Comment 9 for bug 1735407

So, I proposed a fix that adds a late server group policy check to the evacuation code path. However that fix only decrease the window of the race condition between two parallel evacuations.

So the fix [1] eliminates the race between two evacuates if one of the rebuild request reaches the target compute significantly later than the other rebuild request. Therefore the other rebuild request can finish in the driver before the late check introduced in [1] is run for the first request.

In the other hand if the second rebuild request reaches the compute manager while the first rebuild request is handled by the virt driver then the race can still occur as the late check is added before the virt driver rebuild call.

[1] https://review.openstack.org/525242