Having gone through this let me offer some things I've found.

The most prevalent failure I've come across for work items is _somehow_ work items get a parent that is NOT a work item ( specifically, an it was an iteration =:-o ).

You need to enable the ETL driver logging and let the job fail again.  Then go into the driver log and try to find the 'unsuccessful' URL that shows up.
e.g.

https://rtp-rtc8:9443/jazz/rpt/repository/workitem?fields=workitem%2FworkItem%5BprojectArea%2Farchived%3Dfalse%5D%2F%28parent%2FuniqueId%7CprojectArea%2Farchived%7CuniqueId%29&id=_Gt3PUDT6EeScEqaCQ7MUkQ&modifiedsince=2014-09-04T00%3A00%3A00.000-0400&size=100&pos=800

You'd need to take this URL then use something like RESTClient in firefox modify the size/pos values iteratively until the query fails.    Now, what seems to cause the failure is the parent/uniqueId when the work item parent is an iteration.   Remove that from the query and maybe add properties for the workItem/id and resend.

Then go into the project area and clean up.

I found that this miscreant work items seemed to originate from Templates which contained bad references.