Github user joewitt commented on the issue:
https://github.com/apache/nifi/pull/2360
I think the primary observed scenario was that when using
{code}UserGroupInformation.loginUserFromKeytabAndReturnUGI(){code} the UGI
static member 'loginUser' was not being set. The 'user' object's 'login' was
being set in the created UGI instance (non static). As a result explicit
logins were working but implicit (hadoop initiated ones) appeared not be. We
could get the lockup in that case to happen within 30 mins or so. Once we
switched to using the loginFromKeytab/getCurrentUser the static member variable
of UGI was set and the flow would work for hours then again lockup. Now, keep
in mind here we had created a thread that ran every 60 seconds to do an
explicit relogin and the hadoop client which would automatically do it when
necessary. In tracing through that scenario what it appears was happening is a
race condition on the 'subject' object whereby we'd do a relogin which clears
out the subject (as a logout is called) and in that time we likely conducted a
hadoop operat
ion which triggered hadoop subject login but the context was wiped and so it
ended up in the promptForName. That is the working theory.
The results so far are very promising.
One should be able to pretty easily replicate what NiFi is/was doing
outside of NiFi in so far as interaction with HDFS.
---
