Feature inference on a Parquet FileSource raises KeyError for date, categorical and decimal columns #7040

Expected Behavior

With no schema on the feature view, feast apply infers a Field for each source column with the column's type (docs, feature-view.md § Feature inferencing). The DataFrame inference path already maps the same values: datetime.date → UnixTimestamp, category → String, Decimal → Decimal. A Parquet file written by pandas with those columns should infer the same way, or fail with a message that names the column and the unsupported type.

Current Behavior

feast apply fails with a bare KeyError carrying the Arrow type string, as soon as the Parquet file has one such column, even when that column is not wanted as a feature.

Column type as pandas writes it and pyarrow.parquet.ParquetDataset(path).schema reads it back, then schema-less apply of a feature view on a file holding only driver_id, event_timestamp and that column:

The same file applies when the feature view lists its features explicitly, so that is the workaround. decimal256, time32, time64, duration, large_binary and halffloat columns fail the same way. Same result on 0.67.0 and on current master.

Steps to reproduce

feature_store.yaml:

project: cli_repro
registry: data/registry.db
provider: local
online_store:
  type: sqlite
  path: data/online.db
entity_key_serialization_version: 3

make_data.py:

from datetime import date, datetime, timedelta, timezone
import pandas as pd
now = datetime.now(timezone.utc).replace(microsecond=0)
pd.DataFrame({
    "driver_id": [1, 2],
    "event_timestamp": [now - timedelta(hours=1), now - timedelta(hours=2)],
    "trips": [10, 20],
    "signup_date": [date(2026, 1, 1), date(2026, 1, 2)],
}).to_parquet("data/driver_stats.parquet")

definitions.py:

import os
from datetime import timedelta
from feast import Entity, FeatureView, Field, FileSource
from feast.types import Int64

driver = Entity(name="driver", join_keys=["driver_id"])
source = FileSource(path="data/driver_stats.parquet", timestamp_field="event_timestamp")
schema = [Field(name="trips", dtype=Int64)] if os.environ.get("SCHEMA") == "explicit" else None
driver_stats = FeatureView(name="driver_stats", entities=[driver], ttl=timedelta(days=1), schema=schema, source=source)

mkdir -p data && python make_data.py
feast apply                  # exit 1
SCHEMA=explicit feast apply  # exit 0, feature trips: Int64

Tail of the failing run:

  File ".../feast/inference.py", line 269, in _infer_features_and_entities
    fv.batch_source.source_datatype_to_feast_value_type()(
  File ".../feast/type_map.py", line 1661, in pa_to_feast_value_type
    value_type = type_map[pa_type_as_str]
KeyError: 'date32[day]'

Replacing signup_date with pd.Series(["x", "y"], dtype="category") gives KeyError: 'dictionary<values=string, indices=int8, ordered=0>'; with [Decimal("1.10"), Decimal("2.20")] it gives KeyError: 'decimal128(3, 2)'.

Specifications

• Version: 0.67.0; same on current master
• Platform: macOS arm64, Python 3.11, pandas 2.3.3, pyarrow 26.0.0
• Subsystem: Python SDK, feature inference for FileSource (Parquet) during feast apply

Possible Solution

pa_to_feast_value_type in sdk/python/feast/type_map.py looks the type string up in a fixed dict after the timestamp, map, large_string and struct branches, so any other spelling raises. Mapping date32/date64 to UNIX_TIMESTAMP, decimal128/decimal256 to DECIMAL and dictionary<values=T, ...> to the mapping of T would match what the DataFrame path does; failing that, a ValueError naming the type would at least point at the column. I searched issues and PRs for the symptom and the mechanism and found only #6876, which handles the list<element: T> and large_list spellings and does not touch scalar types. One related question: a polars-written string column arrives as large_string and is registered as Json rather than String; is that intended? Happy to open a PR for the mapping.