Same as test_delete_versions_invalidates_held_group, for a dataset that was obtained directly.

A version group obtained before delete_versions() must not keep reading from the stale chunk offsets afterwards.

All VersionedHDF5File handles on a file must be invalidated by delete_versions(), not just the one it was called on.

raw_data must not retain any data from the deleted versions in the padding around the chunks that lie on the edge of the virtual dataset. This inc

Test that delete_versions only needs linear time to find the previous version for the versions that are being kept.

Test that __setitem__ loads from the base slabs only the chunks that are partially selected by the index

In StagedChangesArray.from_array, dtype from base array takes precedence over the one from fill_value

In StagedChangesArray.full, explicitly declared dtype takes precedence over the one from fill_value. However, if omitted, always respect fill_valu

In StagedChangesArray.__init__, dtype from base slabs takes precedence over the one from fill_value

A chunk shorter than the physical slab width must hash only its valid region, never the (possibly garbage) memory past the edge.

A chunk with a zero-length axis hashes just its (empty) shape string.

Test .compression, .compression_opts, .shuffle, and .fletcher32 after DatasetWrapper hot-swaps InMemoryDataset -> InMemoryArrayDataset -> InMe

from_array(..., as_base_slabs=False) Input array is exactly divisible by chunk_size. Slabs are writeable views of the original array.

from_array(..., as_base_slabs=False) Input array is not exactly divisible by chunk_size. All slabs are views of the original array, including

from_array(..., as_base_slabs=False) Only some axes are not exactly divisible by chunk_size.

from_array(..., as_base_slabs=False) Input array is too small to fully cover even a single chunk; the slab is a trimmed writeable view of the

from_array(..., as_base_slabs=False) Input array is a read-only NumPy array. It is CoW-copied on first write.

from_array(..., as_base_slabs=False) Input array is an ArrayProtocol.

astype() lazily converts trimmed staged slabs.

changes() yields edge chunks trimmed to the exact edge size.

commit() hashes, deduplicates, and consolidates trimmed staged slabs.

copy() CoW-deep-copies trimmed staged slabs upon first write access.

rechunk() works on trimmed staged slabs.

refill() rewrites fill_value points on trimmed staged slabs in place.

Enlarging axis 0 while shrinking axis 1 relocates the trimmed last chunks and drops the slabs left unreferenced by the shrink.

Enlarging axis 0 relocates the only chunk of single-chunk trimmed slabs, which drops them; the drop is re-derived from scratch together with the d

Enlarging only along axis 0 relocates trimmed last chunks instead of deep-copying whole slabs.

Full edge chunks living on trimmed staged slabs are filled in place. After shrinking (3, 5) -> (2, 3), the trimmed slabs still span 3 rows but th

resize() deep-copies trimmed staged slabs only when enlarging beyond a trimmed edge; a shrink or a no-op resize leaves them as views. Enlarging af

Relocating trimmed staged slabs also truncates their hash tables, if present.

from_array(..., as_base_slabs=False) Edge slabs are trimmed to the exact edge size, never padded.

_untrim_staged_slabs() does nothing when only axis 0 is being enlarged. Slabs trimmed along axis 0 are left for ResizePlan, which relocates only

hash_slab works on a read-only broadcasted array (the StagedChangesArray full slab) and matches the contiguous materialization.

Check that the diff between two versions returns the expected chunks.

Check that the diff between a version and itself is nothing.

__getitem__ changes chunk dtype upon first access.

Test that InMemoryGroup supports dictionary-like accessors.

Test that InMemoryGroup.get() returns the same as __getitem__ for existing keys.

Test that repr(InMemoryGroup) also shows reprs of child objects.

Test that we're not calling hashlib.hash.digest() on NpyStrings, which could cause hash collisions: https://github.com/numpy/numpy/issues/2922

Test that NpyStrings and object strings produce identical hashes.

Test that chunks in StagedChangesArray have not been converted to object dtype, which means they will reach the hashing function intact

hash_tables stays the same length as slabs, with None wherever a slab is None.

Random tiled chunks carved out of a randomly strided slab.

Test we correctly handle inserting into a multi-dimensional Dataset and shift the existing entries back.

The dtype is not part of the digest, so two chunks of the same shape with the same raw byte image collide even when they mean entirely different t

Check that creating a dataset before making it empty can be done successfully. This test would pass unless the file gets closed/reopened for each

Check that creating a multidimensional dataset before making it empty can be done successfully. See https://github.com/deshaw/versioned-hdf5/

Test that versioned_hdf5 behaves like h5py when it comes to NpyStrings UX.

Test that setting compression via modify_metadata works for third-party compression filters from hdf5plugin or pytables, such as Blosc.

Test changing compression options of a custom compression filter, using hdf5plugin API.

modify_metadata on a dataset whose wrapper carries chunks=None. A dataset assigned with ``sv[name] = array`` is backed by an InMemoryArrayDataset

modify_metadata(chunks=True) on a dense dataset created with explicit chunks. modify_metadata reuses the cached original InMemoryArrayDataset; it

Test that setting compression via modify_metadata works for default compression.