Just to be clear - the loading job keeps running when the kernel error messages first start happening at 21:14:09. But after 21:21:31 the loading job makes no more progress. The kernel error messages continued until 22:19:23 when I killed the job, but nothing further got written to the DB in that intervening hour (based on looking at the filesize, which admittedly may not tell the whole story). The job stats were set to output once per 200,000 ops. It's of course possible that it was still making some progress, but dropping from 3203 ops/second to something less than 55/second is pretty drastic, and at that rate it would have taken weeks to complete the load.

As another note, I also tried to use mdbm_pre_split() at the beginning of the job. That churned for 10 minutes at the beginning of the job before adding any records, then started processing at its normal speed, and then quickly degraded into the same state.