You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Major Changes
This release completes the second phase of the performance overhaul of the commit
stage started in 2.5.0. The creation of the virtual datasets at the end of commit has
been rewritten in Cython; the full commit pipeline now revolves around the StagedChangesArray backend. delete_versions() and rebuild_hashtables() still use
the legacy code paths and will be overhauled in a future version.
modify_metadata() and recreate_dataset() now stream data from disk to disk and no
longer load entire versions of the dataset in memory; this enables using them when the
sum of non-empty chunks of a version is larger than memory.
Minor Changes
Fix ValueError: assignment destination is read-only after astype() or asarray() on a staged dataset.
Fixed crash in resize() followed by commit_staged_changes() when all staged chunks
are deduplicated away.
Fixed crash in resize() after assigning to a dataset with sv[name] = arr.
Fixed bugs when multiple h5py.File objects are open on the same file on disk
np.bytes_ scalars can now be assigned into variable-width string datasets.
Improved rejection of the reserved word versions used as a dataset path component.
modify_metadata() now preserves dataset attributes
Fixed crash in modify_metadata() with fixed-string datasets
Fixed bug where modify_metadata() on variable-width strings datasets would prevent
further versions from being staged.
Developers-only Changes
ASV benchmarks on Linux and macOS now track peak memory usage via memray
asv-compare no longer suppresses its own error messages