Result Caching

Stores function results so that they are not computed again on repetitive calls of the function with the same arguments. Results can be stored either on disk or in memory.

Quick setup

pip install git+https://github.com/mschrimpf/result_caching

Usage example

from result_caching import store

@store()
def f(a, b):
	return a * b
	
y = f(1, 2)  # computed first time, stored on disk
y = f(1, 2)  # not computed again, loaded from disk
y = f(1, 3)  # computed again, different parameters

By default, results will be stored in ~/.result_caching, this can be changed through the environment variable RESULTCACHING_HOME.

cache will only hold results in memory and not write them to disk.

Environment variables

Storage formats

pickle (the default) is fast but tied to the exact pandas/xarray/numpy versions that wrote it. netcdf is portable across environment upgrades, at the cost of flattening MultiIndex coordinates on write and rebuilding them on load.

The default is deliberately still pickle so that existing warm caches stay valid; opt in per environment rather than globally.

Either way an unreadable entry is treated as a miss, never an error — a cache written before a dependency bump is recomputed rather than raising.

S3 backend

Set RESULTCACHING_S3_BUCKET to make xarray results readable and writable across machines, which is what makes the cache useful for ephemeral containers where local disk does not survive the job.

Entries are always netCDF on S3 regardless of RESULTCACHING_FORMAT; the data object is written before its manifest and a read requires the manifest, so an interrupted write is indistinguishable from a miss and needs no temp-key dance (S3 PUTs are already atomic).

boto3 is an optional dependency. If it is missing, the backend degrades to "no cache" rather than breaking the import.