Skip to main content

Fixed an issue where `SSDictCursor` returned a tuple instead of a dict

asyncmy: Fixed a bug where SSDictCursor returned a tuple instead of a dict​

Link: https://github.com/long2ice/asyncmy/pull/164

  • Target: long2ice/asyncmy (default branch dev, HEAD 0000bd0)
  • Type: PR (bug fix + regression testing)
  • Title: fix: return dicts from SSDictCursor instead of tuples

Background​

SceneMakerAI uses asyncmy as the MySQL/MariaDB driver in its FastAPI service and primarily receives results as DictCursor. SSDictCursor, the dictionary version of the unbuffered cursor for bulk queries, is described in the class docstring as “An unbuffered cursor, which returns results as a dictionary,” but in reality, both fetchone and fetchmany / fetchall / async for all return tuples. Since there were no cases in the repository tests using SSCursor / SSDictCursor, this issue went unnoticed.

Cause​

asyncmy/cursors.pyx’s SSCursor.read_next() converts rows to self._conv_row(...). However, _conv_row is declared as the cdef method in both Cursor and SSCursor (both of which are cdef class). Cython dispatches method calls within cdef class directly to the C vtable, and pure Python subclasses cannot override vtable slots. Therefore, in class SSDictCursor(DictCursorMixin, SSCursor), the DictCursorMixin._conv_row method—which comes first in the MRO—is not called; only SSCursor._conv_row, which returns the input as-is, is executed.

The reason DictCursor behaves correctly is that, in the buffered path, the conversion is completed within the Python method DictCursorMixin._do_get_result. For unbuffered cursors, since _rows is None at that point, read_next() is the only conversion point.

This is a regression introduced in the April 2022 commit 8836fc5, where def was changed to cdef relative to git log -S"cdef _conv_row".

Changes​

  • asyncmy/cursors.pyx: Changed Cursor._conv_row and SSCursor._conv_row from cdef to cpdef (2 lines). Since cpdef searches for and calls the Python override when it is a subclass instance, DictCursorMixin._conv_row and the user-defined dict_type are reapplied.
  • tests/test_cursor.py: Added test_ss_dict_cursor (fetchone / fetchmany / fetchall / async for).
  • CHANGELOG.md: Added one line to the 0.2.16 section.

Reproduction Steps​

async with conn.cursor(cursor=SSDictCursor) as cur:
    await cur.execute("SELECT 1 AS a, 'x' AS b UNION ALL SELECT 2, 'y'")
    print(await cur.fetchone())   # Before modification: (1, 'x')   After modification: {'a': 1, 'b': 'x'}

Verification​

  • Ran on a temporary mariadbd instance on 10/11/15 (dedicated datadir, port 33917). Existing containers and volumes were left untouched and cleaned up after the task.
  • Before the fix: New test failed with an AssertionError (a tuple was returned, which did not match the dict).
  • After the fix: All pytest 87 tests passed, 4 skipped. ruff format --check, ruff check, and mypy passed.
  • For stubtest, the 9 existing errors related to contexts.TypeVar in the Python 3.11 environment remained unchanged before and after the fix (CI uses 3.14). No new errors were introduced by this change.
  • I was unable to run tests on a local machine with a MySQL server and a different Python version.