ZODB-friendly BTree-based list implementation
The sequence in this package has a list-like API, but stores its values in individual buckets. This means that, for small changes in large sequences, the sequence could be a big win. For instance, an ordered BTree-based container might want to store order in a sequence, so that moves only cause a bucket or two–around 50 strings or less–to be rewritten in the database, rather than the entire contents (which might be thousands of strings, for instance).
If the sequence is most often completely rearranged, the complexity of the code in this package is not desirable. It only makes sense if changes most frequently are fairly small.
One downside is that reading and writing is more work than with a normal list. If this were to actually gain traction, perhaps writing some or all of it in C would be helpful. However, it still seems pretty snappy.
Another downside is the corollary of the bucket advantage listed initially: with more persistent objects, iterating over it will fill a lot of ZODB’s object cache (which is based on the number of objects cached, rather than the size). Consider specifying a big object cache if you are using these to store a lot of data and are frequently iterating or changing.
These sequences return slices as iterators, and add some helpful iteration methods. It adds a copy method that provides a cheap copy of the blist that shares all buckets and indexes until a write happens, at which point it copies and mutates the affected indexes and buckets.
We’ll take a glance at how these differences work, and then describe the implementation’s basic mechanism, and close with a brief discussion of performance characteristics in the abstract.
This doesn’t need much discussion. Getting slices of all sorts returns iterators.
>>> from zc.blist import BList >>> l = BList(range(1000)) >>> l[345:351] # doctest: +ELLIPSIS <generator object at ...> >>> list(l[345:351]) [345, 346, 347, 348, 349, 350]>>> l[351:345:-1] # doctest: +ELLIPSIS <generator object at ...> >>> list(l[351:345:-1]) [351, 350, 349, 348, 347, 346]>>> l[345:351:2] # doctest: +ELLIPSIS <generator object at ...> >>> list(l[345:351:2]) [345, 347, 349]
iterReversed lets you iterate over the list in reverse order, efficiently, with a given start point. It is used for slices that proceed with a step of -1.
>>> i = l.iterReversed() >>> i.next() 999 >>> i.next() 998 >>> list(i)[-10:] [9, 8, 7, 6, 5, 4, 3, 2, 1, 0]
iterSlice lets you iterate over the list with a slice. It is equivalent to using a slice with __getitem__.
>>> i = l.iterSlice(345, 351, 2) >>> i # doctest: +ELLIPSIS <generator object at ...> >>> list(i) [345, 347, 349]
The copy method produces a cheap copy of the given blist. All buckets and indexes are shared until a change is made to either side. Copies can safely be made of other copies.
>>> c = l.copy() >>> l == c True >>> list(c) == list(l) True >>> del c[10:] >>> list(l) == range(1000) True >>> list(c) [0, 1, 2, 3, 4, 5, 6, 7, 8, 9] >>> l == c False >>> c2 = c.copy() >>> c2 == c True >>> list(c2) [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]
In its implementation, the sequence is an adapted B+ tree. Indexes are keys, but each bucket or branch starts at 0. For instance, a perfectly-balanced bucket sequence with 16 items, and a limit of 3 entries in a bucket or branch, would have “keys” like this. In the diagram, the top three rows are indexes, and the bottom row consists of buckets:
0 8 0 4 0 4 0 2 0 2 0 2 0 2 01 01 01 01 01 01 01 01
So, for instance, you would get the value at position 5 using this process:
- In the top index (the top row, with keys of 0 and 8), find the largest key that is lower than the desired position, and use the associated value (index or bucket, which is in this case the index represented by the first pair of 0 and 4 in the second row) for the next step. In this case, the top index has keys of 0 and 8, so the largest key lower than position 5 is 0. Subtract this key from the position for the next step. This difference will be used as the position for the next step. In this case, the next position will be (5-0=) 5.
- The next index has keys of 0 and 4. The largest key lower than 5 is 4. Use the child index associated with the 4 key for the next step (the second pair of 0 and 2 in the third row), and subtract the key (4) from the position (5) for the position to be used in the next step (=1).
- The next index (the second pair of 0 and 2 in the third row) needs to find position 1. This will return the third pair of 0 1 in the last row. The new position will be (1-0=) 1.
- Finally, position 1 in the bottom bucket stores the actual desired value.
This arrangement minimizes the changes to keys necessary when a new value is inserted low in the sequence: ignoring balancing the tree, only parents and their subsequent siblings must be adjusted. For instance, inserting a new value in the 0 position of the bucketsequence described above (the worst case for the algorithm, in terms of the number of objects touched) would result in the following tree:
0 9 0 5 0 4 0 3 0 2 0 2 0 2 012 01 01 01 01 01 01 01
- __getitem__ is efficient, not loading unnecessary buckets. It handles slices pretty well too, not even loading intermediary buckets if the slice is very large. Slices currently return iterables rather than lists; this may switch to a view of some sort. All that should be assumed right now is that you can iterate over the result of a slice.
- __setitem__ and all the write methods do a pretty good job in terms of efficient loading of buckets, and only writing what they need to. It supports full Python slice semantics.
- copy is cheap: it reuses buckets and indexes so that new inner components are created lazily when they mutate.
- While __contains__, __iter__, index and other methods are brute force and written in Python, they might not load all buckets and items, while with a normal list or tuple, they always will. See also iter, iterReversed, and iterSlice.
- count, __eq__, and other methods load all buckets and items, and are brute force, and in Python. In contrast, lists and tuples will load all items (worse), and is brute force in C (better, but not algorithmically).
- This will create a lot of Persistent objects for one blist, which may cause cache eviction problems depending on circumstances and usage.
- Did I mention that this was in Python, not C? That’s fixable, at least, and in fact doesn’t appear to be too problematic at the moment, at least for the author’s usage.
- fixed: internal data structures were not stored correctly in the ZODB, so BLists loaded from a fresh DB connection would break.
- removed unused code and the dependency on rwproperty
- improved test coverage of the BList API, fixed access to items at index -1 or at the ends of the valid index range