I still wonder whether the hunch of @dliappis with the network problem (e.g. network connection terminated) might guide us into the right direction. Rally just issues one request at the start and one request at the end of the benchmark to retrieve the metrics in question. If there are two hours in between, some component might indeed terminate the connection.
Can you please try to reduce the total benchmark duration to just a few minutes? You can run with --challenge="append-no-conflicts-index-only" --track-params="ingest_percentage:10". This will run an indexing workload (without queries) and only ingest 10% of the corpus.
Daniel