Simple dictionary with LRU behaviour

## LRU Dictionaries

>>> from darts.lib.utils.lru import LRUDict


An LRUDict is basically a simple dictionary, which has a defined maximum capacity, that may be supplied at construction time, or modified at run-time via the capacity property:

>>> cache = LRUDict(1)
>>> cache.capacity
1

The minimum capacity value is 1, and LRU dicts will complain, if someone attempts to use a value smaller than that:

>>> cache.capacity = -1                              #doctest: +ELLIPSIS
Traceback (most recent call last):
...
ValueError: -1 is not a valid capacity
>>> LRUDict(-1)                                      #doctest: +ELLIPSIS
Traceback (most recent call last):
...
ValueError: -1 is not a valid capacity

LRU dictionaries can never contain more elements than their capacity value indicates, so:

>>> cache[1] = "First"
>>> cache[2] = "Second"
>>> len(cache)
1

In order to ensure this behaviour, the dictionary will evict entries if it needs to make room for new ones. So:

>>> 1 in cache
False
>>> 2 in cache
True

The capacity can be adjusted at run-time. Growing the capacity does not affect the number of elements present in an LRU dictionary:

>>> cache.capacity = 3
>>> len(cache)
1
>>> cache[1] = "First"
>>> cache[3] = "Third"
>>> len(cache)
3

but shrinking does:

>>> cache.capacity = 2
>>> len(cache)
2
>>> sorted(list(cache.iterkeys()))
[1, 3]

Note, that the entry with key 2 was evicted, because it was the oldest entry at the time of the modification of capacity. The new oldest entry is the one with key 1, which can be seen, when we try to add another entry to the dict:

>>> cache[4] = "Fourth"
>>> sorted(list(cache.iterkeys()))
[3, 4]

The following operations affect an entry’s priority:

- get
- __getitem__
- __setitem__
- __contains__

Calling any of these operations on an existing key will boost the key’s priority, making it more unlikely to get evicted, when the dictionary needs to make room for new entries. There is a special peek operation, which returns the current value associated to a key without boosting the priority of the entry:

>>> cache.peek(3)
'Third'
>>> cache[5] = "Fifth"
>>> sorted(list(cache.iterkeys()))
[4, 5]

As you can see, even though we accessed the entry with key 3 as the last one, the entry is now gone, because it did not get a priority boost from the call to peek.

The class LRUDict supports a subset of the standard Python dict interface. In particular, we can iterate over the key, values, and items of an LRU dict:

>>> sorted([k for k in cache.iterkeys()])
[4, 5]
>>> sorted([v for v in cache.itervalues()])
['Fifth', 'Fourth']
>>> sorted([p for p in cache.iteritems()])
[(4, 'Fourth'), (5, 'Fifth')]
>>> sorted(list(cache))
[4, 5]

Note, that there is no guaranteed order; in particular, the elements are not generated in priority order or somesuch. Similar to regular dicts, an LRU dict’s __iter__ is actually any alias for iterkeys.

Furthermore, we can remove all elements from the dict:

>>> cache.clear()
>>> sorted(list(cache.iterkeys()))
[]


Instances of class LRUDict are not thread safe. Worse: even concurrent read-only access is not thread-safe and has to be synchronized by the client application.

There is, however, the class SynchronizedLRUDict, which exposes the same interface as plain LRUDict, but fully thread-safe. The following session contains exactly the steps, we already tried with a plain LRUDict, but now using the synchronized version:

>>> from darts.lib.utils.lru import SynchronizedLRUDict
>>> cache = SynchronizedLRUDict(1)
>>> cache.capacity
1
>>> cache.capacity = -1                              #doctest: +ELLIPSIS
Traceback (most recent call last):
...
ValueError: -1 is not a valid capacity
>>> LRUDict(-1)                                      #doctest: +ELLIPSIS
Traceback (most recent call last):
...
ValueError: -1 is not a valid capacity
>>> cache[1] = "First"
>>> cache[2] = "Second"
>>> len(cache)
1
>>> 1 in cache
False
>>> 2 in cache
True
>>> cache.capacity = 3
>>> len(cache)
1
>>> cache[1] = "First"
>>> cache[3] = "Third"
>>> len(cache)
3
>>> cache.capacity = 2
>>> len(cache)
2
>>> sorted(list(cache.iterkeys()))
[1, 3]
>>> cache[4] = "Fourth"
>>> sorted(list(cache.iterkeys()))
[3, 4]
>>> cache.peek(3)
'Third'
>>> cache[5] = "Fifth"
>>> sorted(list(cache.iterkeys()))
[4, 5]
>>> sorted([k for k in cache.iterkeys()])
[4, 5]
>>> sorted([v for v in cache.itervalues()])
['Fifth', 'Fourth']
>>> sorted([p for p in cache.iteritems()])
[(4, 'Fourth'), (5, 'Fifth')]
>>> sorted(list(cache))
[4, 5]
>>> cache.clear()
>>> sorted(list(cache.iterkeys()))
[]

Having some kind of dictionary which is capable of cleaning itself up is nice, but in order to implement caching, there is still something missing: the mechanism, which actually loads something into our dict. This part of the story is implemented by the AutoLRUCache:

>>> from darts.lib.utils.lru import AutoLRUCache

Let’s first define a load function:

>>> def load_resource(key):
...    if key < 10:
...        return "R(%s)" % (key,)

and a cache:

>>> cache = AutoLRUCache(load_resource, capacity=3)
'R(1)'
'R(1)'

As you can see, the first time, an actual element is loaded, the load function provided to the constructor is called, in order to provide the actual resource value. On subsequent calls to load, the cached value is returned.

Internally, the AutoLRUCache class uses an LRUDict to cache values, so:

>>> cache.load(2)
'R(2)'
'R(3)'
'R(4)'
'R(1)'

Note the “Loading 1” line in the last example. The cache has been initialized with a capacity of 3, so the value of key 1 had to be evicted when the one for key 4 was loaded. When we tried to obtain 1 again, the cache had to properly reload it, calling the loader function.

If there is actually no resource for a given key value, the loader function must return None. It follows, that None is never a valid resource value to be associated with some key in an AutoLRUCache.

>>> cache.load(11, 'Oops')
'Oops'


Instances of class AutoLRUCache are fully thread safe. Be warned, though, that the loader function is called outside of any synchronization scope the class may internally use, and has to provide its own synchronization if required.

The cache class actually tries to minimize the number of invocations of the loader by making sure, that no two concurrent threads will try to load the same key value (though any number of concurrent threads might be busy loading the resources associated with different keys).

## Caching and stale entries

There is another AutoLRUCache-like class provided by the LRU module, which gives more control over timing out of entries than AutoLRUCache does.

>>> from darts.lib.utils.lru import DecayingLRUCache
>>> current_time = 0
>>> def tick():
...     global current_time
...     current_time += 1


Here, we defined a simple “clock”. We could have used the system clock, but roling our own here gives us more control over the notion of “time”. Now, let’s define a simple cache entry:

>>> from collections import namedtuple
>>> Entry = namedtuple("Entry", "timestamp payload")


>>> def load(full_key):
...     return make_entry(u"Entry for %r" % (full_key,))


For the following parts, we consider an entry to be “too old”, if it has been created more then two “ticks” ago:

>>> def is_still_current(entry):
...    return current_time - entry.timestamp <= 2


Finally, we create another cache thingy

>>> cache = DecayingLRUCache(load, tester=is_still_current, capacity=3)


The DecayingLRUCache shows much of the same behaviour of the AutoLRUCache, namely:

>>> cache.load(1)


The entry with key 1 had to be reloaded, since the cache has a capacity of 3, and the old entry for 1 was evicted when the entry for 4 was loaded and we needed to make room.

>>> cache.load(3)


>>> tick()


The entry is still available.

>>> tick()
>>> tick()


Note, that eviction is still based on LRU, not on the age test.

## Change Log

### Version 0.5

Added a “from __future__ import with_statement” for Python 2.5 compatibility. Note, that supporting py2.5 is not a real goal, and I did not test the code using that version.