`TIMESTAMP`, `TIMESTAMP_NTZ`, `TINYINT` and the other `TypeDecorator`s don't declare `python_type`

從定義五個TypeDecorator子類別的src/databricks/sqlalchemy/_types.py開始，閱讀它們的實作和現有型別測試。檢查每種型別的python_type是否與其回傳值相符：TIMESTAMP和TIMESTAMP_NTZ對應datetime，TINYINT對應int，DatabricksTimeType對應time，DatabricksStringType對應str。當這些型別在支援的SQLAlchemy版本下回報預期的Python類別時，即表示完成。

Description of the bug

Our use case

We generate SQLAlchemy ORM models from Databricks tables with sqlacodegen. It reflects the tables through this dialect and annotates each column with the column type's python_type, for example Mapped[str | None] for a STRING column.

For TIMESTAMP, TIMESTAMP_NTZ and TINYINT columns we get Mapped[object | None] instead of Mapped[datetime | None] and Mapped[int | None]. These are the types reflection returns for timestamp, timestamp_ntz and tinyint. Type checkers then accept any value for these attributes, and any code reading them sees object.

The same happens to any other tool that reads python_type, and to the two types the dialect uses in place of Time and String through its colspecs (DatabricksTimeType and DatabricksStringType).

The bug

These five types are sqlalchemy.types.TypeDecorator subclasses that don't declare python_type:

TypeDecorator doesn't take python_type from its impl, so these types get the TypeEngine default: object in SQLAlchemy 2.1, and a NotImplementedError in 2.0.

The SQLAlchemy maintainers consider this the intended behaviour. A TypeDecorator can return different values from its impl through process_result_value, so only the decorator knows its Python type, and it's expected to declare python_type itself: link to the SQLAlchemy discussion.

Versions

• databricks-sqlalchemy: 2.0.10 (also on main)
• databricks-sql-connector: 4.6.0
• SQLAlchemy: 2.1.4, and 2.0.54
• Python: 3.12
• OS: Linux

To Reproduce

No connection is needed:

import datetime
from importlib.metadata import version

import sqlalchemy as sa
from databricks.sqlalchemy import TIMESTAMP, TIMESTAMP_NTZ, TINYINT, DatabricksDialect

print("SQLAlchemy", sa.__version__)
print("databricks-sqlalchemy", version("databricks-sqlalchemy"))

dialect = DatabricksDialect()
expected = {
    # The types reflection returns for timestamp, timestamp_ntz and tinyint.
    "TIMESTAMP": (TIMESTAMP(), datetime.datetime),
    "TIMESTAMP_NTZ": (TIMESTAMP_NTZ(), datetime.datetime),
    "TINYINT": (TINYINT(), int),
    # The types the dialect uses in place of Time and String.
    "Time": (sa.Time().dialect_impl(dialect), datetime.time),
    "String": (sa.String().dialect_impl(dialect), str),
}
for name, (type_, python_type) in expected.items():
    print(f"{name:14} {type(type_).__name__:20} python_type: {type_.python_type}")

for name, (type_, python_type) in expected.items():
    assert type_.python_type is python_type, name

Output

With SQLAlchemy 2.1.4:

SQLAlchemy 2.1.4
databricks-sqlalchemy 2.0.10
TIMESTAMP      TIMESTAMP            python_type: <class 'object'>
TIMESTAMP_NTZ  TIMESTAMP_NTZ        python_type: <class 'object'>
TINYINT        TINYINT              python_type: <class 'object'>
Time           DatabricksTimeType   python_type: <class 'object'>
String         DatabricksStringType python_type: <class 'object'>
Traceback (most recent call last):
  ...
AssertionError: TIMESTAMP

With SQLAlchemy 2.0.54, the first python_type raises NotImplementedError.

Expected behaviour

Each type's python_type is the type of the values it returns: datetime for TIMESTAMP and TIMESTAMP_NTZ, int for TINYINT, time for DatabricksTimeType and str for DatabricksStringType.

Possible fix

Declare python_type on each type in src/databricks/sqlalchemy/_types.py, which already imports datetime and time from datetime:

class TIMESTAMP_NTZ(sqlalchemy.types.TypeDecorator):
    ...

    @property
    def python_type(self) -> type:
        return datetime

The same goes for TIMESTAMP (datetime), TINYINT (int), DatabricksTimeType (time) and DatabricksStringType (str).

Additional context

• DatabricksArray, DatabricksMap and DatabricksVariant also report object, because UserDefinedType has the same default. That's right for all three today. VARIANT can hold any value, and ARRAY and MAP have no result_processor, so their values are whatever the connector returns, which depends on its settings:

  databricks-sql-connector 4.6.0	ARRAY	MAP
  default	numpy.ndarray	list of (key, value) tuples
  disable_pandas=True	list	list of (key, value) tuples
  _use_arrow_native_complex_types=False	str (JSON)	str (JSON)

  A result_processor converting the values to list and dict would let these two types declare python_type too, but that changes the values users get, so it's not part of this request.

• Our workaround: when a TypeDecorator reports object, we use its impl's python_type. That's correct for the five types above, but as the SQLAlchemy maintainers point out, it's only a guess in general.

Thank you!