JSON path equality against a non-string value: silently empty on MySQL, raises on PostgreSQL #1564

Table & {"json_attr.field": value} only behaves correctly when value is a string. A Python bool returns the wrong answer on MySQL and raises on PostgreSQL; a Python int raises on PostgreSQL.

Sibling of #1563 — same root (JSON extraction yields text), different side: that one is the :type annotation on projection, this one is the Python value on restriction.

Repro

@schema
class E(dj.Manual):
    definition = """
    name : varchar(32)
    ---
    s : json
    """
E.insert([
    {"name": "a", "s": {"vendor": "Acme",  "ch": 64, "cal": True}},
    {"name": "b", "s": {"vendor": "Other", "ch": 16, "cal": False}},
])

Each row should select ['a']:

The MySQL boolean row is the serious one: no error, no rows, and the natural reading of an empty result is "nothing is calibrated."

Cause

adapter.json_path_expr yields json_value(...) / jsonb_extract_path_text(...), both of which return text. prep_value (condition.py:336) then renders the Python value by its own type, so the comparison becomes text = <non-text>:

• PostgreSQL refuses the comparison outright.
• MySQL coerces for numerics, which is why 64 happens to work, but compares the extracted true against 1 for a bool and matches nothing.

The asymmetry is invisible to a user: the same expression is correct, wrong, or an error depending on the value's Python type and the backend.

Suggested fix

On the JSON-path branch of prep_value, render the comparison value as text — or cast the extraction to the value's type — so that True, 64 and "Acme" all behave the same way on both backends.

Whichever way, a bool must not silently match nothing. If a given comparison cannot be made portable, raising is acceptable; returning the wrong rows is not.

Documentation

This is almost certainly why tutorials/advanced/json-type.ipynb teaches "Filtering on JSON Content — fetch then filter in Python" and lists "Filter in Python" as an inherent property of JSON in its Design Guidelines. Server-side filtering does work, with the value passed as a string; the tutorial's advice reads as a limitation of the type rather than of this behavior. Worth revisiting together.