Instead of redirecting from the project service to the service and back to the model,
put all destruction code in the service. Also removes a possible source of failure
where run_after_commit may not destroy the project.
This ensures that _all_ caches (including any caches normally only
flushed under certain conditions) are flushed whenever a project is
removed. Because cache keys are based on project namespaces (excluding
IDs) not doing so could result in a newly created project re-using old
caches (if the project was re-created using the same name).
Ths commit does next:
* When we remove project we move repository to path+deleted.git
* Then we schedule removal of path+deleted with sidekiq
* If repository move failed we abort project removal
This should help us with NFS issue when project get removed but
repository stayed. The full explanation of problem is below:
* rm -rf project.git
* rm -rf removes project.git/objects/foo
* NFS server renames foo to foo.nfsXXXX because some NFS client (think
* Unicorn) still has the file open
* rm -rf exits, but project.git/objects/foo.nfsXXX still exists
* Unicorn closes the file, the NFS client closes the file (foo), and the
* NFS server removes foo.nfsXXX
* the directory project.git/objects/ still exists => problem
So now we move repository and even if repository removal failed
Repository directory is moved so no bugs with project removed but
repository directory taken. User still able to create new project with
same name. From administrator perspective you can easily find stalled
repositories by searching `*+deleted.git`
Signed-off-by: Dmitriy Zaporozhets <dmitriy.zaporozhets@gmail.com>
