Looks like a RowBatch wasn't destroyed before the DataStreamRecvr was closed.
Close() can arrive while a TransmitData() RPC is in flight, because of the asynchronous nature of the new RPC flow. And recent changes have made the locking in AddBatch() quite fine-grained. So the mem_tracker might be destroyed during AddBatch(), before the RowBatch is finished with. Looking at the code, it could also be torn down before the RowBatch is created, which would lead to an exception as well.
F0320 18:11:23.188339 4263 mem-tracker.cc:194] Check failed: consumption_->current_value() == 0 (24544 vs. 0) DataStreamRecvr
@ 0x1755bab impala::MemTracker::~MemTracker()
@ 0x1740cbc boost::checked_delete<>()
@ 0x173ef55 boost::scoped_ptr<>::~scoped_ptr()
@ 0x173f43f boost::scoped_ptr<>::reset()
@ 0x17aed5f impala::DataStreamRecvr::Close()
@ 0x1ab0969 impala::ExchangeNode::Close()
@ 0x1e95d65 impala::PlanFragmentExecutor::Close()
@ 0x1e8dc83 impala::FragmentInstanceState::Exec()
@ 0x1e98e65 impala::QueryExecMgr::ExecFInstance()
@ 0x1e9bca2 boost::_mfi::mf1<>::operator()()
@ 0x1e9bb2b boost::_bi::list2<>::operator()<>()
@ 0x1e9b65f boost::_bi::bind_t<>::operator()()