Upgrade to release 1.6.0:
- Fix caching of functions whose source cannot be retrieved, such
as functions defined in a notebook cell. Their identity fell back
to str(hash(func.__code__)), which is salted by PYTHONHASHSEED
and so differed between processes. A worker reading the
func_code.py written by another one concluded that the function
had changed and wiped the whole cache directory for it, discarding
results computed by its peers. func_code.py is also no longer
rewritten in place, so a reader can no longer catch it
half-written and draw the same conclusion.
- Drop python 3.9 support. The oldest supported Python version is
now Python 3.10.
- Fix eval_expr (used to evaluate the pre_dispatch argument of
Parallel) to raise a ValueError as documented instead of leaking
a ZeroDivisionError for expressions that divide or take a modulo
by zero.
- MemorizedResult now forwards mmap_mode to its store backend, so a
cached array reconstructed from a location is memory-mapped as
requested instead of being loaded fully into memory.
- Unvendor cloudpickle to more quickly benefit from maintenance
releases of cloudpickle
- Fix Memory.cache for functions with a keyword-only argument that
has a default declared before a keyword-only argument without a
default.
- Fix behavior of filter_args on some precise cases.
- Fix a concurrency error that could happen with unordered
generator.
- Fix: dump() now accepts any input os.PathLike object to be
consistent with load.
- The documentation now uses pydata sphinx theme. Furthermore,
optional dependencies test and docs have been added to
pyproject.toml.
Extend RDEPENDS with the missing runtime dependencies:
python3-cloudpickle.
Signed-off-by: Leon Anavi <leon.anavi@konsulko.com>
Signed-off-by: Khem Raj <khem.raj@oss.qualcomm.com>
