Huge memory leak when using logging with invalid params on py27

This very exotic and invalid input to log function causes the memory usage go all the way up until it's killed by OS (so the worker dies with OOM).

```
import logging
import apache_beam as beam

[0] | beam.Map(lambda x: logging.warning('%203080000332%', 1))

```

Works as expected on python37 (raises TypeError), fails on python27 on macOS, linux and dataflow workers.