clearml/clearml v2.2.0

PyPI v2.2.0 - ClearML

on GitHub

Bug fixes

• Fix nondeterministic type validation error messages in DataModel.assert_isinstance (#1693, thanks @marcus-campbell !)
• Fix cache cleanup evicting files currently served to another process (e.g. sibling DataLoader workers) by protecting cache hits with advisory file locks (#1737)
• Fix shell command argument escaping in the autoscaler cloud driver to prevent syntax errors and injections (#1704)
• Report CPU usage relative to the container's CPU limit (cgroup v2) when running inside a container, instead of machine-wide usage (#1736)

New Features and improvements

• Add sdk.storage.zip_compression configuration option to select the zip compression algorithm (stored, deflated, bzip2, lzma) used for dataset chunks, folder artifacts and offline session packages (#1733)
• Add DataView.to_list() to retrieve all data entries of a DataView as a list (#1734)
• Add CLEARML_API_RETRY_VALUES environment variable and vault support for overriding the API session retry settings at runtime (#1735)
• Add sdk.development.worker.log_stdout, log_stderr and resource_monitoring configuration options to disable stdout/stderr capture and resource monitoring; CLEARML_LOG_LEVEL now also accepts logging level names (#1703)
• Docstring updates for HPO, task scheduler, trigger scheduler, hyperdatasets, Task and HTTP router (#1705, #1706, #1707, #1708, #1709, #1710)
• Remove Python 2 compatibility shims and six usage from the SDK code (abstractproperty, python_2_unicode_compatible, abstract context manager, abstractclassmethod, six.moves, six.BytesIO/six.StringIO, PostImportHookPatching dead code) and no-op from __future__ imports (#1711 to #1720, #1738)