New implementation uses core concurrent job queue, it is more
robust and predictable than the one shipped in Concurrent.
Additionally:
- Trickles through updates during bulk classification
- Reports errors if we fail during a bulk classification
* push concurrency down to 40. 100 feels quite high.