Unpinning 2026

This documentation provides details of the bugs that were discovered during lockfile unpinning in 2026. For ways to understand how to unpin lockfiles, look at the Unpinning Environment Lockfiles documentation.

Dependencies which resulted in issues are Pandas, pytest and Numba. NumPy is dependent on Numba and varies according to the Numba version.

Pandas

Pandas upgrade is the cause of the majority of failures since Pandas is used in regression data, which is used in almost all of the tests. This script below replicates the root causes.

test_cow_to_numpy_copy passes and all others fail in Pandas 3. All tests pass in Pandas 2.

import os
import tempfile

import pandas as pd
import numpy as np


def test_cow_series_values():
    df = pd.DataFrame({"a": [1.0, 2.0, 3.0]})
    arr = df["a"].values
    arr[0] = 99.0


def test_cow_dataframe_values():
    df = pd.DataFrame({"a": [1.0, 2.0], "b": [3.0, 4.0]})
    arr = df.values
    arr[0, 0] = 99.0


def test_cow_to_numpy():
    df = pd.DataFrame({"a": [1.0, 2.0, 3.0]})
    arr = df["a"].to_numpy()
    arr[0] = 99.0


def test_cow_to_numpy_copy():
    df = pd.DataFrame({"a": [1.0, 2.0, 3.0]})
    arr = df["a"].to_numpy(copy=True)
    arr[0] = 99.0


def test_hdf5_bytes_attr():
    import tables

    with tempfile.TemporaryDirectory() as tmpdir:
        path = os.path.join(tmpdir, "test.h5")
        df = pd.DataFrame({"x": [1.0, 2.0, 3.0]})
        df.to_hdf(path, key="mydata")

        with tables.open_file(path, "a") as f:
            node = f.root.mydata
            node._v_attrs.pandas_type = b"frame"

        df2 = pd.read_hdf(path, key="mydata")

pytest

pytest reordered fixtures which also cause similar issues. pytest-dev/pytest#13774 is a related PR. Below is a simple reproducer script which fails in pytest 9 but passes in pytest 8.

import pytest


@pytest.fixture(params=["a0"])
def fix_a(request):
    return request.param


@pytest.fixture
def fix_b(fix_a):
    return fix_a


@pytest.fixture(params=["c0"])
def fix_c(fix_a, request):
    return request.param


@pytest.fixture
def fix_group(fix_b, fix_c):
    return [fix_b, fix_c]


def test_solve(fix_group, request):
    print(f"Got '{request.node.name}'")
    assert request.node.name == "test_solve[c0-a0]"

Numba

Running TARDIS with tracking enabled produces different tracker dataframes in Numba 0.61.2 and Numba 0.63.1. This is seen in line interactions, originating from line_scatter_event in interaction_event_callers.py.

The root cause is order of operations which changes between LLVM 15 and LLVM 20. Numba 0.61.2 uses llvmlite 0.44.0 (LLVM 15), while Numba 0.63.1 uses llvmlite 0.46.0 (LLVM 20).

This is also a fastmath issue, disabling fastmath makes it go away.

Tracing the tracker dataframe leads to line_scatter_event as the root cause. line_scatter_event calls get_doppler_factor, and the combination of equations produced do not match between LLVM 15 and LLVM 20. The bug only appears in line_scatter_event and not in get_doppler_factor, since get_doppler_factor standalone is a smaller equation and does not require LLVM reorganisation.

For more information, please see LLVM 15 Debugging and LLVM 20 Debugging.