Details
Major Description
We've repeatedly observed the jetty acceptor thread silently dying in the DNs. The webhdfs servlet may also "disappear" and jetty returns non-json 404s.
One approach to make webhdfs more resilient to bad DNs is dfsclient-like fetching of block locations to directly access the DNs instead of relying on a NN redirect that may repeatedly send the client to the same faulty DN(s).