[WIP] fast GitHub match with GetCommit API

[smola]

What?

• Adds faster GitHub matching based on repositories API instead of search.
• The current approach works like this:
  • On FindPeople (when retrieving from DB, not when using raw cache), it gets an arbitrary commit hash for each (repo, author_name, author_email).
  • The repo and commit hash are used to call the GitHub API and retrieve the GitHub username, which is then cached.
  • On ReducePeople, the previous cache is used to get the GitHub email to username mapping, and a subsequent API call is still done to retrieve the name.

Why?

This removes the need for the search API, which has much lower API quota limits. For an organization like pytorch, it can fit the process within 1 hour limit, as opposed to a few hours.

How to improve?

This pull request is a completely dirty approach that does not work properly with the in-disk caches (they should be removed before running) and doesn't respect the separation of responsibilities of each step. However, it can be improve to better fit the process as follows:

• During FindPeople step, save the arbitrary commit along (repo, author_name, author_email).
• During ReducePeople, pass repo and commit to MatchByEmail, so it can do the GetCommit call as needed.

Note that this requires that repository_id from gitbase is a proper GitHub URL and not an arbitrary path. This is the case for source{d} CE imported repo, for example, but may not be the case in other scenarios. So probably MatchByEmail should fallback to the Search API when the given repo does not match the GitHub pattern.

[vmarkovtsev]

So the plan is to scan the distinct commit signatures to pre-populate the cache, thus the search API will not have to be executed at all + remember about https://github.com/src-d/identity-matching/pull/63/files#r322251594 I will start from scratch this week.

Also spotted:

• Remove the HTTP headers from the logs