Python handlers: Numba loop vectorization disabled by default on ARM warehouses (Pending)¶

This change affects Python UDFs, UDTFs, and stored procedures that use Numba on ARM-based warehouses. x86-based warehouses are unaffected.

Before the change:

Numba loop vectorization is enabled by default. Some floating-point calculations can produce different results across warehouse generations or repeated runs.

After the change:

Numba loop vectorization is disabled by default, removing this source of numerical variation. Affected calculations might return different results than before.

Preparing for the change¶

No code changes are required. Before enabling the bundle in production, test affected workloads and review numerical differences before updating stored reference results.

To re-enable loop vectorization, add the following to your Python handler before importing Numba:

import os
os.environ["NUMBA_LOOP_VECTORIZE"] = "1"

import numba
numba.config.LOOP_VECTORIZE = 1

Re-enabling vectorization can reintroduce the numerical variation described above.

Ref: 2439