It seems that they download different sets of data files.
However, the benchmarks/bench_test.go source walks over the contents of .cache/benchdata and blindly loads all files there to use as the datasets. This means that our benchmarks are not consistent, and dependent on whether we ran the integration test, the download_benchdata.bash script, or both.
The benchdata directory is 451MiB. One consideration is to move it out of .cache and perhaps into benchmarks/.benchdata. Doing so will allow Travis-CI to avoid transferring that data every time the integration test is run.