The description is not entirerly correct. The issue was follow-up errors
  where the first error is not caught - in which case it's often a system
  error with errcode < 1000 (which are mapped by default to 'HY000'). In this
  case the error state is different from what was assumed in the execution
  loop.