Caching for Analytic Computations

Humans repeat stuff. Caching helps.

Normal caching policies like LRU aren't well suited for analytic computations where both the cost of recomputation and the cost of storage routinely vary by one million or more. Consider the following computations

# Want thisnp.std(x) # tiny result, costly to recompute# Don't want thisnp.transpose(x) # huge result, cheap to recompute

Cachey tries to hold on to values that have the following characteristics

1. Expensive to recompute (in seconds)
2. Cheap to store (in bytes)
3. Frequently used
4. Recenty used

It accomplishes this by adding the following to each items score on each access

score += compute_time / num_bytes * (1 + eps) ** tick_time

For some small value of epsilon (which determines the memory halflife.) This has units of inverse bandwidth, has exponential decay of old results and roughly linear amplification of repeated results.

Example

>>>fromcacheyimportCache>>>c=Cache(1e9, 1) # 1 GB, cut off anything with cost 1 or less>>>c.put('x', 'some value', cost=3)
>>>c.put('y', 'other value', cost=2)
>>>c.get('x')
'some value'

This also has a memoize method

>>>memo_f=c.memoize(f)

Install

Cachey is on PyPI and Conda-forge:

$ pip install cachey # option 1
$ conda install cachey -c conda-forge # option 2

Or install from source

$ python setup.py install # option 1
$ pip install -e .# option 2 (best for development)

Status

Cachey is new and not robust.