RedactUrlQueryParamsFilter crashes with TypeError: NoneType is not iterable when MLflow's urllib3 filter runs first #946

RedactUrlQueryParamsFilter is attached to the global urllib3.connectionpool logger at import (databricks/sql/__init__.py:51). Its filter() iterates record.args with no None guard:

else:
    record.args = tuple(
        (self.redact(arg) if isinstance(arg, str) else arg)
        for arg in record.args        # TypeError when record.args is None
    )

Normal logging never sets record.args = None (an args-less call yields ()), so this looks safe in isolation. But MLflow attaches its own credential-redaction filter to the same logger (SensitiveQueryParamFilter), and when it redacts a URL it explicitly nulls the args:

if redacted != message:
    record.msg = redacted
    record.args = None

When both libraries are imported (MLflow first, so its filter runs first), any urllib3.connectionpool record carrying a credentialed URL, e.g. logged on a connection retry, is redacted by MLflow, which sets record.args = None, and then RedactUrlQueryParamsFilter crashes on it.

Minimal repro

import logging, mlflow, databricks.sql
logging.getLogger("urllib3.connectionpool").warning(
    "GET %s", "https://bucket.s3/obj?X-Amz-Signature=deadbeef"
)
# TypeError: 'NoneType' object is not iterable

Real-world trigger: MLflow search_traces(..., include_spans=True) against a Databricks tracking store, on a retry of a presigned/credentialed URL during span loading.

Two bugs in the filter:

1. for arg in record.args has no None guard
2. Line 39, record.arg[k] should be record.args[k]: an AttributeError in the dict-args branch.

Both of these would have been caught with a basic type annotation for the record argument.

Version

databricks-sql-connector==4.4.0; unchanged on main.