Right now, in Hadoop processors, there is a re-login period parameter that will define when the Kerberos ticket must be renewed by the processor. This re-login method will not take into account the lifetime of the generated ticket as the Hadoop library does (when requiring a new ticket, one will be actually generated only after 80% of the current ticket lifetime).
With the current situation, in very specific sets of parameters for relogin period, ticket lifetime and processor scheduling, we can be in a situation where we try to execute the code without actually renewing the ticket. The re-login parameter / code should be removed to rely on the one provided by the underlying Hadoop library.
Workaround is to lower the re-login period parameter.