Bplustree

An on-disk B+tree for Python 3.

It feels like a dict, but stored on disk. When to use it?

When the data to store does not fit in memory
When the data needs to be persisted
When keeping the keys in order is important

This project is under development: the format of the file may change between versions. Do not use as your primary source of data.

Quickstart

Install Bplustree with pip:

pip install bplustree

Create a B+tree index stored on a file and use it with:

>>> from bplustree import BPlusTree
>>> tree = BPlusTree('/tmp/bplustree.db', order=50)
>>> tree[1] = b'foo'
>>> tree[2] = b'bar'
>>> tree[1]
b'foo'
>>> tree.get(3)
>>> tree.close()

Keys and values

Keys must have a natural order and must be serializable to bytes. Some default serializers for the most common types are provided. For example to index UUIDs:

>>> import uuid
>>> from bplustree import BPlusTree, UUIDSerializer
>>> tree = BPlusTree('/tmp/bplustree.db', serializer=UUIDSerializer(), key_size=16)
>>> tree.insert(uuid.uuid1(), b'foo')
>>> list(tree.keys())
[UUID('48f2553c-de23-4d20-95bf-6972a89f3bc0')]

Values on the other hand are always bytes. They can be of arbitrary length, the parameter value_size=128 defines the upper bound of value sizes that can be stored in the tree itself. Values exceeding this limit are stored in overflow pages. Each overflowing value occupies at least a full page.

Iterating

Since keys are kept in order, it is very efficient to retrieve elements in order:

>>> for i in tree:
...     print(i)
...
1
2
>>> for key, value in tree.items():
...     print(key, value)
...
1 b'foo'
2 b'bar'

It is also possible to iterate over a subset of the tree by giving a Python slice:

>>> for key, value in tree.items(slice(start=0, stop=10):
...     print(key, value)
...
1 b'foo'
2 b'bar'

Both methods use a generator so they don't require loading the whole content in memory, but copying a slice of the tree into a dict is also possible:

>>> tree[0:10]
{1: b'foo', 2: b'bar'}

Concurrency

The tree is thread-safe, it follows the multiple readers/single writer pattern.

It is safe to:

Share an instance of a BPlusTree between multiple threads

It is NOT safe to:

Share an instance of a BPlusTree between multiple processes
Create multiple instances of BPlusTree pointing to the same file

Durability

A write-ahead log (WAL) is used to ensure that the data is safe. All changes made to the tree are appended to the WAL and only merged into the tree in an operation called a checkpoint, usually when the tree is closed. This approach is heavily inspired by other databases like SQLite.

If tree doesn't get closed properly (power outage, process killed...) the WAL file is merged the next time the tree is opened.

Performances

Like any database, there are many knobs to finely tune the engine and get the best performance out of it:

order, or branching factor, defines how many entries each node will hold
page_size is the amount of bytes allocated to a node and the length of read and write operations. It is best to keep it close to the block size of the disk
cache_size to keep frequently used nodes at hand. Big caches prevent the expensive operation of creating Python objects from raw pages but use more memory

Some advices to efficiently use the tree:

Insert elements in ascending order if possible, prefer UUID v1 to UUID v4
Insert in batch with tree.batch_insert(iterator) instead of using tree.insert() in a loop
Let the tree iterate for you instead of using tree.get() in a loop
Use tree.checkpoint() from time to time if you insert a lot, this will prevent the WAL from growing unbounded
Use small keys and values, set their limit and overflow values accordingly
Store the file and WAL on a fast disk

License

MIT