[Zhuoxi2000]

Description

put_dataproto / get_dataproto changed integer list fields of non_tensor_batch to float64 if any row was an empty list. Rows [[1, 2], [], [2**53 + 1]] came back as [[1.0, 2.0], [], [9007199254740992.0]] with codec dtype float64. Both the element type and values above 2**53 changed. Without the empty row, the same field round-trips as int64. None rows are already excluded from inference.

Cause: np.asarray([]).dtype is float64. _choose_leaf_codec, in its list/tuple branch, and _encode_typed_ragged_values, when no dtype_hint is given, both pass that dtype to np.result_type along with the real rows.

Fix: a small helper, _typed_ragged_row_dtypes, leaves empty list/tuple rows out of promotion. It uses all rows only when every row is empty, so the all-empty case behaves as before. Empty np.ndarray rows still count, because their dtype is explicit. Both inference sites use the helper.

Behaviour after the change (main shown in brackets where it differs):

rows	result	dtype
[[1, 2], [], [2**53 + 1]]	[[1, 2], [], [9007199254740993]]	int64 (was float64, value changed)
[(1, 2), ()]	[[1, 2], []]	int64 (was float64)
[[True, False], []]	[[True, False], []] via msgpack_ragged, same as without the [] row	(was typed_ragged float64, [[1.0, 0.0], []])
[[1.5], []]	unchanged	float64
[np.array([], int32), np.array([1], int32)]	unchanged	int32
[[], []]	unchanged	unchanged

Overlap: rebased onto the relocated module from #3633 (python/mooncake/structured_object_store.py, tests in python/tests/store/). The Rust store copy added in #4254 (python/mooncake/store/rs/structured_object_store.py) has no typed_ragged inference, so nothing to port there. #4603 (integer map keys in msgpack_ragged decode, still open) touches a different function but adds its test near the same spot in test_structured_object_store.py, so the second of the two may have a trivial test-file conflict. I found no issue or PR that reports this bug.

Module

• Transfer Engine (mooncake-transfer-engine)
• Mooncake Store (mooncake-store)
• Mooncake Conductor (mooncake-conductor)
• Reshard (mooncake-reshard)
• Mooncake EP (mooncake-ep)
• Mooncake PG (mooncake-pg)
• Integration (mooncake-integration)
• P2P Store (mooncake-p2p-store)
• Python Wheel (mooncake-wheel)
• Common (mooncake-common)
• Mooncake RL (mooncake-rl)
• CI/CD
• Docs
• Other

Type of Change

• Bug fix
• New feature
• Refactor
• Breaking change
• Documentation update
• Performance improvement
• Other

How Has This Been Tested?

Tests in python/tests/store/test_structured_object_store.py:

• test_dataproto_helper_int_list_rows_with_empty_row_keep_int_dtype: [[1, 2], [], [2**53 + 1]] round trip keeps int values and int64 dtype.
• test_choose_leaf_codec_ignores_empty_list_rows_for_dtype: _choose_leaf_codec picks typed_ragged with dtype int64 for int list/tuple rows with an empty row, and msgpack_ragged for [[True, False], []], the same as for [[True, False]].
• test_dataproto_helper_bool_list_rows_with_empty_row_keep_bool_values: [[True, False], []] round trip uses msgpack_ragged and keeps bool values.

Test commands:

PYTHONPATH=python:mooncake-wheel pytest -q python/tests/store/test_structured_object_store.py \
  -k 'empty_row or ignores_empty_list_rows'
PYTHONPATH=python:mooncake-wheel pytest -q python/tests/store/test_structured_object_store.py \
  mooncake-wheel/tests/test_dataproto_catalog.py -k 'not with_real_store'

Test results:
Run on top of main at 19e0ab2 (Python 3.12, numpy 2.5.4, msgpack 1.2.3, pytest 9.1.1). On macOS the native mooncake._store extension is not built, so I ran with a small local sitecustomize stub that only provides ReplicateConfig for the import; the tests above use InMemoryStore, and the *_with_real_store tests are deselected.

• Main plus the new tests: all 3 new tests fail ([9007199254740992.0] != [9007199254740993], 'float64' == 'int64', 'typed_ragged' == 'msgpack_ragged'); the two files give 3 failed, 150 passed, 27 skipped.
• Only the _choose_leaf_codec change reverted: all 3 new tests fail. The codec test fails on 'float64' == 'int64'. The round trips fail too, because the decision's float64 dtype overrides the encoder's metadata, so int64 bytes decode as float64 ([5e-324, 1e-323] != [1, 2]) and the bool payload fails to decode (buffer size must be a multiple of element size).
• With the fix: the two files give 153 passed, 27 skipped (main: 150 passed, 27 skipped). The skips are the same tests that skip on main.
• Unit tests pass
• Integration tests pass (if applicable)
• Manual testing done: round-tripped the inputs in the table above before and after the change

Checklist

• I have performed a self-review of my own code
• I have formatted my code using ./scripts/code_format.sh (no C/C++ changes)
• I have run pre-commit on the files changed in this PR: ruff check 0.6.9, scripts/ci/check_changed_python_format.py origin/main, ruff format --check and codespell 2.2.6 all pass on both files.
• I have updated the documentation (if applicable)
• I have added tests to prove my changes are effective
• For changes >500 LOC: I have filed an RFC issue

AI Assistance Disclosure

• No AI tools were used
• AI tools were used (specify below)

AI assistance: this change was drafted with an AI coding assistant (Claude) and verified locally with the tests above. I reviewed every changed line and can answer for the change.

[zxpdemonio]

Overall, the implementation looks sound: filtering empty Python list/tuple rows while retaining empty ndarray dtypes and the all-empty fallback keeps codec selection and encoding consistent. I also verified the integer-precision case and the surrounding edge cases against current main.

Suggestion (non-blocking): Please add a regression case for [[True, False], []] (or directly assert that _choose_leaf_codec returns msgpack_ragged). The existing integer test would still pass if the _choose_leaf_codec change were removed, because _encode_typed_ragged_values independently recomputes the dtype. As a result, the current test does not protect the codec-selection branch changed by this PR.

The PR currently conflicts with main after the file relocation, so it will also need a rebase and a fresh CI run.

[Zhuoxi2000]

Thanks for the review. I rebased onto current main, so the change now lives in python/mooncake/structured_object_store.py and the tests in python/tests/store/test_structured_object_store.py.

I added two tests for the codec-selection branch:

• test_choose_leaf_codec_ignores_empty_list_rows_for_dtype asserts _choose_leaf_codec returns typed_ragged with dtype int64 for [[1, 2], [], [2**53 + 1]] and [(1, 2), ()], and msgpack_ragged for [[True, False], []].
• test_dataproto_helper_bool_list_rows_with_empty_row_keep_bool_values round-trips [[True, False], []] and checks the codec is msgpack_ragged and the values stay bool.

With only the _choose_leaf_codec change reverted, both new tests fail, and the codec test fails on 'float64' == 'int64'. They pass with the full change. One note: the existing int test also fails in that case, because _encode_structured_leaf merges decision.metadata over the encoder's metadata, so the int64 payload is decoded as float64. The direct codec test now makes that dependency explicit.

[zxpdemonio]

LGTM.