[LONG] Re: [Twisted-Python] Why do I get a _Dereference instance?

Brian Warner warner at lothar.com
Wed Jul 9 04:51:49 EDT 2003

"Patrik Blommaskog" <pb_twisted at olga.mine.nu> writes:

> 	def setCopyableState(self, state):
> 	    self.__dict__.update(state)
> gives the strange behavior

> While:
> 	def setCopyableState(self, state):
> 	    self.__dict__ = state
> Gives the expected

> This feels like the twilight zone. Can somebody please tell me what is going
> on? I have attached the code for my experiment.

You're seeing an unfortunate artifact that results from the way circular
references are handled in the standard Jelly code. The problem has been
addressed somewhat in newjelly, but there are deeper issues which mean the
problem may not be possible to fix completely.

The following is a longish discussion of how jelly works and why this is
mostly unavoidable, followed by some design possibilities for other ways to
deal with it (newbanana enthusiasts are asked to read and comment upon the
last bit).

1: Mutable Containers

To start with, think about how a self-referential list must be unserialized.
This is a "mutable container", a category of objects which also includes
dictionaries. Something like:
  >>> a = [1,2]
  >>> a.append(a)
  >>> a.append(4)
  >>> a
  [1, 2, [...], 4]

When this is serialized, jelly (just like pickle) gets to the third element
and recognizes that this is something it has serialized before (or at least
it has started the serialization process). Rather than engage in pointless
infinite recursion, it inserts a 'dereference' marker that tells the
unserializer to insert a reference to something it already knows about.

(jelly, newjelly, and pickle all handle this slightly differently. oldjelly
puts explicit 'reference' markers on things that will need to be used later,
while newjelly uses an implicit reference counter that is incremented by
one. I use newjelly examples here because they are slightly easier to
follow.. oldjelly behaves pretty much the same way)

The resulting sexp stream (for newjelly) is:
  >>> from twisted.spread.newjelly import jelly
  >>> jelly(a)
  ['list', 1, 2, ['dereference', 0], 4]

The unserialization process handles one token at a time. It sees the "list"
token and creates an empty list, and stashes a reference in the "dereference
table" at index 0. It sees the number tokens and appends "1" and "2" to that
previously-empty list. Then it gets the "dereference" marker, looks up the
index in the dereference table, and ends up appending that list reference to
the list. Then it sees the "4" and appends a number, then it sees the
closing bracket and finishes up. This technique works for
arbitrarily-deeply-nested self-referential mutable container objects.

Note that the list is not really "complete" until the closing bracket, but
it needs to be referenceable before that point. Because we can create empty
lists and add things to them later, the list is referenceable starting with
the open bracket, and complete at the closing bracket.

2: Immutable Containers

Now think about a somewhat self-referential tuple. This is a "immutable
container". Of course it can't be directly self-referential: the cycle (in
graph-theory terms) must pass through a mutable container at some point,
otherwise the object could never be constructed in the first place.
  >>> b = [1,2]
  >>> c = (b,)
  >>> b.append(c)
  >>> c
  ([1, 2, ([...],)],)

This is serialized with similar dereference markers:

  >>> jelly(c)
  ['tuple', ['list', 1, 2, ['dereference', 0]]]

The difference this time is the immutable tuple (which needs to be
referenced later). It is not possible to create an empty tuple and add its
contents later. It is "referenceable" and "complete" at the same time,
unlike the list object.

To deal with this, we need a placeholder object. In jelly this is a
"_Tuple", which derives from twisted.persisted.crefutil.NotKnown . It acts
like a list while we add contents, but if we ever add the _Tuple to another
collection (like a dict or a list), the _Tuple remembers the collection
object and index to which it was added. When the closing bracket is
received, the _Tuple turns its list into a real tuple (making it "complete"
and "referenceable"), then walks through all the places that point to it and
fix them up. It does this by replacing the old _Tuple reference with a
reference to the new, complete tuple object. If the container that needed to
be fixed up is itself a _Tuple (or other NotKnown subclass), it might now be
complete, so it is given a chance to fix up *its* references, etc.

Cycles in the object graph are thus resolved in a chain, starting with the
node closest to the root object (the original thing passed to jelly()) and
propogating outwards to the deepest as-yet-unresolved container.

The salient points are: tuples and other "immutable containers" (which
currently includes bound instance methods) cannot be referenced until they
are complete, which requires all of their contents to be referenceable.
Placeholder objects and a post-completion fix-up step are used to deal with
this frustrating property.

3: Class Instances

Now consider regular instance objects, which can be described by a fully
qualified class name and a state dictionary. Clearly they are mutable
containers: you can trivially create cycles of nothing but instances:

  >>> class A: pass
  >>> a1 = A(); a2 = A()
  >>> a1.a2 = a2; a2.a1 = a1
  >>> a1
  <__main__.A instance at 0x81abea4>
  >>> jelly(a1)
  ['__main__.A', ['dictionary', ['a2', ['__main__.A', ['dictionary', ['a1', ['dereference', 0]]]]]]]

The particular problem you're seeing is a misbehavior in oldjelly that was
addressed in newjelly. Oldjelly uses explicit reference tokens to mark
things that will be pointed to by dereference tokens later. The syntax of
those reference tokens is:

 ['reference', number, [thing being referenced]]

The oldjelly serialization of that pair of objects looks like this:
  >>> from twisted.spread.jelly import jelly as oldjelly
  >>> oldjelly(a1)
  ['reference', 1, ['__main__.A', ['dictionary', ['a2', ['__main__.A', ['dictionary', ['a1', ['dereference', 1]]]]]]]]

When oldjelly sees the 'reference' tag, it unserializes the "thing being
referenced", then stuffs a reference to it in the table. While that
unserialization is taking place (when the stack of things being unserialized
is actually: (reference 1), (instance a1), (a1.__dict__), (key 'a2'),
(instance a2), (a2.__dict__), (key 'a1')), the unjellying code hits the
"dereference 1" marker. Now, that reference hasn't been created yet: it
won't be until the enclosing object is finished. Oldjelly deals with this
using a _Deference object, which is another NotKnown subclass like _Tuple.
When the object is finally available, the _Dereference walks through the
places that point to it (in this case, a2.a1) and fix them up to point to
the correct object.

Until that point, a2.a1 is a _Dereference object. After the object is
complete, it is the correct A instance.

To perform that fixup, the _Dereference knows the dictionary it was added
to. Of course, it doesn't know about any *other* dictionaries that the
_Dereference might have been copied into, such as the one you used here:

> 	def setCopyableState(self, state):
> 	    self.__dict__.update(state)

The original 'state' dictionary is fixed-up with the correct instance
reference, but the independent __dict__ is not, so you wind up seeing the
_Dereference object that should have been hidden.

Newjelly deals with this differently, but is vulnerable to a similar

In newjelly, there are no explicit 'reference' tags. Instead, every sexp
that creates a new container object will increment the implicit reference
number. We mainly did this to remove the overhead of those reference tags,
both space (consumed in the bytestream), and computation (spent trying to
decide whether the object was going to be referenced more than once before
including the reference tag). It was far easier to simply track references
to *everything*, on both sides of the wire.

An important side-effect was that dereference tags will now *always* have a
matching reference, which is inserted into the reference table as soon as
the opening bracket is processed. The new unjelly_dereference() code
eliminated _Dereference altogether.

The (incomplete) instance object is referenceable as soon as the opening
bracket has been processed. The same is true for the state dictionary. This
means that the state handed to the .unjellyFor() method (and therefore to
.setCopyableState()) will reference the correct (although incomplete)
instance objects. No fixups are necessary, so there isn't a chance to fix-up
the wrong dictionary if "self.__dict__.update(state)" is used.

As a result, this pair of objects, when transported by newjelly, should
arrive without weird _Dereference artifacts, fixing your immediate problem.

4: the Remaining Problem

Newjelly is still vulnerable to a similar problem that results when an
immutable container (which cannot be fully resolved when the instance is
complete) is part of the instance's state:
  >>> a = A()
  >>> t = (1,2,a,4)
  >>> a.t = t
  >>> jelly(t)
  ['tuple', 1, 2, ['__main__.A', ['dictionary', ['t', ['dereference', 0]]]], 4]

Here, the tuple cannot be created until the '4' has been processed, so it is
neither complete nor referenceable when the instance's state is being
created. As a result, that state will have a _Tuple object that will
eventually be replaced by the real tuple. This means that the _Tuple needs
to remember that collection object "a", key "t" (a.t) needs to be fixed up
when the _Tuple is resolved. As before, if the _Tuple reference is copied
into a different dictionary, this fix-up will miss the real target, and the
instance will be left with a spurious _Tuple instance.

So, even though newjelly fixes the "instance.foo = instance" problem, it is
still vulnerable to certain "instance.foo = tuple" problems. Of course, this
is only when you use "self.__dict__.update(state)", so the simplest answer
is to just not do that :).

The more fundamental problem here is in presenting a coherent object graph
to the functions that are involved in creating that object graph. When
.setCopyableState() is called, the state dictionary may contain unresolved
references, and it may contain references to objects that are not yet
complete (other instances which have not yet had their own .setCopyableState
called). This puts a restriction upon those methods: they may not be able to
use all of their own state, or look inside related objects, or call methods
on related objects.

5: Possibilities  (newbanana people wake up)

I think this is a Hard Problem, and it isn't generally possible to figure
out a "correct" order in which to run these functions. It is very much like
running finalization routines during object destruction. It may be possible
to implement phases of deserialization.. .setCopyableState() could be run as
soon as the state dictionary is full of referenceable objects (putting it
off until all tuples were created, for example), and then call some other
method after all the objects are complete. Perhaps having a more uniform
scheme for indicating "incomplete" objects would help, one where instances
could register callbacks to be run after a certain list of other objects are
complete. The unserializer would then have to evaluate the callback requests
and come up with a legal sequence in which they could be executed, a
potentially impossible task if there are cycles in the dependency graph.

Some smaller-scale solutions might be worth exploring:

 Instead of things like _Tuple just storing a (collection,key) pair and
 updating that reference blindly, perhaps they should call a
 .resolvedReference(key, value) method. This would give the instance a
 chance to run some code and do something useful with the newly available

 gc.get_referrers() could be used to find the places which point to the
 _Tuple, then any which were instances could have a .resolvedReference
 method called (and any which were lists/dicts could be searched for the
 places that use the _Tuple, and then updated in place). This would probably
 fix the __dict__.update problem, at the expense of performance
 (get_referrers is very slow) and the inclusion of an uncomfortable amount
 of magic.

 The newbanana scheme (sandbox/warner/slicer.py) uses Deferreds as
 placeholder objects. Arbitrary callbacks can be added, which might make it
 easier to handle the "self.__dict__.update(state)" pattern.

Anyway, that was probably a much longer answer than you were looking for,
but I thought it important to explain the larger problem (which remains
unfixed) rather than simply saying "it's been fixed in newjelly", or "don't
use __dict__.update".

hope that helps more than it confuses,

More information about the Twisted-Python mailing list