Joblib provides a set of tools for using long-running Python functions as pipeline jobs:
- transparent disk-caching of the output values and lazy re-evaluation
- logging and tracing of the execution
The original focus was on scientific-computing scripts, but any long-running succession of operations can profit from the tools provided by joblib.
joblib is BSD-licensed.
Joblib came out of long-running data-analysis Python scripts. The long term vision is to provide tools for scientists to achieve better reproducibility when running jobs, without changing the way numerical code looks like. However, Joblib can also be used to provide a light-weight make replacement.
The main problems identified are:
- Lazy evaluation: People need to rerun over and over the same script as it is tuned, but end up commenting out steps, or uncommenting steps, as they are needed, as they take long to run.
- Persistance: It is difficult to persist in an efficient way arbitrary objects containing large numpy arrays. In addition, hand-written persistence to disk does not link easily the file on disk to the corresponding Python object it was persists from in the script. This leads to people not a having a hard time resuming the job, eg after a crash and persistence getting in the way of work.
The approach take by Joblib to address these problems is not to build a heavy framework and coerce user into using it (e.g. with an explicit pipeline). It strives to leave your code and your flow control as unmodified as possible.
The tools that have been identified and developped so far are:
Transparent and fast disk-caching of output value: a make-like functionality for Python functions that works well with large numpy arrays. The goal is to separate operations in a set of steps with well-defined inputs and outputs, that are saved and reran only if necessary, by using standard Python functions:
>>> from joblib import Memory >>> mem = Memory(cachedir='/tmp/joblib') >>> import numpy as np >>> a = np.vander(np.arange(3)) >>> square = mem.cache(np.square) >>> b = square(a) ________________________________________________________________________________ [Memory] Calling square... square(array([[0, 0, 1], [1, 1, 1], [4, 2, 1]])) ___________________________________________________________square - 0.0s, 0.0min >>> c = square(a) >>> # The above call did not trigger an evaluationLogging/tracing: The functionalities described above will progressively acquire better logging mechanism to help track what has been ran, and capture I/O easily. In addition, Joblib will provide a few I/O primitives, to easily define define logging and display streams, and maybe provide a way of compiling a report. In the long run, we would like to be able to quickly inspect what has been run.
As stated on the project page, currently the project is in alpha quality. I am testing heavily all the features, as I care more about robustness than having plenty of features. On the other side, I expect to be playing with the API and features for a while before I can figure out what is the right set of functionalities to expose.
The code is hosted on github for the good reason that cloning the project and publishing it along-side my branch is dead-easy.
| Memory | A context object for caching a function’s return value each time it is called with the same input arguments. |
| dump (value, filename) | Persist a Python into a filename, with numpy arrays saved as separate .npy files. |
| load (filename[, mmap_mode]) | Reconstruct a Python object and the numpy arrays it contains from a persisted file. |
| PrintTime | Print and log messages while keeping track of time. |
| Logger | Base class for logging messages. |
| hash (obj[, hash_name, coerce_mmap]) | Quick calculation of a hash to identify uniquely Python objects containing numpy arrays. |