Hitting Maximum Recursion Depth Using Python's Pickle / cPickle

Question

The background: I'm building a trie to represent a dictionary, using a minimal construction algorithm. The input list is 4.3M utf-8 strings, sorted lexicographically. The resulting graph is acyclic and has a maximum depth of 638 nodes. The first line of my script sets the recursion limit to 1100 via sys.setrecursionlimit().

The problem: I'd like to be able to serialize my trie to disk, so I can load it into memory without having to rebuild from scratch (roughly 22 minutes). I have tried both pickle.dump() and cPickle.dump(), with both the text and binary protocols. Each time, I get a stack-trace that looks like the following:

  File "/System/Library/Frameworks/Python.framework/Versions/2.5/lib/python2.5/pickle.py", line 649, in save_dict
    self._batch_setitems(obj.iteritems())
  File "/System/Library/Frameworks/Python.framework/Versions/2.5/lib/python2.5/pickle.py", line 663, in _batch_setitems
    save(v)
  File "/System/Library/Frameworks/Python.framework/Versions/2.5/lib/python2.5/pickle.py", line 286, in save
    f(self, obj) # Call unbound method with explicit self
  File "/System/Library/Frameworks/Python.framework/Versions/2.5/lib/python2.5/pickle.py", line 725, in save_inst
    save(stuff)
  File "/System/Library/Frameworks/Python.framework/Versions/2.5/lib/python2.5/pickle.py", line 286, in save
    f(self, obj) # Call unbound method with explicit self
  File "/System/Library/Frameworks/Python.framework/Versions/2.5/lib/python2.5/pickle.py", line 648, in save_dict
    self.memoize(obj)
RuntimeError: maximum recursion depth exceeded

My data structures are relatively simple: trie contains a reference to a start state, and defines some methods. dfa_state contains a boolean field, a string field, and a dictionary mapping from label to state.

I'm not very familiar with the inner workings of pickle - does my max recursion depth need to be greater/equal n times the depth of the trie for some n? Or could this be caused by something else I'm unaware of?

Update: Setting the recursion depth to 3000 didn't help, so this avenue doesn't look promising.

Update 2: You guys were right; I was being short-sighted in assuming that pickle would use a small nesting depth due to default recursion limitations. 10,000 did the trick.

Answer 1

From the docs:

Trying to pickle a highly recursive data structure may exceed the maximum recursion depth, a RuntimeError will be raised in this case. You can carefully raise this limit with sys.setrecursionlimit().

Although your trie implementation may be simple, it uses recursion and can lead to issues when converting to a persistent data structure.

My recommendation would be continue raising the recursion limit to see if there is an upper bound for the data you are working with and the trie implementation you are using.

Other then that, you can try changing your tree implementation to be "less recursive", if possible, or write an additional implementation that has data persistence built-in (use pickles and shelves in your implementation). Hope that helps

Answer 2

Double-check that your structure is indeed acyclic.

You could try bumping up the limit even further. There's a hard maximum that's platform dependent, but trying 50000 would be reasonable.

Also try pickling a trivially small version of your trie. If pickle dies even though it's only storing a couple three-letter words, then you know there's some fundamental problem with your trie and not pickle. But if it only happens when you try storing 10k words, then it might be the fault of a platform limitation in pickle.

Answer 3

Pickle does need to recursively walk your trie. If Pickle is using just 5 levels of function calls to do the work your trie of depth 638 will need the level set to more than 3000.

Try a much bigger number, the recursion limit is really just there to protect users from having to wait too long if the recursion falls in an infinite hole.

Pickle handles cycles ok, so it doesn't matter even if your trie had a cycle in there