable objects are the simplest kind of PB object. You can call methods on them and return them from methods to provide access to other objects’ methods.

    However, when a method is called on a Referenceable, it’s not possible to tell who called it.

    Remotely accessible methods on Referenceables are named with the remote_ prefix.

  • twisted.spread.pb.Viewable

    Viewable objects are remotely referenceable objects which have the additional requirement that it must be possible to tell who is calling them. The argument list to a Viewable’s remote methods is modified in order to include the Perspective representing the calling user.

    Remotely accessible methods on Viewables are named with the view_ prefix.

Things you can Copy Remotely

In addition to returning objects that you can call remote methods on, you can return structured copies of local objects.

There are 2 basic flavors that allow for copying objects remotely. Again, you can use these by subclassing them. In order to specify what state you want to have copied when these are serialized, you can either use the Python default __getstate__ or specialized method calls for that flavor.

  • twisted.spread.pb.Copyable

    This is the simpler kind of object that can be copied. Every time this object is returned from a method or passed as an argument, it is serialized and unserialized.

    Copyable provides a method you can override, getStateToCopyFor(perspective) , which allows you to decide what an object will look like for the perspective who is requesting it. The perspective argument will be the perspective which is either passing an argument or returning a result an instance of your Copyable class.

    For security reasons, in order to allow a particular Copyable class to actually be copied, you must declare a RemoteCopy handler for that Copyable subclass. The easiest way to do this is to declare both in the same module, like so:

    from twisted.spread import flavors
    class Foo(flavors.Copyable):
        pass
    class RemoteFoo(flavors.RemoteCopy):
        pass
    flavors.setUnjellyableForClass(Foo, RemoteFoo)
    

    In this case, each time a Foo is copied between peers, a RemoteFoo will be instantiated and populated with the Foo’s state. If you do not do this, PB will complain that there have been security violations, and it may close the connection.

  • twisted.spread.pb.Cacheable

    Let me preface this with a warning: Cacheable may be hard to understand. The motivation for it may be unclear if you don’t have some experience with real-world applications that use remote method calling of some kind. Once you understand why you need it, what it does will likely seem simple and obvious, but if you get confused by this, forget about it and come back later. It’s possible to use PB without understanding Cacheable at all.

    Cacheable is a flavor which is designed to be copied only when necessary, and updated on the fly as changes are made to it. When passed as an argument or a return value, if a Cacheable exists on the side of the connection it is being copied to, it will be referred to by ID and not copied.

    Cacheable is designed to minimize errors involved in replicating an object between multiple servers, especially those related to having stale information. In order to do this, Cacheable automatically registers observers and queries state atomically, together. You can override the method getStateToCacheAndObserveFor(self, perspective, observer) in order to specify how your observers will be stored and updated.

    Similar to getStateToCopyFor , getStateToCacheAndObserveFor gets passed a perspective. It also gets passed an observer , which is a remote reference to a “secret” fourth referenceable flavor: RemoteCache .

    A RemoteCache is simply the object that represents your Cacheable on the other side of the connection. It is registered using the same method as RemoteCopy , above. RemoteCache is different, however, in that it will be referenced by its peer. It acts as a Referenceable, where all methods prefixed with observe_ will be callable remotely. It is recommended that your object maintain a list (note: library support for this is forthcoming!) of observers, and update them using callRemote when the Cacheable changes in a way that should be noticeable to its clients.

    Finally, when all references to a Cacheable from a given perspective are lost, stoppedObserving(perspective, observer) will be called on the Cacheable , with the same perspective/observer pair that getStateToCacheAndObserveFor was originally called with. Any cleanup remote calls can be made there, as well as removing the observer object from any lists which it was previously in. Any further calls to this observer object will be invalid.

Footnotes

[1]Most of Twisted is like this. Hell, most of Unix is like this: if you think it would be useful, someone else has probably thought that way in the past, and acted on it, and you can take advantage of the tool they created to solve the same problem you’re facing now.