Did you activate parquet as internal storage option? If so you might deactivate it. I am still wary of this feature and the parquet integration into KNIME.
If this does not help. Next idea could be to switch out snappy for GZIP or none or tell the node to store the table in memory.
Since you are experienced unfortunately it might not be easy to track such a strange error in the middle of an extensive loop.
If stability continues to be an issue you could think about strategies to save intermediate results and be able to restore a loop based on that results. That is not super straightforward but might save the day.
On KNIME server a strategy is possible to have a schedule run every hour or so an pick up where a crashed workflow might have stopped (with the option not to run if the previous job is still on).