Overview

When programming with the ZODB, Python dictionaries aren’t always what you need. The most important case is where you want to store a very large mapping. When a Python dictionary is accessed in a ZODB, the whole dictionary has to be unpickled and brought into memory. If you’re storing something very large, such as a 100,000-entry user database, unpickling such a large object will be slow. BTrees are a balanced tree data structure that behave like a mapping but distribute keys throughout a number of tree nodes. The nodes are stored in sorted order (this has important consequences – see below). Nodes are then only unpickled and brought into memory as they’re accessed, so the entire tree doesn’t have to occupy memory (unless you really are touching every single key).

Behaviour

The keys, values(), and items() methods on BTree and TreeSet types do not materialize a list with all of the data. Instead, they return lazy sequences that fetch data from the BTree as needed. They also support optional arguments to specify the minimum and maximum values to return, often called “range searching”. Because all these types are stored in sorted order, range searching is very efficient.

The keys(), values(), and items() methods on Bucket and Set types do return lists with all the data. Starting in ZODB 3.3, there are also iterkeys(), itervalues(), and iteritems() methods that return iterators (in the Python 2.2 sense). Those methods also apply to BTree and TreeSet objects.

A BTree object supports all the methods you would expect of a mapping, with a few extensions that exploit the fact that the keys are sorted. The example below demonstrates how some of the methods work. The extra methods are minKey() and maxKey(), which find the minimum and maximum key value subject to an optional bound argument, and byValue(), which should probably be ignored (it’s hard to explain exactly what it does, and as a result it’s almost never used – best to consider it deprecated). The various methods for enumerating keys, values and items also accept minimum and maximum key arguments (“range search”), and (new in ZODB 3.3) optional Boolean arguments to control whether a range search is inclusive or exclusive of the range’s endpoints.

>>> from BTrees.OOBTree import OOBTree
>>> t = OOBTree()
>>> t.update({1: "red", 2: "green", 3: "blue", 4: "spades"})
>>> len(t)
4
>>> t[2]
'green'
>>> s = t.keys() # this is a "lazy" sequence object
>>> s
<...TreeItems object at ...>
>>> len(s)  # it acts like a Python list
4
>>> s[-2]
3
>>> list(s) # materialize the full list
[1, 2, 3, 4]
>>> list(t.values())
['red', 'green', 'blue', 'spades']
>>> list(t.values(1, 2)) # values at keys in 1 to 2 inclusive
['red', 'green']
>>> list(t.values(2))    # values at keys >= 2
['green', 'blue', 'spades']
>>> list(t.values(min=1, max=4))  # keyword args new in ZODB 3.3
['red', 'green', 'blue', 'spades']
>>> list(t.values(min=1, max=4, excludemin=True, excludemax=True))
['green', 'blue']
>>> t.minKey()     # smallest key
1
>>> t.minKey(1.5)  # smallest key >= 1.5
2
>>> [k for k in t.keys()]
[1, 2, 3, 4]
>>> [k for k in t]    # new in ZODB 3.3
[1, 2, 3, 4]
>>> [pair for pair in t.iteritems()]  # new in ZODB 3.3
[(1, 'red'), (2, 'green'), (3, 'blue'), (4, 'spades')]
>>> t.has_key(4)  # returns a true value
True
>>> t.has_key(5)
False
>>> 4 in t  # new in ZODB 3.3
True
>>> 5 in t  # new in ZODB 3.3
False
>>>

Each of the modules also defines some functions that operate on BTrees – difference(), union(), and intersection(). The difference() function returns a Bucket, while the other two methods return a Set. If the keys are integers, then the module also defines multiunion(). If the values are integers or floats, then the module also defines weightedIntersection() and weightedUnion(). The function doc strings describe each function briefly.

Interfaces defines the operations, and is the official documentation. Note that the interfaces don’t define the concrete types returned by most operations, and you shouldn’t rely on the concrete types that happen to be returned: stick to operations guaranteed by the interface. In particular, note that the interfaces don’t specify anything about comparison behavior, and so nothing about it is guaranteed. In ZODB 3.3, for example, two BTrees happen to use Python’s default object comparison, which amounts to comparing the (arbitrary but fixed) memory addresses of the BTrees. This may or may not be true in future releases. If the interfaces don’t specify a behavior, then whether that behavior appears to work, and exactly happens if it does appear to work, are undefined and should not be relied on.

Total Ordering and Persistence

The BTree-based data structures differ from Python dicts in several fundamental ways. One of the most important is that while dicts require that keys support hash codes and equality comparison, the BTree-based structures don’t use hash codes and require a total ordering on keys.

Total ordering means three things:

  1. Reflexive. For each x, x == x is true.

  2. Trichotomy. For each x and y, exactly one of x < y, x == y, and x > y is true.

  3. Transitivity. Whenever x <= y and y <= z, it’s also true that x <= z.

The default comparison functions for most objects that come with Python satisfy these rules, with some crucial cautions explained later. Complex numbers are an example of an object whose default comparison function does not satisfy these rules: complex numbers only support == and != comparisons, and raise an exception if you try to compare them in any other way. They don’t satisfy the trichotomy rule, and must not be used as keys in BTree-based data structures (although note that complex numbers can be used as keys in Python dicts, which do not require a total ordering).

Examples of objects that are wholly safe to use as keys in BTree-based structures include ints, longs, floats, 8-bit strings, Unicode strings, and tuples composed (possibly recursively) of objects of wholly safe types.

It’s important to realize that even if two types satisfy the rules on their own, mixing objects of those types may not. For example, 8-bit strings and Unicode strings both supply total orderings, but mixing the two loses trichotomy; e.g., 'x' < chr(255) and u'x' == 'x', but trying to compare chr(255) to u'x' raises an exception. Partly for this reason (another is given later), it can be dangerous to use keys with multiple types in a single BTree-based structure. Don’t try to do that, and you don’t have to worry about it.

Another potential problem is mutability: when a key is inserted in a BTree- based structure, it must retain the same order relative to the other keys over time. This is easy to run afoul of if you use mutable objects as keys. For example, lists supply a total ordering, and then

>>> L1, L2, L3 = [1], [2], [3]
>>> from BTrees.OOBTree import OOSet
>>> s = OOSet((L2, L3, L1))  # this is fine, so far
>>> list(s.keys())           # note that the lists are in sorted order
[[1], [2], [3]]
>>> s.has_key([3])           # and [3] is in the set
True
>>> L2[0] = 5                # horrible -- the set is insane now
>>> s.has_key([3])           # for example, it's insane this way
False
>>> s.__class__
<class 'BTrees.OOBTree.OOSet'>
>>> list(s)
[[1], [5], [3]]

Key lookup relies on that the keys remain in sorted order (an efficient form of binary search is used). By mutating key L2 after inserting it, we destroyed the invariant that the OOSet is sorted. As a result, all future operations on this set are unpredictable.

A subtler variant of this problem arises due to persistence: by default, Python does several kinds of comparison by comparing the memory addresses of two objects. Because Python never moves an object in memory, this does supply a usable (albeit arbitrary) total ordering across the life of a program run (an object’s memory address doesn’t change). But if objects compared in this way are used as keys of a BTree-based structure that’s stored in a database, when the objects are loaded from the database again they will almost certainly wind up at different memory addresses. There’s no guarantee then that if key K1 had a memory address smaller than the memory address of key K2 at the time K1 and K2 were inserted in a BTree, K1’s address will also be smaller than K2’s when that BTree is loaded from a database later. The result will be an insane BTree, where various operations do and don’t work as expected, seemingly at random.

Now each of the types identified above as “wholly safe to use” never compares two instances of that type by memory address, so there’s nothing to worry about here if you use keys of those types. The most common mistake is to use keys that are instances of a user-defined class that doesn’t supply its own __cmp__() method. Python compares such instances by memory address. This is fine if such instances are used as keys in temporary BTree-based structures used only in a single program run. It can be disastrous if that BTree-based structure is stored to a database, though.

>>> class C:
...     pass
...
>>> a, b = C(), C()
>>> print(a < b)   # this may print 0 if you try it
True
>>> del a, b
>>> a, b = C(), C()
>>> print(a < b)   # and this may print 0 or 1
False
>>>

That example illustrates that comparison of instances of classes that don’t define __cmp__() yields arbitrary results (but consistent results within a single program run).

Another problem occurs with instances of classes that do define __cmp__(), but define it incorrectly. It’s possible but rare for a custom __cmp__() implementation to violate one of the three required formal properties directly. It’s more common for it to “fall back” to address-based comparison by mistake. For example,

>>> class Mine:
...     def __cmp__(self, other):
...         if other.__class__ is Mine:
...             return cmp(self.data, other.data)
...         else:
...             return cmp(self.data, other)

It’s quite possible there that the else clause allows a result to be computed based on memory address. The bug won’t show up until a BTree-based structure uses objects of class Mine as keys, and also objects of other types as keys, and the structure is loaded from a database, and a sequence of comparisons happens to execute the else clause in a case where the relative order of object memory addresses happened to change.

This is as difficult to track down as it sounds, so best to stay far away from the possibility.

You’ll stay out of trouble by follwing these rules, violating them only with great care:

  1. Use objects of simple immutable types as keys in BTree-based data structures.

  2. Within a single BTree-based data structure, use objects of a single type as keys. Don’t use multiple key types in a single structure.

  3. If you want to use class instances as keys, and there’s any possibility that the structure may be stored in a database, it’s crucial that the class define a __cmp__() method, and that the method is carefully implemented.

    Any part of a comparison implementation that relies (explicitly or implicitly) on an address-based comparison result will eventually cause serious failure.

  4. Do not use Persistent objects as keys, or objects of a subclass of Persistent.

That last item may be surprising. It stems from details of how conflict resolution is implemented: the states passed to conflict resolution do not materialize persistent subobjects (if a persistent object P is a key in a BTree, then P is a subobject of the bucket containing P). Instead, if an object O references a persistent subobject P directly, and O is involved in a conflict, the states passed to conflict resolution contain an instance of an internal PersistentReference stub class everywhere O references P. Two PersistentReference instances compare equal if and only if they “represent” the same persistent object; when they’re not equal, they compare by memory address, and, as explained before, memory-based comparison must never happen in a sane persistent BTree. Note that it doesn’t help in this case if your Persistent subclass defines a sane __cmp__() method: conflict resolution doesn’t know about your class, and so also doesn’t know about its __cmp__() method. It only sees instances of the internal PersistentReference stub class.

Iteration and Mutation

As with a Python dictionary or list, you should not mutate a BTree-based data structure while iterating over it, except that it’s fine to replace the value associated with an existing key while iterating. You won’t create internal damage in the structure if you try to remove, or add new keys, while iterating, but the results are undefined and unpredictable. A weak attempt is made to raise RuntimeError if the size of a BTree-based structure changes while iterating, but it doesn’t catch most such cases, and is also unreliable. Example

>>> from BTrees.IIBTree import IISet
>>> s = IISet(range(10))
>>> list(s)
[0, 1, 2, 3, 4, 5, 6, 7, 8, 9]
>>> for i in s:  # the output is undefined
...     print(i)
...     s.remove(i)
0
2
4
6
8
Traceback (most recent call last):
  File "<stdin>", line 1, in ?
RuntimeError: the bucket being iterated changed size
>>> list(s)      # this output is also undefined
[1, 3, 5, 7, 9]
>>>

Also as with Python dictionaries and lists, the safe and predictable way to mutate a BTree-based structure while iterating over it is to iterate over a copy of the keys. Example

>>> from BTrees.IIBTree import IISet
>>> s = IISet(range(10))
>>> for i in list(s.keys()):  # this is well defined
...     print(i)
...     s.remove(i)
0
1
2
3
4
5
6
7
8
9
>>> list(s)
[]
>>>

BTree node sizes

BTrees (and TreeSets) are made up of a tree of Buckets (and Sets) and internal nodes. There are maximum sizes of these notes configured for the various key and value types (unsigned and quad unsigned follow integer and long, respectively):

Key Type

Value Type

Maximum Bucket or Set Size

Maximum BTree or TreeSet Size

Integer

Float

120

500

Integer

Integer

120

500

Integer

Object

60

500

Long

Float

120

500

Long

Long

120

500

Long

Object

60

500

Object

Integer

60

250

Object

Long

60

250

Object

Object

30

250

For your application, especially when using object keys or values, you may want to override the default sizes. You can do this by subclassing any of the BTree (or TreeSet) classes and specifying new values for max_leaf_size or max_internal_size in your subclass:

>>> import BTrees.OOBTree

>>> class MyBTree(BTrees.OOBTree.BTree):
...     max_leaf_size = 500
...     max_internal_size = 1000

As of version 4.9, you can also set these values directly on an existing BTree class if you wish to tune them across your entire application.

max_leaf_size is used for leaf nodes in a BTree, either Buckets or Sets. max_internal_size is used for internal nodes, either BTrees or TreeSets.

BTree Diagnostic Tools

A BTree (or TreeSet) is a complex data structure, really a graph of variable- size nodes, connected in multiple ways via three distinct kinds of C pointers. There are some tools available to help check internal consistency of a BTree as a whole.

Most generally useful is the check module. The check() function examines a BTree (or Bucket, Set, or TreeSet) for value-based consistency, such as that the keys are in strictly increasing order. See the function docstring for details. The display() function displays the internal structure of a BTree.

BTrees and TreeSets also have a _check() method. This verifies that the (possibly many) internal pointers in a BTree or TreeSet are mutually consistent, and raises AssertionError if they’re not.

If a check() or _check() call fails, it may point to a bug in the implementation of BTrees or conflict resolution, or may point to database corruption.

Repairing a damaged BTree is usually best done by making a copy of it. For example, if self.data is bound to a corrupted IOBTree,

>>> self.data = IOBTree(self.data)

usually suffices. If object identity needs to be preserved,

>>> acopy = IOBTree(self.data)
>>> self.data.clear()
>>> self.data.update(acopy)

does the same, but leaves self.data bound to the same object.