fix registry._setup_database.lock remove issue and rollback the session after UNIQUE constraint failed exception

[harrisonfeng]

It will try to remove non-existing registry._setup_database.lock
when search_backend setting to sqlalchemy

Signed-off-by: Harrison Feng [email protected]

[dmp42]

I don't get it.
Under what condition may that file not exist?

https://github.com/harrisonfeng/docker-registry/blob/master/docker_registry/toolkit.py#L325

[harrisonfeng]

when search_backend sets to sqlalchemy, first start the container, then the container will stop automatically as that error happened. Actually, that error will happen every time. So I cannot run
docker registry container at all. With my fix, I can run it happily.

[dmp42]

Ok. The better fix here would be to use lockfile (http://pythonhosted.org/lockfile/). Would you be willing to try this?
Thanks.

[pires]

@dmp42 why is lockfile a better solution?

[docwhat]

I'm not sure what @dmp42 is thinking, but as a general rule -- if you can let someone else do the tricky bits, the better. And lockfiles are notoriously tricky.

[dmp42]

@docwhat @pires the current implementation is a parody of a proper file locking mechanism, and it fails miserably as this report demonstrates...
So yeah, let's use something that does work :)

[pires]

Got it, thanks ;-)

[pwais]

+1 if this change reduces crashes at least in the interim. Currently docker run -p 5000:5000 registry is broken out-of-the-box (try running it several times).

[hans-d]

+1. First quick fix the broken part (done with this PR), then in a separate PR improve locking