Fix crossed-bound XCom slices returning wrong rows via negative SQL L…

[bramhanandlingala]

Summary

Fixes the XCom sequence-slice endpoint (get_mapped_xcom_by_slice) returning incorrect rows — or erroring, depending on the database backend — when a requested slice has crossed bounds (i.e. the effective stop comes before the effective start, which Python evaluates as an empty result).

Root cause

The endpoint maps Python slice semantics onto SQLAlchemy's Query.slice(low, high), which compiles to:

OFFSET low
LIMIT (high - low)

When high <= low, this produces a negative LIMIT, which SQLAlchemy does not clamp:

• SQLite treats a negative LIMIT as "no limit" and silently returns the wrong (extra) rows instead of an empty result.
• PostgreSQL / MySQL reject the query outright, since LIMIT cannot be negative.

This affects 4 of the branches in get_mapped_xcom_by_slice — specifically where both start and stop are provided and a COUNT query is needed to resolve a mixed-sign bound (e.g. start=2, stop=-3), since that's where a crossed slice can occur without either individual bound being invalid on its own.

Fix

Added a small helper, _sliced_or_empty(), that checks for crossed bounds before calling .slice() and returns .limit(0) instead:

def _sliced_or_empty(query: Select, low: int, high: int) -> Select:
    if high <= low:
        return query.limit(0)
    return query.slice(low, high)

Replaced the 4 affected query.slice(...) calls with _sliced_or_empty(query, ...). No other logic changes — non-crossed slices behave exactly as before.

Files changed

• airflow-core/src/airflow/api_fastapi/execution_api/routes/xcoms.py
• airflow-core/tests/unit/api_fastapi/execution_api/versions/head/test_xcoms.py

Testing

Added tests covering all 4 previously-affected branches, verified against real Python slice semantics.

Fixes #69204

Gen-AI disclosure: I used a generative AI tool to help identify the root
cause, write tests, and draft the PR description. I reviewed, tested, and
verified all changes locally before submitting.

Was generative AI tooling used to co-author this PR?

• Yes — Claude

Generated-by: Claude following the guidelines

[shahar1]

See #70437 (review)