percache is a Python module to persistently cache results of functions (or callables in general) using decorators.

It is somehow similar to the Memoize Example from the Python Decorator Library but with the advantage that results are stored persistently in a cache. percache provides memoization across multiple invocations of the Python interpreter.

Install with pip install percache. percache works with Python 2.6, 2.7, and 3.3 and has no dependencies outside the standard library.

>>> import percache
>>> cache = percache.Cache("/tmp/my-cache")
>>>
>>> @cache
... def longtask(a, b):
...     print("running a long task")
...     return a + b
...
>>> longtask(1, 2)
running a long task
3
>>>
>>> longtask(1, 2)
3
>>> cache.close() # writes new cached results to disk

As you can see at the missing output after the second invocation, longtask has been called once only. The second time the result is retrieved from the cache. The key feature of this module is that this works across multiple invocations of the Python interpreter.

A requirement on the results to cache is that they are pickable.

Each cache file can be used for any number of differently named callables.

By default percache uses a shelve as its cache back-end. Alternative back-ends may be used if they are given as dictionary-like objects with a close() and sync() method:

>>> class FooCache(dict):
...     def sync(self):
...         ...
...     def close(self):
...         ...
>>> fc = FooCache()
>>> cache = percache.Cache(fc, livesync=True)

In this example a cache is created in live-sync mode, i.e. results immediately are stored permanently. Normally this happens not until a cache's close() method has been called or until it gets finalized. Note that the live-sync mode may slow down your percache-decorated functions (though it reduces the risk of "loosing" results).

When caching the result of a callable, a SHA1 hash based on the callable's name and arguments is used as a key to store the result in the cache file.

The hash calculation does not work directly with the arguments but with their representations, i.e. the string returned by applying repr(). Argument representations are supposed to differentiate values sufficiently for the purpose of the function but identically across multiple invocations of the Python interpreter. By default the built-in function repr() is used to get argument representations. This is just perfect for basic types, lists, tuples and combinations of them but it may fail on other types:

>>> repr(42)
42                                  # good
>>> repr(["a", "b", (1, 2L)])
"['a', 'b', (1, 2L)]"               # good
>>> o = object()
>>> repr(o)
'<object object at 0xb769a4f8>'     # bad (address is dynamic)
>>> repr({"a":1,"b":2,"d":4,"c":3})
"{'a': 1, 'c': 3, 'b': 2, 'd': 4}"  # bad (order may change)
>>> class A(object):
...     def __init__(self, a):
...         self.a = a
...
>>> repr(A(36))
'<__main__.A object at 0xb725bb6c>' # bad (A.a not considered)
>>> repr(A(35))
'<__main__.A object at 0xb725bb6c>' # bad (A.a not considered)

A bad representation is one that is not identically across Python invocations (all bad examples) or one that does not differentiate values sufficiently (last 2 bad examples).

To use such types anyway you can either

1. implement the type's __repr__() method accordingly or
2. provide a custom representation function using the repr keyword of the Cache constructor.

To pass dictionaries to percache decorated functions, you could wrap them in an own dictionary type with a suitable __repr__() method:

>>> class mydict(dict):
...     def __repr__(self):
...         items = ["%r: %r" % (k, self[k]) for k in sorted(self)]
...         return "{%s}" % ", ".join(items)
...
>>> repr(mydict({"a":1,"b":2,"d":4,"c":3}))
"{'a': 1, 'b': 2, 'c': 3, 'd': 4}"  # good (always same order)

The following example shows how to use a custom representation function to get a suitable argument representation of file objects:

>>> def myrepr(arg):
...     if isinstance(arg, file):
...         # return a string with file name and modification time
...         return "%s:%s" % (arg.name, os.fstat(arg.fileno())[8])
...     else:
...         return repr(arg)
...
>>> cache = percache.Cache("/some/path", repr=myrepr)

• Make sure to delete the cache file whenever the behavior of a cached function has changed!
• To prevent the cache from getting larger and larger you can call the clear() method of a Cache instance. By default it clears all results from the cache. The keyword maxage my be used to specify a maximum number of seconds passed since a cached result has been used the last time. Any result not used (written or accessed) for maxage seconds gets removed from the cache.

• Support Python 3.3 (next to 2.6 and 2.7)

• Add missing README to PyPi package.

• Automatically close (i.e. sync) the cache on finalization.
• Optionally sync the cache on each change.
• Support for alternative back-ends (others than shelve).
• Cache object are callable now, which makes the explicit check() method obsolete (though the old interface is still supported).

• Fix wrong usage age output of command line interface.
• Meet half way with pylint.

• Initial release