Pri-2 in the absence of a priority. 

My CL at https://codereview.chromium.org/2054763004/ has a tool for generating random directory structures with a given "profile". IE A large number of small files, a mixture of large / small files, a small number of large files. I have been using to test the performance of the directory walking code in the same CL.

These tests take on the order of tens of minutes, so are not really suitable to be run as part of things like the "git cl upload" like most of the other tests. They probably make sense to run on some type of performance bots.

Assigning to tansell@ as he already started creating the framework. We'd need one for 100k and an option to have one for 1m. For the CI, we do not need one that slow, likely one at 10k with mixed load would be sufficient; as the desire is to catch major regressions.

The python code is already tested with local_smoke_test.py in //appengine/swarming. The problem is with the Go code.