mlflow/mlflow · error · MlflowException

Failed to convert column {name} from type {values.dtype} to

Error message

Failed to convert column {name} from type {values.dtype} to {t}.

What it means

For object-dtype or python-datetime columns expected to be datetime, MLflow attempts values.astype(datetime64[ns], errors='raise'). If any element cannot be parsed into a datetime (e.g. malformed strings, mixed types), ValueError is raised and wrapped in this exception naming the column and both types. This is the column-level conversion path for PySpark date columns converted to object dtype.

Source

Thrown at mlflow/models/utils.py:808

    if t == DataType.datetime and values.dtype.kind == t.to_numpy().kind:
        # NB: datetime values have variable precision denoted by brackets, e.g. datetime64[ns]
        # denotes nanosecond precision. Since MLflow datetime type is precision agnostic, we
        # ignore precision when matching datetime columns.
        try:
            return values.astype(np.dtype("datetime64[ns]"))
        except TypeError as e:
            raise MlflowException(
                "Please ensure that the input data of datetime column only contains timezone-naive "
                f"datetime objects. Error: {e}"
            )

    if t == DataType.datetime and (values.dtype == object or values.dtype == t.to_python()):
        # NB: Pyspark date columns get converted to object when converted to a pandas
        # DataFrame. To respect the original typing, we convert the column to datetime.
        try:
            return values.astype(np.dtype("datetime64[ns]"), errors="raise")
        except ValueError as e:
            raise MlflowException(
                f"Failed to convert column {name} from type {values.dtype} to {t}."
            ) from e

    if t == DataType.boolean and values.dtype == object:
        # Should not convert type otherwise it converts None to boolean False
        return values

    if t == DataType.double and values.dtype == decimal.Decimal:
        # NB: Pyspark Decimal column get converted to decimal.Decimal when converted to pandas
        # DataFrame. In order to support decimal data training from spark data frame, we add this
        # conversion even we might lose the precision.
        try:
            return pd.to_numeric(values, errors="raise")
        except ValueError:
            raise MlflowException(
                f"Failed to convert column {name} from type {values.dtype} to {t}."
            )

View on GitHub (pinned to 6a27f2decc)

Solutions

Pre-convert the column with pd.to_datetime(df[name], errors='coerce') and inspect/handle rows that became NaT
Fix or drop rows with invalid date values before calling predict
Parse strings with an explicit format: pd.to_datetime(df[name], format='%Y-%m-%d')
If the column genuinely isn't datetime, correct the model signature or the incoming data type

Example fix

# before
df["date"] = df["date"].astype(str)  # mixed/unparsable strings
model.predict(df)

// after
df["date"] = pd.to_datetime(df["date"], format="%Y-%m-%d", errors="coerce")
df = df.dropna(subset=["date"])
model.predict(df)

Defensive patterns

Strategy: validation

Validate before calling

import pandas as pd
for col, t in schema.column_types().items():
    if str(t).endswith("datetime"):
        bad = pd.to_datetime(df[col], errors="coerce").isna() & df[col].notna()
        assert not bad.any(), f"unparsable datetime values in {col}: {df.loc[bad, col].head().tolist()}"

Type guard

def is_parsable_datetime(v) -> bool:
    return isinstance(v, (datetime.datetime, datetime.date, str)) and bool(str(v).strip()) and pd.notna(pd.to_datetime(v, errors="coerce"))

Try / catch

try:
    preds = model.predict(df)
except MlflowException as e:
    if "Failed to convert column" in str(e):
        col = str(e).split("column ")[1].split(" ")[0]
        df[col] = pd.to_datetime(df[col], errors="coerce")
        preds = model.predict(df.dropna(subset=[col]))
    else:
        raise

Prevention

Convert date columns with pd.to_datetime(..., errors='coerce') right after loading data
Log and drop rows with invalid dates before scoring
Keep date columns as datetime from source instead of strings
Test your serving path with the model's saved input_example

When it happens

Trigger: Column declared datetime in the signature but data contains unparsable strings like 'not-a-date', mixed str/datetime objects, or None mixed into an object column being converted.

Common situations: Pyspark DataFrames converted to pandas where date columns become object dtype; CSV reads leaving mixed-type date columns; user input forms sending free-text dates.

Related errors

AI-assisted analysis of mlflow/mlflow@6a27f2decc (2026-08-29). Data as JSON: /api/errors/72a66fc7ae644dab. Report an issue: GitHub.