treelock

Fast read/write sub-tree locking for asyncio Python. Suitable for large trees, when it's not feasible or desired to have the entire tree in memory at once.

Inspired by the work of Ritik Malhotra.

Installation

pip install treelock

Usage

Each instance of TreeLock is callable, and returns an asynchronous context manager. In order to acquire a read (shared) lock on the sub-trees with root nodes in the iterable read_roots; and to acquire a write (exclusive) lock of the sub-trees with root nodes in the iterable write_roots, you must pass them to the instance of TreeLock:

from treelock import TreeLock

lock = TreeLock()

async def access(read_roots, write_roots):
  async with lock(read=read_roots, write=write_roots):
    # access the sub-trees

The lock is not re-entrant: the same task attempting to enter multiple context managers with incompatible sub-trees will deadlock. Hence the locks for all the required sub-trees must be requested up-front.

A typical use-case will be for read/write (shared/exclusive) locking of a path in a filesystem hierarchy. For example, if treating S3 as a filesystem, but allowing what-whould-be non-atomic operations on folders.

For example, you could define delete, write, rename, copy and read operations on folders at certain paths, e.g. instances of PurePosixPath. A read lock of such a path should allow reads of the corresponding folder, but block all operations that would change it. A write lock should prevent all other access to that folder. You can do this using TreeLock, noting that each path is in fact a node in the tree of all possible paths.

from treelock import TreeLock

lock = TreeLock()

async def delete(path):
  async with lock(read=[], write=[path]):
    ...

async def write(path, ...):
  async with lock(read=[], write=[path]):
    ...

async def rename(path_from, path_to):
  async with lock(read=[], write=[path_from, path_to]):
    ...

async def copy(path_from, path_to):
  async with lock(read=[path_from], write=[path_to]):
    ...

async def read(path):
  async with lock(read=[path], write=[]):
    ...

There is more information on this usage, as well as details of the underlying algorithm, at https://charemza.name/blog/posts/python/asyncio/s3-path-locking/.

Required properties of the nodes

These are a subset of the properties of PurePosixPath.

• Each defines the __cmp__ and __hash__ methods. These are used for dictionary and sets internally, so __hash__ must be reasonable enough to to acheive constant-time behaviour.

• Each must define the __lt__ method. This must be well-behaved, i.e. defines a total order between all possible nodes, otherwise deadlock can occur.

• Each has a property parents that is an iterator to the ancestors of the node, in any order. This is a slightly mis-named property, but this is consistent with PurePosixPath.

Note that a node does not need to be aware of its child nodes. This makes TreeLock suitable for locking sub-trees below a node without knowledge of the descendants of that node.

Fast locking and unlocking

The number of operations to lock or unlock a node only depends on the ancestors of a node. Specifically, it does not increase as the number of descendants increase, nor does it increase with the number of locks currently being held.

Running tests

python setup.py test