Instance quota exceeded with many self.driver.spawn errors

Hello, I want to ask a question about this issue:

I added "raise Exception" before self.driver.spawn() to find this error, like follows:

raise Exception
self.driver.spawn(context, instance, image_meta,
                              injected_files, admin_password,
                              self._legacy_nw_info(network_info),
                              block_device_info)

then, I booted a new instance, and got the following result:

$nova list
+--------------------------------------+-------+--------+----------+
| ID | Name | Status | Networks |
+--------------------------------------+-------+--------+----------+
| 3da5acb5-4a45-46e6-a80a-519bf43a050a | test1 | ERROR | |
+--------------------------------------+-------+--------+----------+

$ mysql> select resource,in_use from quota_usages;
+-----------+--------+
| resource | in_use |
+-----------+--------+
| instances | 1 |
| ram | 512 |
| cores | 1 |
+-----------+--------+

then, I deleted this instance:
$ nova delete 3da5acb5-4a45-46e6-a80a-519bf43a050a

and the quota_usages is reset to 0 like normally:
$ mysql> select resource,in_use from quota_usages;
+-----------+--------+
| resource | in_use |
+-----------+--------+
| instances | 0 |
| ram | 0 |
| cores | 0 |
+-----------+--------+

What I want to ask is that it is a bug ? Even if we got an error instance, but it is still an instance. After you delete it, the resources belonging to it will be released, and quota_usages table will be normal too.

Am I right ?

If the instance isn't running on the hypervisor when it is in the error state then I feel it shouldn't count towards the quota.

As a user, I think error instance should not count quota. Because in most case the error instances are not caused by me. It's the system's fault, I just want a health instance to play. Couting the quota means I used the resource in the meantime I should pay for it. But error instance is useless to me that I don't need to pay it.

OK, thanks, I got it, I will try to fix it.

Error state instances should be counted against the quotas, we don't 100% know the state of what happened so we default to safe. Also having some error state count and some not count is even more confusing.