Comment 13 for bug 1587285

(Reprint from gerrit)

I have rethought how to resolve this problem after submitting patch-set 9.
But, I couldn't bethink better ideas.
I think that the current idea is the only feasible one.

The other ideas below have big problems.
--------------
(1)Delete garbage BDM records when new BDM record is created.

 pros:
* This idea is very simple.
* This idea guards BDM table from corruption.

 cons:
* There will be several corrected BDM records if we specify a multi-attached volume.[1] In this case, distinguishing gaarbage records from in-progress records is impossible.

[1]"multi-attach-volume" bp (https://blueprints.launchpad.net/nova/+spec/multi-attach-volume) is in-progress.

--------------
(2)Cinder decide to process or not.
Cinder decide to process or not by self when it receives terminate_connection API/detach API.

I think that we can use volume_attachment records for above decision.
But, volume_attachment record is currently created at detach API process. This is too late.
volume_attachment record should be created at first of initialize_connection API process for above decision.

And, we should add "instance_uuid" parameter to initialize_connection API.
Otherwise we couldn't find out the volume_attachment record created at initialize_connection API in detach API process.

pros:
* Cinder already decide to change status or not in detach process. So adding conditional branches to cinder is looks natural to me.

cons:
* Changing initialize-connection API interface doesn't support backward compatibility.

--------------