This chapter focuses on how to use PB to pass complex types (specifically class instances) to and from a remote process. The first section is on simply copying the contents of an object to a remote process (pb.Copyable ). The second covers how to copy those contents once, then update them later when they change (Cacheable ).
From the previous chapter , you’ve seen how to pass basic types to a remote process, by using them in the arguments or return values of a callRemote function. However, if you’ve experimented with it, you may have discovered problems when trying to pass anything more complicated than a primitive int/list/dict/string type, or another pb.Referenceable object. At some point you want to pass entire objects between processes, instead of having to reduce them down to dictionaries on one end and then re-instantiating them on the other.
The most obvious and straightforward way to send an object to a remote process is with something like the following code. It also happens that this code doesn’t work, as will be explained below.
class LilyPond: def __init__(self, frogs): self.frogs = frogs pond = LilyPond(12) ref.callRemote("sendPond", pond)
If you try to run this, you might hope that a suitable remote end which
remote_sendPond method would see that method get
invoked with an instance from the
LilyPond class. But instead,
you’ll encounter the dreaded InsecureJelly exception. This is
Twisted’s way of telling you that you’ve violated a security restriction,
and that the receiving end refuses to accept your object.
What’s the big deal? What’s wrong with just copying a class into another process’ namespace?
Reversing the question might make it easier to see the issue: what is the problem with accepting a stranger’s request to create an arbitrary object in your local namespace? The real question is how much power you are granting them: what actions can they convince you to take on the basis of the bytes they are sending you over that remote connection.
Objects generally represent more power than basic types like strings and dictionaries because they also contain (or reference) code, which can modify other data structures when executed. Once previously-trusted data is subverted, the rest of the program is compromised.
The built-in Python “batteries included” classes are relatively
tame, but you still wouldn’t want to let a foreign program use them to
create arbitrary objects in your namespace or on your computer. Imagine a
protocol that involved sending a file-like object with a
method that was supposed to used later to retrieve a document. Then imagine
what if that object were created with
os.fdopen("~/.gnupg/secring.gpg") . Or an instance of
telnetlib.Telnet("localhost", "chargen") .
Classes you’ve written for your own program are likely to have far more
power. They may run code during
__init__ , or even have special
meaning simply because of their existence. A program might have
User objects to represent user accounts, and have a rule that
User objects in the system are referenced when
authorizing a login session. (In this system,
would probably add the object to a global list of known users). The simple
act of creating an object would give access to somebody. If you could be
tricked into creating a bad object, an unauthorized user would get
So object creation needs to be part of a system’s security design. The dotted line between “trusted inside” and “untrusted outside” needs to describe what may be done in response to outside events. One of those events is the receipt of an object through a PB remote procedure call, which is a request to create an object in your “inside” namespace. The question is what to do in response to it. For this reason, you must explicitly specify what remote classes will be accepted, and how their local representatives are to be created.
Another basic question to answer before we can do anything useful with an incoming serialized object is: what class should we create? The simplistic answer is to create the “same kind” that was serialized on the sender’s end of the wire, but this is not as easy or as straightforward as you might think. Remember that the request is coming from a different program, using a potentially different set of class libraries. In fact, since PB has also been implemented in Java, Emacs-Lisp, and other languages, there’s no guarantee that the sender is even running Python! All we know on the receiving end is a list of two things which describe the instance they are trying to send us: the name of the class, and a representation of the contents of the object.
This function takes a remote/sender class reference (either the
fully-qualified name as used by the sending end, or a class object from
which the name can be extracted), and a local/recipient class (used to
create the local representation for incoming serialized objects). Whenever
the remote end sends an object, the class name that they transmit is looked
up in the table controlled by this function. If a matching class is found,
it is used to create the local object. If not, you get the
In general you expect both ends to share the same codebase: either you
control the program that is running on both ends of the wire, or both
programs share some kind of common language that is implemented in code
which exists on both ends. You wouldn’t expect them to send you an object of
the MyFooziWhatZit class unless you also had a definition for that class. So
it is reasonable for the Jelly layer to reject all incoming classes except
the ones that you have explicitly marked with
setUnjellyableForClass . But keep in mind that the sender’s idea
User object might differ from the recipient’s, either
through namespace collisions between unrelated packages, version skew
between nodes that haven’t been updated at the same rate, or a malicious
intruder trying to cause your code to fail in some interesting or
potentially vulnerable way.
Ok, enough of this theory. How do you send a fully-fledged object from one side to the other?
#!/usr/bin/env python # Copyright (c) Twisted Matrix Laboratories. # See LICENSE for details. from twisted.spread import pb, jelly from twisted.python import log from twisted.internet import reactor class LilyPond: def setStuff(self, color, numFrogs): self.color = color self.numFrogs = numFrogs def countFrogs(self): print "%d frogs" % self.numFrogs class CopyPond(LilyPond, pb.Copyable): pass class Sender: def __init__(self, pond): self.pond = pond def got_obj(self, remote): self.remote = remote d = remote.callRemote("takePond", self.pond) d.addCallback(self.ok).addErrback(self.notOk) def ok(self, response): print "pond arrived", response reactor.stop() def notOk(self, failure): print "error during takePond:" if failure.type == jelly.InsecureJelly: print " InsecureJelly" else: print failure reactor.stop() return None def main(): from copy_sender import CopyPond # so it's not __main__.CopyPond pond = CopyPond() pond.setStuff("green", 7) pond.countFrogs() # class name: print ".".join([pond.__class__.__module__, pond.__class__.__name__]) sender = Sender(pond) factory = pb.PBClientFactory() reactor.connectTCP("localhost", 8800, factory) deferred = factory.getRootObject() deferred.addCallback(sender.got_obj) reactor.run() if __name__ == '__main__': main()
# Copyright (c) Twisted Matrix Laboratories. # See LICENSE for details. """ PB copy receiver example. This is a Twisted Application Configuration (tac) file. Run with e.g. twistd -ny copy_receiver.tac See the twistd(1) man page or http://twistedmatrix.com/documents/current/howto/application for details. """ import sys if __name__ == '__main__': print __doc__ sys.exit(1) from twisted.application import service, internet from twisted.internet import reactor from twisted.spread import pb from copy_sender import LilyPond, CopyPond from twisted.python import log #log.startLogging(sys.stdout) class ReceiverPond(pb.RemoteCopy, LilyPond): pass pb.setUnjellyableForClass(CopyPond, ReceiverPond) class Receiver(pb.Root): def remote_takePond(self, pond): print " got pond:", pond pond.countFrogs() return "safe and sound" # positive acknowledgement def remote_shutdown(self): reactor.stop() application = service.Application("copy_receiver") internet.TCPServer(8800, pb.PBServerFactory(Receiver())).setServiceParent( service.IServiceCollection(application))
The sending side has a class called
LilyPond . To make this
eligible for transport through
callRemote (either as an
argument, a return value, or something referenced by either of those [like a
dictionary value]), it must inherit from one of the four Serializable classes. In this section,
we focus on Copyable .
The copyable subclass of
LilyPond is called
CopyPond . We create an instance of it and send it through
callRemote as an argument to the receiver’s
remote_takePond method. The Jelly layer will serialize
(“jelly” ) that object as an instance with a class name of”copy_sender.CopyPond” and some chunk of data that represents the
pond.__class__.__name__ are used to derive the class name
string. The object’s getStateToCopy method is
used to get the state: this is provided by pb.Copyable , and the default just retrieves
self.__dict__ . This works just like the optional
__getstate__ method used by
pickle . The pair of
name and state are sent over the wire to the receiver.
The receiving end defines a local class named
to represent incoming
LilyPond instances. This class derives
from the sender’s
LilyPond class (with a fully-qualified name
copy_sender.LilyPond ), which specifies how we expect it to
behave. We trust that this is the same
LilyPond class as the
sender used. (At the very least, we hope ours will be able to accept a state
created by theirs). It also inherits from pb.RemoteCopy , which is a requirement for all
classes that act in this local-representative role (those which are given to
the second argument of
RemoteCopy provides the methods that tell the Jelly layer how
to create the local object from the incoming serialized state.
setUnjellyableForClass is used to register the two
classes. This has two effects: instances of the remote class (the first
argument) will be allowed in through the security layer, and instances of
the local class (the second argument) will be used to contain the state that
is transmitted when the sender serializes the remote object.
When the receiver unserializes (“unjellies” ) the object, it will
create an instance of the local
ReceiverPond class, and hand
the transmitted state (usually in the form of a dictionary) to that object’s
This acts just like the
__setstate__ method that
pickle uses when unserializing an object.
setCopyableState are distinct from
__setstate__ to allow objects to be
persisted (across time) differently than they are transmitted (across
When this is run, it produces the following output:
[-] twisted.spread.pb.PBServerFactory starting on 8800 [-] Starting factory <twisted.spread.pb.PBServerFactory instance at 0x406159cc> [Broker,0,127.0.0.1] got pond: <__builtin__.ReceiverPond instance at 0x406ec5ec> [Broker,0,127.0.0.1] 7 frogs
$ ./copy_sender.py 7 frogs copy_sender.CopyPond pond arrived safe and sound Main loop terminated. $
setCopyableState , you can control how the object is transmitted
over the wire. For example, you might want perform some data-reduction:
pre-compute some results instead of sending all the raw data over the wire.
Or you could replace references to a local object on the sender’s side with
markers before sending, then upon receipt replace those markers with
references to a receiver-side proxy that could perform the same operations
against a local cache of data.
Another good use for
getStateToCopy is to implement “local-only” attributes: data that is only accessible by the local
process, not to any remote users. For example, a
attribute could be removed from the object state before sending to a remote
system. Combined with the fact that
Copyable objects return
unchanged from a round trip, this could be used to build a
challenge-response system (in fact PB does this with
pb.Referenceable objects to implement authorization as
described here ).
getStateToCopy returns from the sending object will
be serialized and sent over the wire;
whatever comes over the wire and is responsible for setting up the state of
the object it lives in.
#!/usr/bin/env python # Copyright (c) Twisted Matrix Laboratories. # See LICENSE for details. from twisted.spread import pb class FrogPond: def __init__(self, numFrogs, numToads): self.numFrogs = numFrogs self.numToads = numToads def count(self): return self.numFrogs + self.numToads class SenderPond(FrogPond, pb.Copyable): def getStateToCopy(self): d = self.__dict__.copy() d['frogsAndToads'] = d['numFrogs'] + d['numToads'] del d['numFrogs'] del d['numToads'] return d class ReceiverPond(pb.RemoteCopy): def setCopyableState(self, state): self.__dict__ = state def count(self): return self.frogsAndToads pb.setUnjellyableForClass(SenderPond, ReceiverPond)
#!/usr/bin/env python # Copyright (c) Twisted Matrix Laboratories. # See LICENSE for details. from twisted.spread import pb, jelly from twisted.python import log from twisted.internet import reactor from copy2_classes import SenderPond class Sender: def __init__(self, pond): self.pond = pond def got_obj(self, obj): d = obj.callRemote("takePond", self.pond) d.addCallback(self.ok).addErrback(self.notOk) def ok(self, response): print "pond arrived", response reactor.stop() def notOk(self, failure): print "error during takePond:" if failure.type == jelly.InsecureJelly: print " InsecureJelly" else: print failure reactor.stop() return None def main(): pond = SenderPond(3, 4) print "count %d" % pond.count() sender = Sender(pond) factory = pb.PBClientFactory() reactor.connectTCP("localhost", 8800, factory) deferred = factory.getRootObject() deferred.addCallback(sender.got_obj) reactor.run() if __name__ == '__main__': main()
#!/usr/bin/env python # Copyright (c) Twisted Matrix Laboratories. # See LICENSE for details. from twisted.application import service, internet from twisted.internet import reactor from twisted.spread import pb import copy2_classes # needed to get ReceiverPond registered with Jelly class Receiver(pb.Root): def remote_takePond(self, pond): print " got pond:", pond print " count %d" % pond.count() return "safe and sound" # positive acknowledgement def remote_shutdown(self): reactor.stop() application = service.Application("copy_receiver") internet.TCPServer(8800, pb.PBServerFactory(Receiver())).setServiceParent( service.IServiceCollection(application))
In this example, the classes are defined in a separate source file, which
also sets up the binding between them. The
ReceiverPond are unrelated save for this binding: they happen
to implement the same methods, but use different internal instance variables
to accomplish them.
The recipient of the object doesn’t even have to import the class
definition into their namespace. It is sufficient that they import the class
definition (and thus execute the
statement). The Jelly layer remembers the class definition until a matching
object is received. The sender of the object needs the definition, of
course, to create the object in the first place.
When run, the
copy2 example emits the following:
$ twistd -n -y copy2_receiver.py [-] twisted.spread.pb.PBServerFactory starting on 8800 [-] Starting factory <twisted.spread.pb.PBServerFactory instance at 0x40604b4c> [Broker,0,127.0.0.1] got pond: <copy2_classes.ReceiverPond instance at 0x406eb2ac> [Broker,0,127.0.0.1] count 7
$ ./copy2_sender.py count 7 pond arrived safe and sound Main loop terminated.
setUnjellyableForClassmust refer to the class as known by the sender . The sender has no way of knowing about how your local
importstatements are set up, and Python’s flexible namespace semantics allow you to access the same class through a variety of different names. You must match whatever the sender does. Having both ends import the class from a separate file, using a canonical module name (no “sibling imports” ), is a good way to get this right, especially when both the sending and the receiving classes are defined together, with the
setUnjellyableForClassimmediately following them.
pb.RemoteCopy. This will also make it possible to send the same class symmetrically back and forth over the wire. But don’t get confused about when it is coming (and using
setCopyableState) versus when it is going (using
errbackhandler. If you do not add one to the
callRemote, then you will never receive notification of the problem.
__init__method that takes no arguments. All setup must be performed in the
setCopyableStatemethod. As the docstring on RemoteCopy says, don’t implement a constructor that requires arguments in a subclass of
pb.Copyableis mostly implemented in
twisted.spread.flavors, and the docstrings there are the best source of additional information.
Copyableis also used in twisted.web.distrib to deliver HTTP requests to other programs for rendering, allowing subtrees of URL space to be delegated to multiple programs (on multiple machines).
Sometimes the object you want to send to the remote process is big and slow. “big” means it takes a lot of data (storage, network bandwidth, processing) to represent its state. “slow” means that state doesn’t change very frequently. It may be more efficient to send the full state only once, the first time it is needed, then afterwards only send the differences or changes in state whenever it is modified. The pb.Cacheable class provides a framework to implement this.
pb.Cacheable is derived from pb.Copyable , so it is based upon the idea of an object’s state being captured on the sending side, and then turned into a new object on the receiving side. This is extended to have an object “publishing” on the sending side (derived from pb.Cacheable ), matched with one”observing” on the receiving side (derived from pb.RemoteCache ).
To effectively use
pb.Cacheable , you need to isolate changes
to your object into accessor functions (specifically “setter”
functions). Your object needs to get control every single time some
attribute is changed  .
You derive your sender-side class from
pb.Cacheable , and you
add two methods: getStateToCacheAndObserveFor
and stoppedObserving . The first
is called when a remote caching reference is first created, and retrieves
the data with which the cache is first filled. It also provides an
object called the “observer”  that points at that receiver-side cache. Every time the state of the object
is changed, you give a message to the observer, informing them of the
change. The other method,
stoppedObserving , is called when the
remote cache goes away, so that you can stop sending updates.
On the receiver end, you make your cache class inherit from pb.RemoteCache , and implement the
setCopyableState as you would for a
object. In addition, you must implement methods to receive the updates sent
to the observer by the
pb.Cacheable : these methods should have
names that start with
observe_ , and match the
callRemote invocations from the sender side just as the usual
perspective_* methods match normal
The first time a reference to the
pb.Cacheable object is
sent to any particular recipient, a sender-side Observer will be created for
it, and the
getStateToCacheAndObserveFor method will be called
to get the current state and register the Observer. The state which that
returns is sent to the remote end and turned into a local representation
setCopyableState just like
described above (in fact it inherits from that class).
After that, your “setter” functions on the sender side should call
callRemote on the Observer, which causes
methods to run on the receiver, which are then supposed to update the
receiver-local (cached) state.
When the receiver stops following the cached object and the last
reference goes away, the
pb.RemoteCache object can be freed.
Just before it dies, it tells the sender side it no longer cares about the
original object. When that reference count goes to zero, the
Observer goes away and the
pb.Cacheable object can stop
announcing every change that takes place. The stoppedObserving method is
used to tell the
pb.Cacheable that the Observer has gone
classes in place, bound together by a call to
pb.setUnjellyableForClass , all that remains is to pass a
reference to your
pb.Cacheable over the wire to the remote end.
pb.RemoteCache object will automatically be
created, and the matching methods will be used to keep the receiver-side
slave object in sync with the sender-side master object.
Here is a complete example, in which the
controlled by the sending side, and the
SlaveDuckPond is a
cache that tracks changes to the master:
#!/usr/bin/env python # Copyright (c) Twisted Matrix Laboratories. # See LICENSE for details. from twisted.spread import pb class MasterDuckPond(pb.Cacheable): def __init__(self, ducks): self.observers =  self.ducks = ducks def count(self): print "I have [%d] ducks" % len(self.ducks) def addDuck(self, duck): self.ducks.append(duck) for o in self.observers: o.callRemote('addDuck', duck) def removeDuck(self, duck): self.ducks.remove(duck) for o in self.observers: o.callRemote('removeDuck', duck) def getStateToCacheAndObserveFor(self, perspective, observer): self.observers.append(observer) # you should ignore pb.Cacheable-specific state, like self.observers return self.ducks # in this case, just a list of ducks def stoppedObserving(self, perspective, observer): self.observers.remove(observer) class SlaveDuckPond(pb.RemoteCache): # This is a cache of a remote MasterDuckPond def count(self): return len(self.cacheducks) def getDucks(self): return self.cacheducks def setCopyableState(self, state): print " cache - sitting, er, setting ducks" self.cacheducks = state def observe_addDuck(self, newDuck): print " cache - addDuck" self.cacheducks.append(newDuck) def observe_removeDuck(self, deadDuck): print " cache - removeDuck" self.cacheducks.remove(deadDuck) pb.setUnjellyableForClass(MasterDuckPond, SlaveDuckPond)
#!/usr/bin/env python # Copyright (c) Twisted Matrix Laboratories. # See LICENSE for details. from twisted.spread import pb, jelly from twisted.python import log from twisted.internet import reactor from cache_classes import MasterDuckPond class Sender: def __init__(self, pond): self.pond = pond def phase1(self, remote): self.remote = remote d = remote.callRemote("takePond", self.pond) d.addCallback(self.phase2).addErrback(log.err) def phase2(self, response): self.pond.addDuck("ugly duckling") self.pond.count() reactor.callLater(1, self.phase3) def phase3(self): d = self.remote.callRemote("checkDucks") d.addCallback(self.phase4).addErrback(log.err) def phase4(self, dummy): self.pond.removeDuck("one duck") self.pond.count() self.remote.callRemote("checkDucks") d = self.remote.callRemote("ignorePond") d.addCallback(self.phase5) def phase5(self, dummy): d = self.remote.callRemote("shutdown") d.addCallback(self.phase6) def phase6(self, dummy): reactor.stop() def main(): master = MasterDuckPond(["one duck", "two duck"]) master.count() sender = Sender(master) factory = pb.PBClientFactory() reactor.connectTCP("localhost", 8800, factory) deferred = factory.getRootObject() deferred.addCallback(sender.phase1) reactor.run() if __name__ == '__main__': main()
#!/usr/bin/env python # Copyright (c) Twisted Matrix Laboratories. # See LICENSE for details. from twisted.application import service, internet from twisted.internet import reactor from twisted.spread import pb import cache_classes class Receiver(pb.Root): def remote_takePond(self, pond): self.pond = pond print "got pond:", pond # a DuckPondCache self.remote_checkDucks() def remote_checkDucks(self): print "[%d] ducks: " % self.pond.count(), self.pond.getDucks() def remote_ignorePond(self): # stop watching the pond print "dropping pond" # gc causes __del__ causes 'decache' msg causes stoppedObserving self.pond = None def remote_shutdown(self): reactor.stop() application = service.Application("copy_receiver") internet.TCPServer(8800, pb.PBServerFactory(Receiver())).setServiceParent( service.IServiceCollection(application))
When run, this example emits the following:
$ twistd -n -y cache_receiver.py [-] twisted.spread.pb.PBServerFactory starting on 8800 [-] Starting factory <twisted.spread.pb.PBServerFactory instance at 0x40615acc> [Broker,0,127.0.0.1] cache - sitting, er, setting ducks [Broker,0,127.0.0.1] got pond: <cache_classes.SlaveDuckPond instance at 0x406eb5ec> [Broker,0,127.0.0.1]  ducks: ['one duck', 'two duck'] [Broker,0,127.0.0.1] cache - addDuck [Broker,0,127.0.0.1]  ducks: ['one duck', 'two duck', 'ugly duckling'] [Broker,0,127.0.0.1] cache - removeDuck [Broker,0,127.0.0.1]  ducks: ['two duck', 'ugly duckling'] [Broker,0,127.0.0.1] dropping pond
$ ./cache_sender.py I have  ducks I have  ducks I have  ducks Main loop terminated.
Points to notice:
There is one
Observer for each remote program that holds
an active reference. Multiple references inside the same program don’t
matter: the serialization layer notices the duplicates and does the
appropriate reference counting  .
Multiple Observers need to be kept in a list, and all of them need to be updated when something changes. By sending the initial state at the same time as you add the observer to the list, in a single atomic action that cannot be interrupted by a state change, you insure that you can send the same status update to all the observers.
observer.callRemote calls can still fail. If the
remote side has disconnected very recently and
stoppedObserving has not yet been called, you may get a
DeadReferenceError . It is a good idea to add an errback to
callRemote s to throw away such an error. This is a
observer.callRemote('foo', arg).addErrback(lambda f: None)
getStateToCacheAndObserverFormust return some object that represents the current state of the object. This may simply be the object’s
__dict__attribute. It is a good idea to remove the
pb.Cacheable-specific members of it before sending it to the remote end. The list of Observers, in particular, should be left out, to avoid dizzying recursive Cacheable references. The mind boggles as to the potential consequences of leaving in such an item.
perspectiveargument is available to
getStateToCacheAndObserveFor, as well as
stoppedObserving. I think the purpose of this is to allow viewer-specific changes to the way the cache is updated. If all remote viewers are supposed to see the same data, it can be ignored.
Cacheable, and might be a source of further information.
|||Note that, in this context, “unjelly” is
a verb with the opposite meaning of “jelly” . The verb “to jelly”
means to serialize an object or data structure into a sequence of bytes (or
other primitive transmittable/storable representation), while “to unjelly” means to unserialize the bytestream into a live object in the
receiver’s memory space. “Unjellyable” is a noun, (not an
adjective), referring to the class that serves as a destination or
recipient of the unjellying process. “A is unjellyable into B” means
that a serialized representation A (of some remote object) can be
unserialized into a local object of type B. It is these objects “B”
that are the “Unjellyable” second argument of the
|||pb.RemoteCopy is actually defined
in twisted.spread.flavors , but
|||Of course you could be clever and
add a hook to |
|||This is actually a RemoteCacheObserver , but it isn’t very
useful to subclass or modify, so simply treat it as a little demon that sits
in your |
|||This applies to multiple references through the same Broker . If you’ve managed to make multiple TCP connections to the same program, you deserve whatever you get.|