We launch 50 microdesc requests, spread out over just three guards?
Our choice of how many microdescs to fetch per directory request is really totally irrational since the network has grown so much since we chose those parameters.
Oct 11 13:32:44.779 [info] launch_descriptor_downloads(): Launching 50 requests for 4574 microdescs, 92 at a time
Seriously, we're making 50 requests and spreading them over 3 guards? I guess it's nice that we can get partial answers, but generally partial answers don't help us.
We should fetch many more at a time if it's going over a begindir channel.