Projects::TreeController#show Gitaly migration

~Conversation: #220 (closed)

Investigate moving git functionality in the Projects::TreeController#show endpoint to Gitaly.

Information

• Method Call Wall Time Charts
• Associated issue in gitlab-ce: Projects::TreeController#show is slow because of multiple things: https://gitlab.com/gitlab-org/gitlab-ce/issues/27158

TODO

• Investigate required changes in gitlab-rails

  • According to graphana's WALL time chart, instantiating the repository is the slowest Rugged operation in this action. This is done for all project actions in a before_action :repository on app/controllers/projects/application_controller.rb. This means that even for non-repository related project actions, such as Edit Project, the git repository is loaded into memory. There seems to be an actual reason for this: just to render the Graphs tab in the UI menu there's a call to @repository.try(:root_ref). Migrate root_ref to Gitaly and remove the instantiation of a Rugged repository as a before_action on Project controllers. This should have a positive performance effect on many project actions across the board. We will have to check change @repository access in these controllers that are expecting the variable to be instantiated to use the attr_accessor repository.
  • With the previous point as a precondition, remove the initialization of the Rugged repository in this action once all operations used in this action are migrated to gitaly.
  • The next slowest Rugged operation is Repository#tag_names. The source of this was not clear to me from the beginning, so I ran a profiling on the action. The relevant part is:

So, the source is lib/extracts_path.rb's method extract_ref, which is the one that in a path like foo/bar/baz/my.file determines that foo/bar is the branch name and baz/my.file is the file path (for example). It does this through Repository#ref_names by iterating through branch_names + tag_names. Repository#tag_names, the biggest offender, spends most of its time running Rugged::TagCollection#each. So we should implement ref_names in Gitaly.

• Implement required changes in gitaly MR to follow
  • Implement GRPC call SmartHTTP#root_ref
  • Implement GRPC call SmartHTTP#ref_names. This probably means we can implement SmartHTTP#tag_names and SmartHTTP#branch_names at the same time.
• Implement required changes in gitlab-rails MR to follow

/cc @andrewn @jacobvosmaer-gitlab