This issue collects the results of a Claude-assisted review of src/zarr/ (at 86dabd5) covering four dimensions: bugs, performance, simplifications, and modernizations. Each finding was verified against the code by the reviewing agent before inclusion. Draft PRs addressing groups of these findings reference this issue.

Note: This issue lives on the d-v-b fork so the findings and draft PRs can be triaged before anything is proposed upstream.

Draft PRs

Not actioned in a PR: S7 (medium-effort overload design), S9 (blocked on create() deprecation), M6/M7 (ruff rule enablement — judgement calls), M8 (needs a decision on intent), and P6 (refuted, see below).

🐛 Bugs

Storage & codec pipeline

• B1 — MemoryStore returns wrong bytes for SuffixByteRequest when suffix exceeds object size (src/zarr/storage/_utils.py:155, medium). _normalize_byte_range_index computes start = len(data) - suffix with no lower clamp; a negative start is interpreted by numpy slicing as an offset from the end, silently returning too few bytes. For a 5-byte value, suffix=7 returns 2 bytes while suffix=10 returns all 5 — non-monotonic, and inconsistent with ZipStore/LocalStore which clamp with max(0, ...). Fix: clamp start = max(0, len(data) - suffix).
• B2 — LoggingStore.get_partial_values exhausts a generator argument before forwarding it (src/zarr/storage/_logging.py:182, medium). The log line ",".join([k[0] for k in key_ranges]) fully consumes a one-shot iterator, so the wrapped store receives an empty iterable and returns []. Fix: materialize key_ranges = list(key_ranges) once.
• B3 — getsize_prefix over-counts on MemoryStore/ObjectStore due to substring prefix matching (src/zarr/abc/store.py:539 + src/zarr/storage/_memory.py:202, medium). Store.getsize_prefix calls list_prefix(prefix) without a trailing /, and the substring-matching stores then also count sibling nodes like foobar/... when asked about foo. Affects Array.nbytes_stored. Fix: normalize the prefix with a trailing / in getsize_prefix (matching delete_dir/is_empty/clear).
• B4 — ZipStore.close() raises AttributeError if the store was never opened (src/zarr/storage/_zip.py:121, low). self._lock is only created in _sync_open(), but close() unconditionally takes it, so with ZipStore(path): pass raises. Fix: guard close() on self._is_open or create the lock in __init__.
• B5 — codecs_from_list builds an error message for invalid BytesBytes-after-ArrayArray ordering but never raises it (src/zarr/core/codec_pipeline.py:675, low). Sibling branches raise TypeError(msg); this branch assigns msg and falls through, silently accepting a spec-invalid codec order. Fix: add the missing raise TypeError(msg).

Core

• B6 — DefaultChunkKeyEncoding.decode_chunk_key does not round-trip with encode_chunk_key (src/zarr/core/chunk_key_encodings.py:79, low). encode_chunk_key((0, 1)) → "c/0/1", but decode_chunk_key strips only the "c" character, leaving a leading separator, so int('') raises ValueError for every non-empty coordinate key. Fix: drop the "c" token after splitting instead of slicing one character.
• B7 — Attribute updates mutate frozen-dataclass metadata in place (src/zarr/core/array.py:6011, src/zarr/core/group.py:1219, low). _update_attributes mutates metadata.attributes in place even though the metadata classes are frozen and hash by to_dict(), so an instance's hash can change after construction. Fix: build a new attributes dict and use dataclasses.replace, as update_shape already does.
• B8 — Attributes.put clears live metadata before persisting (non-atomic despite docstring) (src/zarr/core/attributes.py:53, low). put empties the in-memory attributes dict before the store write; if save_metadata raises, memory and store disagree. Fix: compute the replacement dict first and apply it with a replace-based update. (Shares root cause with B7.)

🚀 Performance

• P1 — ShardingCodec rebuilds codec pipelines on every shard access (src/zarr/codecs/sharding.py:367, high impact). codec_pipeline is a plain @property calling get_pipeline_class().from_codecs(...) per shard, and _shard_index_size/_decode_shard_index/_encode_shard_index each rebuild a fresh index pipeline per shard. Fix: cache on the frozen codec instance (cached_property/instance-keyed lru_cache).
• P2 — default_buffer_prototype() reconstructed per inner chunk inside shards (src/zarr/codecs/sharding.py:105, medium). A debug assert prototype == default_buffer_prototype() runs config + registry lookups for every inner-chunk fetch on the hot read path. Fix: hoist or drop the assertion.
• P3 — Per-shard rebuild of the inner chunk grid + indexer, identical across shards (src/zarr/codecs/sharding.py:429,561, medium). Each shard decode/encode reconstructs ChunkGrid.from_sizes, a BasicIndexer, and its materialized projection list, though they're identical for every shard. Fix: memoize keyed by (shard_shape, chunk_shape).
• P4 — Repeated donfig config.get("async.concurrency") lookups on the per-chunk codec path (src/zarr/core/codec_pipeline.py:398,508,574,607,626, src/zarr/abc/codec.py:252,289,496, medium-low). With the default batch_size=1, each chunk incurs ~4 donfig dict traversals that cannot change mid-operation. Fix: read once per read/write call and thread through.
• P5 — Deep group traversal serializes across sibling subgroups (src/zarr/core/group.py:3322, medium). _iter_members_deep drains each subgroup iterator sequentially, so latency does not overlap across branches on high-latency stores. Fix: fan out sibling iterators concurrently under the existing semaphore.
• P6 — StorePath.__truediv__ re-normalizes already-normalized paths per chunk — refuted during implementation: __truediv__ calls _join_paths (a plain "/".join), not normalize_path; the join is already cheap. No action.

🧹 Simplifications

• S1 — Duplicate overwrite-check logic in _create_v2/_create_v3/_create_array_metadata (src/zarr/core/array.py:579,660,4404). Three copies of the overwrite/supports_deletes dance; extract one helper.
• S2 — Module-level _get_orthogonal_selection/_get_mask_selection/_get_coordinate_selection are single-caller pass-through layers (src/zarr/core/array.py:5526-5695). ~110 net lines removable by inlining indexer construction into the AsyncArray.get_* methods (the sync Array already bypasses them).
• S3 — Repeated zarr_format == 2 ? metadata.dtype : metadata.data_type inline dispatch (src/zarr/core/array.py:5393,5739,968). Add a unifying zdtype property on both metadata classes.
• S4 — Array.cdata_shape and Array._chunk_grid_shape duplicate the same delegation (src/zarr/core/array.py:2162-2182).
• S5 — "Not yet implemented" warning boilerplate repeated in open_group and create (src/zarr/api/asynchronous.py:816,1013). Compress to one helper iterating {name: value}.
• S6 — V3 branches of contains_array/contains_group duplicate _contains_node_v3 (src/zarr/storage/_common.py:591-661). Rewrite as (await _contains_node_v3(sp)) == "array" / "group". Implementation note: only contains_group was rewritten — contains_array raises ValueError (rather than returning False) when the node is a group, so the rewrite would change observable behavior.
• S7 — parse_zarr_format defined three times (metadata/v2.py:280, metadata/v3.py:49, group.py:86). Unify with overloads for literal narrowing (medium effort).
• S8 — default_filters_v3 is a trivially inlinable no-op (src/zarr/core/array.py:4836). Body is return () with one caller.
• S9 — AsyncArray._create/_create_v2/_create_v3 trio (~140 lines) exists solely for the deprecated create() path (src/zarr/core/array.py:368-696). Removable when the legacy API is removed — tracked here, not actionable yet.
• S10 — *_like functions re-check isinstance after _like_args already did (src/zarr/api/asynchronous.py:139,1091-1313). Fold fill_value extraction into _like_args.

✨ Modernizations (Python ≥ 3.12)

• M1 — Hand-rolled batched() duplicates itertools.batched (src/zarr/core/codec_pipeline.py:45-50). Delete and use the stdlib version (3.12+).
• M2 — typing_extensions.Buffer → collections.abc.Buffer (src/zarr/codecs/crc32c_.py:8,45,69). Stdlib since 3.12.
• M3 — Unused P = ParamSpec("P") (src/zarr/core/sync.py:22). Defined, never referenced; delete.
• M4 — Hand-rolled product()/reduce(operator.mul, ...)/np.prod() on int tuples → math.prod() (src/zarr/core/common.py:85, src/zarr/core/indexing.py:1190, src/zarr/core/array.py:908, src/zarr/core/chunk_grids.py:855). math.prod avoids numpy scalar round-trips; keep product = math.prod shim for existing importers. Implementation note: chunk_grids.py:689,704,711 operate on a numpy float array (not int tuples) and numcodecs/_codecs.py:21 is inside a docstring — both excluded.
• M5 — os.path.exists/os.path.getsize in the otherwise pathlib-first _local.py (src/zarr/storage/_local.py:366,372). Use Path.exists()/Path.stat().st_size and drop import os.
• M6 — Enable ruff PT011 (pytest.raises without match=) (pyproject.toml:348, already a TODO; 19 test files affected). Medium effort, tests only.
• M7 — Un-ignore ruff TC006 (quoted cast() literals, 93 auto-fixable occurrences). Cosmetic; large mechanical diff — judgement call.
• M8 — ceildiv zero-guard is ambiguous (src/zarr/core/common.py:88). if a == 0: return 0 only matters for ceildiv(0, 0); clarify intent or tighten to a divisor check.

Reviewed, no action proposed

• @dataclass(slots=True) sweep: blocked by cached_property usage on codec classes and downstream-subclassing compat on public classes.
• Removing from __future__ import annotations (60+ files): needs a per-file runtime-introspection audit; not worth the risk now.
• typing_extensions.ReadOnly and generic TypedDict: correctly pinned to typing_extensions until Python 3.13 is the minimum / extra_items lands in stdlib.
• Refuted as bugs after investigation: boundary-chunk is_complete_chunk semantics, chunk-grid boundary math, v2/v3 NaN fill-value round-trips, sync.py double-checked locking, LocalStore atomic writes, sharding index offset correction, BytesCodec endianness, MemoryStore.set_if_not_exists atomicity.

🤖 Generated with Claude Code

Round 2 (deeper sweep: dtype subsystem, indexing, group internals, fsspec/obstore, sync API)

Draft PRs (round 2)

Not actioned: S13/S14/S21 (catalogued, medium-risk design passes).

🐛 Bugs (round 2)

• B9 — Nested structured dtypes fail v2 JSON round-trip (src/zarr/core/dtype/common.py:63-84, high). check_structured_dtype_v2_inner recurses into itself for a nested field whose last element is a list of field pairs, so match_json raises ValueError: No Zarr data type found for data that Struct.to_json(zarr_format=2) itself just wrote. v3 is fine. Fix: recurse with the list-of-fields checker when the inner element is a non-string sequence.
• B10 — Structured dtypes lose field offsets / padding / alignment on round-trip (src/zarr/core/dtype/npy/structured.py:186-216, high). from_native_dtype keeps only (name, dtype) pairs; to_native_dtype re-packs contiguously, so np.dtype([...], align=True) (itemsize 16) comes back as itemsize 9 and stored chunk bytes are misinterpreted. Fix: preserve offsets/itemsize, or loudly reject non-packed layouts.
• B11 — _iter_members_deep drops use_consolidated_for_children on recursion (src/zarr/core/group.py:3343, high). Only immediate children get fresh metadata; grandchildren keep stale consolidated blobs, which consolidate_metadata then bakes into the new consolidation. Empirically confirmed. Implementation found two further holes in the same flag's plumbing: AsyncGroup._members short-circuited to consolidated metadata even when the flag was False, and sync Group.members never forwarded the flag at all — all three fixed in Fix consolidated-metadata staleness in deep iteration, require_array dtype check, attribute-update semantics, and sync timeout cancellation #190.
• B12 — Group.update_attributes_async overwrites attributes while every other update path merges (src/zarr/core/group.py:1965 vs :1219, medium). Two "update" entry points on the same object disagree about whether existing keys survive. Fix: align on merge semantics.
• B13 — require_array with default dtype=None spuriously rejects existing non-float64 arrays (src/zarr/core/group.py:1195-1201, medium). np.dtype(None) → float64, then np.can_cast(complex128, float64) fails. Fix: skip the dtype check when the caller didn't specify one.
• B14 — Deprecation warning points users at a removed config key (src/zarr/storage/__init__.py:35-37, low). The default_compressor warning recommends array.v2_default_compressor.numeric, which is itself removed/deprecated. Fix: update the message.
• B15 — sync() timeout leaves the coroutine running (src/zarr/core/sync.py:151-153, low). On TimeoutError the future is never cancelled, so I/O continues after the caller believes it aborted. Fix: cancel the future on the timeout branch.
• B16 — CPU/GPU NDBuffer.create(fill_value=None) divergence (src/zarr/core/buffer/cpu.py:157 vs gpu.py:179, low/latent). CPU zero-fills on None; GPU returns uninitialized memory. Unreachable on current pipelines but a parity trap. Fix: make the backends agree.
• B17 — UInt64.from_native_dtype is defined twice in the same class (src/zarr/core/dtype/npy/int.py:1100 and :1514, latent). Second definition silently shadows the first. Will be eliminated by the S11 refactor.

🚀 Performance (round 2)

• P7 — _NumcodecsArrayBytesCodec._decode_single forces a full-chunk to_bytes() copy per chunk — refuted during implementation: numcodecs PCodec.decode/ZFPY.decode call ensure_bytes(), which passes bytes through unchanged but copies ndarray input — so to_bytes() is already the optimal form, and the proposed change measured slower (1.07µs vs 0.75µs per 16KB chunk). No action.
• P8 — BytesCodec._decode_sync rebuilds the native dtype via dataclasses.replace(...) per chunk (src/zarr/codecs/bytes.py:89-93, medium). ~1µs/chunk, identical result for every chunk of a read. Fix: cache keyed on (dtype, endian).
• P9 — zarr.open() with zarr_format=None probes array metadata, then re-probes the group on the same path (src/zarr/api/asynchronous.py:360-374, medium on high-latency stores). Opening a group via open() fetches zarr.json/.zattrs twice. Fix: single combined probe dispatching to array vs group; needs care to preserve v2/v3 precedence and warning semantics.
• P10 — BasicIndexer.__iter__ builds 4 tuples per chunk in separate passes (src/zarr/core/indexing.py:611-619, low-medium; ~1.75µs/chunk, ~70ms over 40k chunks). Folded into the S20 dedup (single-pass shared helper).

🧹 Simplifications (round 2)

• S11 — Lift int dtype boilerplate into BaseInt (src/zarr/core/dtype/npy/int.py:240-1551, ~450-550 lines). BaseFloat/BaseComplex already centralize from_native_dtype/to_native_dtype/_from_json_v2/_from_json_v3/to_json; the 8 int classes copy-paste all five. Real variation: item_size, HasEndianness (absent on Int8/UInt8), 2 Windows _check_native_dtype overrides. Also fixes B17 and a eeeeeeeeeeeeeeeee docstring typo (~line 1424).
• S12 — Group.create's 120-line docstring is a verbatim copy of Group.create_array's (src/zarr/core/group.py:2380-2522). Replace with a one-line alias docstring (~115 lines, docs-only).
• S13 — Trivial async _decode_single/_encode_single wrappers liftable into BaseCodec (7 codecs delegate identically, 3 differ only by asyncio.to_thread; ~130-160 lines). Catalogued — medium risk on the hot async path; not actioned this round.
• S14 — Duplicated from_dict across 10 codecs (~40-50 lines + 10 type: ignores). Catalogued — needs a _codec_name ClassVar design pass; not actioned this round.
• S15 — Three byte-identical _parse_*_codec functions (src/zarr/registry.py:197-254, ~40 lines). Collapse to one generic helper.
• S16 — Seven copies of the value-must-be-Buffer check across stores (_memory.py, _local.py, _fsspec.py, _zip.py, ~25 lines). One Store._check_value helper on the ABC; keep message text identical.
• S17 — _get_node_v2/_get_node_v3 + _read_group_metadata_v2/_v3 needless per-format splits (src/zarr/core/group.py:3402-3551, ~45 lines). Fold each pair into its dispatcher.
• S18 — vindex coordinate/mask dispatch + error message copy-pasted 4× (src/zarr/core/indexing.py:1337-1608, ~30 lines). Shared classifier helper + message constant.
• S19 — Dead code: _with_semaphore, collect_aiterator, scalar_failed_type_check_msg (sync.py:192-226, dtype/wrapper.py:280-292, ~35 lines). Grep-verified zero callers in src and tests.
• S20 — Identical ChunkProjection.__iter__ bodies in Basic/Block indexers (src/zarr/core/indexing.py:611-619, :1119-1127). Extract one shared single-pass helper — implementing P10 at the same time.
• S21 — Duplicated JSON-based __eq__/__hash__ in v2/v3 metadata (~20 lines). Catalogued — touches equality on a shared dataclass base; not actioned this round.
• S22 — Double-parse: _parse_deprecated_compressor applied in both Group.create_array and AsyncGroup.create_array (group.py:2646 + :1131). Drop the sync-side pre-parse.

✨ Modernizations (round 2)

• M9 — asyncio.ensure_future → asyncio.create_task (src/zarr/core/common.py:110). Soft-deprecated since 3.10; drop-in here.
• M10 — Enable ruff ASYNC rules; fix 2 real blocking calls in async methods (pyproject.toml + src/zarr/storage/_local.py:366,372). LocalStore.move/getsize call os.path.exists/os.path.getsize without asyncio.to_thread, unlike every other I/O in the file. (Supersedes the pathlib-only change at the same lines in the round-1 modernization PR.)
• M11 — Un-ignore ruff TC006 — corrects round-1 M7: only 12 violations (in _cli/cli.py, _wrapper.py, testing/stateful.py, testing/utils.py), all auto-fixable, not 93. TC006 wants casts quoted, the opposite of what round 1 assumed.
• M12 — Apply ruff PT011 (add match= to bare pytest.raises(ValueError)) — corrects round-1 M6: 15 violations in 2 files (tests/test_store/test_core.py, tests/test_store/test_zip.py), not 19 files.

Reviewed, no action proposed (round 2)

• Indexing deep-dive came back clean: orthogonal/coordinate/mask/block read+write paths, duplicate/unsorted indices, boundary chunks, resize/append, all_equal NaN/signed-zero handling (the -0.0 != 0.0 bit-pattern branch is intentional, PR Write chunks with negative zero values and a zero fill value zarr-developers/zarr-python#3216) — all verified against numpy empirically.
• Byte-range semantics verified consistent across FsspecStore and ObjectStore (exclusive ends match zarr's RangeByteRequest); obstore delete_dir segment matching correct.
• sync/async API twins: no default drift, no missing return wrapping (verified via inspect).
• registry.get_codec_class lazy-load ordering, _NumcodecsCodec thread offloading (already offloaded), get_array_metadata concurrent v2 probes, metadata __eq__ cost (cold path), dtype registry O(n) matching (per-open, not per-chunk) — all fine.
• Float/datetime/complex/vlen fill-value round-trips all correct, including NaN payloads (v3), signed zeros, subnormals, NaT sentinels, embedded null bytes.
• ruff FURB/PERF: already enabled and clean; GitHub Actions: current majors, SHA-pinned; numpy scalar aliases: none in use.