This is something I stumbled on when I was having connection issues with data.astropy.org but not the mirror: when using download_file(), my download times were very long because it tries to connect several times before giving up and switching to the mirror. When it then finally decides to use the mirror, it works, downloads the file, and caches it. But because the file gets cached using the mirror url, if I run download_file() again with the same exact filename, it won't read it from the cache because the first URL it tries is with data.astropy.org, so then the same timeout issues happen again.

This isn't a serious issue, because I suspect the failure mode is pretty rare: this will only happen if connections to data.astropy.org timeout. If the connection is explicitly blocked, it should quickly fall back to the mirror. But it should probably first check the cache for all possible URLs before doing any internet connection.

cc @saimn @pllim