In 1.4, code was added to fetch delegation tokens for Hive metastores and HBase masters. That code is run regardless of whether the user app actually needs those tokens, since there's no way for Spark to know otherwise.
In the best case, that's just unnecessary overhead. In some cases, the Hadoop / Hive / HBase configuration being loaded may contain references to things that Spark doesn't have access to (e.g. Sentry classes), and the Spark application might fail to start.
Adding explicit configuration options to disable those would help, since currently the only way is to set some specific Hive / HBase configuration values, which is kinda brittle.