Hadoop does a bad job closing files and we end up holding too many caps
(Just scribbling this in the tracker for a record.)
We've had several reports on the mailing list of Hadoop clients persistently holding too many caps. This seems to be related to the Hadoop infrastructure doing a bad job of closing files (possibly because it's relying on finalizers instead of explicitly calling close?). We should find a way to work around it in our bindings since Hadoop's job files are pretty safe for us to drop and re-open on-demand.