Cells: race condition when saving an instance

[Andrew Laski]

When an instance is saved in a parent cell the db update happens there and then the update is sent to a child cell. When the child cells updates its database it then sends that update back up to the parent to be save there again. The propagation of the change back up to the parent can overwrite subsequent changes there causing data to be lost. Updates from a parent->child or child->parent should go one direction only and not propagate back to the originating cell.

[Andrew Laski (alaski)]

We have seen this cause issues with the Tempest resetState test, and the ec2_idempotent test.

[Andrew Laski (alaski)]

Partially addressed by https://review.openstack.org/#/c/177356/ which keeps updates from going back up to the parent if they originate at the parent. Just waiting on a block for the other direction.

[Andrew Laski (alaski)]

https://review.openstack.org/#/c/176518/ is the other half.

[Andrew Laski (alaski)]

The update was missed when the fix merged.