As seen in the attached log snippet of impala and llama there are cases where the query is cancelled so impala releases resources yet later still calls into expand CPU resources for llama. Hence we end up returning UNKNOWN_RESERVATION_FOR_EXPANSION and an error is printed in the logs. We don't think this is causing any real problems because the query is already being torn down.
This only happens when a query is cancelled (internally or externally, e.g. if an error occurs) because the fragments are sent a cancellation request asynchronously and then immediately after the coordinator shuts down the query, including releasing the reservation. The coordinator should wait for fragments to finish closing before releasing the reservation.