Which part is unclear? I added [1] where the physical network 
structure is explained a bit.
The major advantage of using provider networks is that the compute 
nodes are themselves physically attached to the external (provider) 
network, so VM traffic doesn't have to be routed across 
control/network nodes. That mitigates a potential bottleneck if for 
example you only have on control node, lots of VMs and all of them 
send their traffic through that one control node. Spreading that load 
out across the compute nodes reduces the load on the control node(s).
So if our control nodes fail the VMs in the provider networks are not 
affected.
Apart from that, if your DHCP agents misbehave I definitely would look 
into that and fix it, independent of provider networks. The logs might 
reveal something, for example if the agents are overloaded. We saw 
that last year in a customer cloud where only one agent was available 
due to power outage on the second control node (third one was in 
maintenance). We also increased "dhcp_agents_per_network" to 2 
(default 1) which seems to have helped there as well. We haven't had 
any issues since then.

[1] 
https://docs.openstack.org/install-guide/launch-instance-networks-provider.html

Zitat von Nguyễn Hữu Khôi <nguyenhuukhoinw@gmail.com>:

> Hello Eugen,
>
> I got what you mean, but "So as long as the compute nodes
> and the network infrastructure are running we can work, the less
> important VMs are not reachable during that time, of course, but it
> doesn't really hurt anybody."
>
> Can you explain this to me?
>
> Nguyen Huu Khoi
>
>
> On Sun, Aug 13, 2023 at 1:13 AM Eugen Block <eblock@nde.ag> wrote:
>
>> That's depending on your use-case. We have two types of VMs in our
>> cloud, production VMs and the rest (self-service tenant networks). All
>> production VMs are in provider networks, launched with config-drive
>> and independent of the control plane. So as long as the compute nodes
>> and the network infrastructure are running we can work, the less
>> important VMs are not reachable during that time, of course, but it
>> doesn't really hurt anybody.
>> So if you have machines that are required to work independent of
>> network nodes (control nodes) you should consider using such provider
>> networks. Or move the VMs to those networks if you already have been
>> using provider networks.
>>
>> Zitat von Nguyễn Hữu Khôi <nguyenhuukhoinw@gmail.com>:
>>
>> > Yes. I agree with you but when all controller is down and some users need
>> > to reboot their instances then it won't have ip. This is my worry.
>> >
>> > Nguyen Huu Khoi
>> >
>> >
>> > On Fri, Aug 11, 2023 at 5:00 AM Clark Boylan <cboylan@sapwetik.org>
>> wrote:
>> >
>> >> On Thu, Aug 10, 2023, at 12:00 AM, Nguyễn Hữu Khôi wrote:
>> >> > Hello guys.
>> >> >
>> >> > Sometimes I see that my instances won't get IP from Neutron DHCP
>> Agent.
>> >> >
>> >> > Is it good if we set a fixed port for instance and set a static ip in
>> >> > the instance's netplan configuration also?
>> >>
>> >> We do this for IPv6 addresses on some of our servers as we've had
>> problems
>> >> with RAs. It seems to work fine in that case. Ideally, this shouldn't be
>> >> necessary for any cloud instance though and it might be worth digging
>> into
>> >> why dhcp isn't working so that it can be corrected.
>> >>
>> >> >
>> >> >
>> >> > Nguyen Huu Khoi
>> >>
>>
>>
>>
>>
>>