Switch to error rate to determine retries, rather than absolute error count

[nkemnitz]

Related to #591.
For long running jobs with uninterrupted workers, the current absolute number of retries eventually is exhausted and will needlessly take down the entire job.