Using Perspective Broker

Basic Example

The first example to look at is a complete (although somewhat trivial) application. It uses BrokerFactory() on the server side, and pb.getObjectAt() on the client side.

First we look at the server. This defines an Echoer class (derived from pb.Root), with a method called remote_echo(). pb.Root objects (because of their inheritance of pb.Referenceable, described later) can define methods with names of the form remote_*; a client which obtains a remote reference to that pb.Root object will be able to invoke those methods.

The pb.Root-ish object is given to a pb.BrokerFactory(). This is a Factory object like any other: the Protocol objects it creates for new connections know how to speak the PB protocol. The object you give to pb.BrokerFactory() becomes the root object, which simply makes it available for the client to retrieve. The client may only request references to the objects you want to provide it: this helps you implement your security model. Because it is so common to export just a single object (and because a remote_* method on that one can return a reference to any other object you might want to give out), the simplest example is one where the BrokerFactory is given the root object, and the client retrieves it.

The client side calls pb.getObjectAt to make a connection to a given port. This is a convenience function (not a method) which runs through the PB protocol steps necessary to retrieve the root object from a BrokerFactory sitting at the given port.

Because .getObjectAt() has to make a network connection and exchange some data, it may take a while, so it returns a Deferred, to which the gotObject() callback is attached. (See the documentation on Deferring Execution for a complete explanation of Deferreds). If and when the connection succeeds and a reference to the remote root object is obtained, this callback is run. The first argument passed to the callback is a remote reference to the distant root object. (you can give other arguments to the callback too, see the other parameters for .addCallback() and .addCallbacks()).

The callback does:

object.callRemote("echo", "hello network")

which causes the server's .remote_echo() method to be invoked. (running .callRemote("boom") would cause .remote_boom() to be run, etc). Again because of the delay involved, callRemote() returns a Deferred. Assuming the remote method was run without causing an exception (including an attempt to invoke an unknown method), the callback attached to that Deferred will be invoked with any objects that were returned by the remote method call.

In this example, the server's Echoer object has a method invoked, exactly as if some code on the server side had done:

echoer_object.remote_echo("hello network")

and from the definition of remote_echo() we see that this just returns the same string it was given: hello network.

From the client's point of view, the remote call gets another Deferred object instead of that string. callRemote() always returns a Deferred. This is why PB is described as a system for translucent remote method calls instead of transparent ones: you cannot pretend that the remote object is really local. Trying to do so (as some other RPC mechanisms do, coughCORBAcough) breaks down when faced with the asynchronous nature of the network. Using Deferreds turns out to be a very clean way to deal with the whole thing.

The remote reference object (the one given to getObjectAt()'s success callback) is an instance the RemoteReference class. This means you can use it to invoke methods on the remote object that it refers to. Only instances of RemoteReference eligible for .callRemote(). The RemoteReference object is the one that lives on the remote side (the client, in this case), not the local side (where the actual object is defined).

In our example, the local object is that Echoer() instance, which inherits from pb.Root, which inherits from pb.Referenceable. It is that Referenceable class that makes the object eligible to be available for remote method callsThere are a few other classes that can bestow this ability, but pb.Referenceable is the easiest to understand; see 'flavors' below for details on the others.. If you have an object that is Referenceable, then any client that manages to get a reference to it can invoke any remote_* methods they please.

You don't have to be a pb.Root to be remotely callable, but you do have to be pb.Referenceable. (Objects that inherit from pb.Referenceable but not from pb.Root can be remotely called, but only pb.Root-ish objects can be given to the BrokerFactory.)

Complete Example

A service is the global state associated with your application, which can contain things such as support for archiving objects, basic abstractions common to all users, and collections of domain-specific objects. A perspective is the representation of a user with respect to a particular service. For PB, a Perspective is where all interaction begins. When a user logs in for the first time, all the methods they can initially call are methods of their Perspective. The Perspective's methods can return objects which themselves have methods that you can call, as well as copies of objects, as described later.

For examples of these, we're returning to the TwistedQuotes project discussed in the Writing Plugins. The PB Service for TwistedQuotes is pretty small. The only thing it needs to keep track of for itself is the quoter object; PB's service, that we will inherit from, already keeps track of perspectives.

The perspective is a QuoteReader, which publishes one method. By subclassing Perspective, we are declaring that all methods with the perspective_ prefix are remotely accessible.

In order to get this Service published, so that we can actually connect to it, we need to re-visit the TAP building plugin, so we can actually get an Application that has a PB broker factory listening on a port. (The default port for PB is 8787.)

In the TAP builder, we create a QuoteService that wraps the quoter. We then create a QuoteReader perspective and attach it to the QuoteService, through the createPerspective call inherited from Service. Finally, we register with the QuoteService's authorizer.

Accessing this through a client is fairly easy, as we can use the pb.connect convenience function.

pb.connect will handle all the details of creating a connection and authenticating. It returns a Deferred, which will have its callback called when pb.connect connects to a perspective, and have its errback called when the object-connection fails for any reason, whether it's host lookup failure, connection refusal, or incorrect authentication credentials.

In this example, the connected callback should be made when the script is run. Looking at the code, it should be clear that in the event of a connection success, the client will print out a quote and exit. If you start up a server, you can see:

% mktap qotd --pb 8787
Saving qotd application to qotd.tap...
Saved.
% twistd -f qotd.tap
% python -c 'import TwistedQuotes.pbquoteclient'
An apple a day keeps the doctor away.

The argument to this callback, perspective, is a RemoteReference. The perspective reference represents a reference to a QuoteReader perspective object.

RemoteReference objects have one method which is their purpose for being: callRemote. This method allows you to call a remote method on the object being referred to by the Reference. RemoteReference.callRemote, like pb.connect, returns a Deferred. When a response to the method-call being sent arrives, the Deferred's callback or errback will be made, depending on whether an error occurred in processing the method-call.

This introduction to PB does not showcase all of the features that it provides, but hopefully it gives you a good idea of where to get started setting up your own application. Here are some of the othe building blocks you can use.

Passing more references

Here is an example of using pb.Referenceable in a second class. The second Referenceable object can have remote methods invoked too, just like the first. In this example, the initial root object has a method that returns a reference to the second object.

The root object has a method called remote_getTwo, which returns the Two() instance. On the client end, the callback gets a RemoteReference to that instance. The client can then invoke two's .remote_three() method.

You can use this technique to provide access to arbitrary sets of objects. Just remember that any object that might get passed over the wire must inherit from Referenceable (or one of the other flavors). If you try to pass a non-Referenceable object (say, by returning one from a remote_* method), you'll get an InsecureJelly exceptionThis can be overridden, by subclassing one of the Serializable flavors and defining custom serialization code for your class. See XXX for details..

References can come back to you

If your server gives a reference to a client, and then that client gives the reference back to the server, the server will wind up with the same object it gave out originally. The serialization layer watches for returning reference identifiers and turns them into actual objects. You need to stay aware of where the object lives: if it is on your side, you do actual method calls. If it is on the other side, you do .callRemote()The binary nature of this local vs. remote scheme works because you cannot give RemoteReferences to a third party. If you could, then your object A could go to B, B could give it to C, C might give it back to you, and you would be hard pressed to tell if the object lived in C's memory space, in B's, or if it was really your own object, tarnished and sullied after being handed down like a really ugly picture that your great aunt owned and which nobody wants but which nobody can bear to throw out. Ok, not really like that, but you get the idea..

The server gives a Two() instance to the client, who then returns the reference back to the server. The server compares the two given with the two received and shows that they are the same, and that both are real objects instead of remote references.

A few other techniques are demonstrated in pb2client.py. One is that the callbacks are are added with .addCallback instead of .addCallbacks. As you can tell from the Deferred documentation, .addCallback is a simplified form which only adds a success callback. The other is that to keep track of state from one callback to the next (the remote reference to the main One() object), we create a simple class, store the reference in an instance thereof, and point the callbacks at a sequence of bound methods. This is a convenient way to encapsulate a state machine. Each response kicks off the next method, and any data that needs to be carried from one state to the next can simply be saved as an attribute of the object.

Remember that the client can give you back any remote reference you've given them. Don't base your zillion-dollar stock-trading clearinghouse server on the idea that you trust the client to give you back the right reference. The security model inherent in PB means that they can only give you back a reference that you've given them for the current connection (not one you've given to someone else instead, nor one you gave them last time before the TCP session went down, nor one you haven't yet given to the client), but just like with URLs and HTTP cookies, the particular reference they give you is entirely under their control.

References to client-side objects

Anything that's Referenceable can get passed across the wire, in either direction. The client can give a reference to the server, and then the server can use .callRemote() to invoke methods on the client end. This fuzzes the distinction between client and server: the only real difference is who initiates the original TCP connection; after that it's all symmetric.

In this example, the client gives a reference to its own object to the server. The server then invokes a remote method on the client-side object.

Raising Remote Exceptions

Everything so far has covered what happens when things go right. What about when they go wrong? The Python Way is to raise an exception of some sort. The Twisted Way is the same.

The only special thing you do is to define your Exception subclass by deriving it from pb.Error. (You do define Exception subclasses, right? String exceptions are, like, so 5 minutes ago. Get with the new century, ok?). When any remotely-invokable method (like remote_* or perspective_*) raises a pb.Error-derived exception, a serialized form of that Exception object will be sent back over the wireTo be precise, the Failure will be sent if any exception is raised, not just pb.Error-derived ones. But the server will print ugly error messages if you raise ones that aren't derived from pb.Error.. The other side (which did callRemote) will have the errback callback run with a Failure object that contains a copy of the exception object. This Failure object can be queried to retrieve the error message and a stack traceback.

Failure is a special class, defined in twisted/python/failure.py, created to make it easier to handle asynchronous exceptions. Just as exception handlers can be nested, errback functions can be chained. If one errback can't handle the particular type of failure, it can be passed along to a errback handler further down the chain.

For simple purposes, think of the Failure as just a container for remotely-thrown Exception objects. To extract the string that was put into the exception, use its .getErrorMessage() method. To get the type of the exception (as a string), look at its .type attribute. The stack traceback is available too. The intent is to let the errback function get just as much information about the exception as Python's normal try: clauses do, even though the exception occurred in somebody else's memory space at some unknown time in the past.

% ./exc_client.py 
got remote Exception
 .__class__ = twisted.spread.pb.CopiedFailure
 .getErrorMessage() = fall down go boom
 .type = __main__.MyError
Main loop terminated.

Oh, and what happens if you raise some other kind of exception? Something that isn't subclassed from pb.Error? Well, those are called unexpected exceptions, which make Twisted think that something has really gone wrong. These will raise an exception on the server side. This won't break the connection (the exception is trapped, just like most exceptions that occur in response to network traffic), but it will print out an unsightly stack trace on the server's stderr with a message that says Peer Will Receive PB Traceback, just as if the exception had happened outside a remotely-invokable method. (This message will go the current log target, if log.startLogging was used to redirect it). The client will get the same Failure object in either case, but subclassing your exception from pb.Error is the way to tell Twisted that you expect this sort of exception, and that it is ok to just let the client handle it instead of also asking the server to complain. Look at exc_client.py and change it to invoke broken2() instead of broken() to see the change in the server's behavior.

If you don't add an errback function to the Deferred, then a remote exception will still send a Failure object back over, but it will get lodged in the Deferred with nowhere to go. When that Deferred finally goes out of scope, the side that did callRemote will emit a message about an Unhandled error in Deferred, along with an ugly stack trace. It can't raise an exception at that point (after all, the callRemote that triggered the problem is long gone), but it will emit a traceback. So be a good programmer and always add errback handlers, even if they are just calls to log.err.

Try/Except blocks and Failure.trap

To implement the equivalent of the Python try/except blocks (which can trap particular kinds of exceptions and pass others up to higher-level try/except blocks), you can use the .trap() method in conjunction with multiple errback handlers on the Deferred. Re-raising an exception in an errback handler serves to pass that new exception to the next handler in the chain. The trap method is given a list of exceptions to look for, and will re-raise anything that isn't on the list. Instead of passing unhandled exceptions up to an enclosing try block, this has the effect of passing the exception off to later errback handlers on the same Deferred. The trap calls are used in chained errbacks to test for each kind of exception in sequence.

% ./trap_client.py 
callOne: call with safe object
 method successful, response: response
callTwo: call with dangerous object
 InsecureJelly: you tried to send something unsafe to them
callThree: call that raises remote exception
 remote raised a MyException
telling them to shut down
callFour: call on stale reference
 stale reference: the client disconnected or crashed
% 

In this example, callTwo tries to send an instance of a locally-defined class through callRemote. The default security model implemented by pb.Jelly on the remote end will not allow unknown classes to be unserialized (i.e. taken off the wire as a stream of bytes and turned back into an object: a living, breathing instance of some class): one reason is that it does not know which local class ought to be used to create an instance that corresponds to the remote object

The naive approach of simply doing import SomeClass to match a remote caller who claims to have an object of type SomeClass could have nasty consequences for some modules that do significant operations in their __init__ methods (think telnetlib.Telnet(host='localhost', port='chargen'), or even more powerful classes that you have available in your server program). Allowing a remote entity to create arbitrary classes in your namespace is nearly equivalent to allowing them to run arbitrary code.

The pb.InsecureJelly exception arises because the class being sent over the wire has not been registered with the serialization layer (known as jelly). The easiest way to make it possible to copy entire class instances over the wire is to have them inherit from pb.Copyable, and then to use setUnjellyableForClass(remoteClass, localClass) on the receiving side. See XXX for an example.

Remember that trap re-raises exceptions that it wasn't asked to look for. You can only check for one set of exceptions per errback handler: all others must be checked in a subsequent handler. check_MyException shows how multiple kinds of exceptions can be checked in a single errback: give a list of exception types to trap, and it will return the matching member. In this case, the kinds of exceptions we are checking for (MyException and MyOtherException) may be raised by the remote end: they inherit from pb.Error.

The handler can return None to terminate processing of the errback chain (to be precise, it switches to the callback that follows the errback; if there is no callback then processing terminates). It is a good idea to put an errback that will catch everything (no trap tests, no possible chance of raising more exceptions, always returns None) at the end of the chain. Just as with regular try: except: handlers, you need to think carefully about ways in which your errback handlers could themselves raise exceptions. The extra importance in an asynchronous environment is that an exception that falls off the end of the Deferred will not be signalled until that Deferred goes out of scope, and at that point may only cause a log message (which could even be thrown away if log.startLogging is not used to point it at stdout or a log file). In contrast, a synchronous exception that is not handled by any other except: block will very visibly terminate the program immediately with a noisy stack trace.

callFour shows another kind of exception that can occur while using callRemote: pb.DeadReferenceError. This one occurs when the remote end has disconnected or crashed, leaving the local side with a stale reference. This kind of exception happens to be reported right away (XXX: is this guaranteed? probably not), so must be caught in a traditional synchronous try: except pb.DeadReferenceError block.

Yet another kind that can occur is a pb.PBConnectionLost exception. This occurs (asynchronously) if the connection was lost while you were waiting for a callRemote call to complete. When the line goes dead, all pending requests are terminated with this exception. Note that you have no way of knowing whether the request made it to the other end or not, nor how far along in processing it they had managed before the connection was lost. XXX: explain transaction semantics, find a decent reference.