[ekrichbaum]

Previous pull wasn't added before changes and Hetzner source. Rebased and separated Ovirt and Openstack into separate pulls although common elements exist.

[semx]

Tested against a DevStack all-in-one (stable/2026.2, openstacksdk 4.22.0) syncing into NetBox 4.7.3. Test servers covered: same name in two projects, two networks, floating IP, attached volume, boot-from-volume, and SHUTOFF.

1. No VMs are created when the compute host also runs nova-conductor/scheduler

GET /os-availability-zone/detail lists the host in two zones:

• internal: nova-scheduler, nova-conductor
• nova: nova-compute

add_cluster records every host of every zone (L837-842). get_cluster_for_host then returns the first match (L474-478), which is internal. As a result, the hypervisor device lands in cluster internal, while the VMs go to nova with device set to that hypervisor. NetBox rejects every VM:

POST /api/virtualization/virtual-machines/ Bad Request
{'device': ['The selected device (nbtest-devstack) is not assigned to this cluster (nova).']}

The interface and virtual-disk POSTs fail afterwards ('virtual_machine': ['This field is required.']).

• Result: 0 VMs in NetBox, the IPs are created unassigned, and the process still exits 0.
• With cluster_exclude_filter = internal, all VMs sync.
• I did not test a multi-node cloud, where compute nodes are not members of internal.

2. Servers with the same name in different projects

With all_projects = True, the second server is skipped:

WARNING: Virtual machine 'nbA-dup' for cluster 'nova' already parsed. Make sure to use unique VM names. Skipping

The duplicate check only uses AZ + name (L1057-1066). After stopping the server that won, the next run rewrote the same NetBox VM with the other server's data:

Virtual machine 'nbA-dup' attribute 'comments' changed from '96804584-…' to 'b93ceea0-…'
Virtual machine interface 'vNIC1 (private) (nbA-dup)' attribute 'primary_mac_address' changed from 'FA:16:3E:CA:51:3D' to 'FA:16:3E:D7:0C:C3'

3. Interface names depend on position

Names are vNIC{count} ({network}) with one counter across all networks (L1157-1175). After removing the first network from a server, the next run renamed the remaining interface:

Updating NetBox 'virtual machine interface' object 'vNIC1 (nbA-net2) (nbA-multi)' with data: {'name': 'vNIC1 (nbA-net2)', ...}

Its previous name was vNIC2 (nbA-net2).

4. The hypervisor interface is synthetic and overwrites the existing type

The hypervisor always gets eth0 with type virtual (L971-977). I changed that interface to 1000base-t in NetBox, and the next run reset it:

Interface 'eth0 (nbtest-devstack)' attribute 'type' changed from '1000base-t' to 'virtual'

5. Primary IPv4 is the last fixed IP

Every fixed address replaces the previous candidate (L1192-1197). The server on private (10.0.0.12) and nbA-net2 (10.77.0.154) got primary_ip4 = 10.77.0.154/32.

6. docs/source_openstack.md vs code

• L21-24 documents --build-arg INSTALL_OPENSTACK=true, but the Dockerfile in this PR has no such argument.
• L115 and L181 refer to an "AWS implementation", but there is no AWS source in the repo.
• L165-166 says disk sizes are GiB before 4.1 and MiB from 4.1. The code multiplies GB by 1000 (L1210): a 2 GB volume is stored as 2000.

7. Removed-settings block copied from vmware

sync_tags = True produces:

Setting 'sync_tags' has been removed ... You need to switch to 'host_tag_source', 'vm_tag_source' or 'cluster_tag_source'

None of those options exist in this source.

8. Smaller points

• Hypervisor platform: it is created as QEMU 8002002, i.e. the raw hypervisor_version.
• Base branch: the PR targets main; other changes go to development.
• Conflicts: it conflicts with #599 in README.md, file_output.py, sources/__init__.py, pyproject.toml, settings-example.ini and uv.lock.
• Duplicated method: it carries its own object_synced_by_other_source, while #599 moves the same method into SourceBase.
• Version range: openstacksdk>=4.0,<5 is a range, while requirements.txt pins exact versions.
• Commit message: commit fe0fe8f is titled "Change greeting from 'Hello World' to 'Goodbye World'" but only changes uv.lock.

What worked

• Lockfile: every uv.lock entry added by the PR matches the PyPI sha256.
• Test suite: pytest merged into development: 392 passed, 8 skipped (vcsim tests included).
• Errors: a wrong password or unreachable Keystone exits 1 with pruning disabled.
• Idempotency: a rerun without changes writes nothing.
• Disks and floating IPs: root/volume/boot-from-volume disks are correct, and floating IPs land on the right interface with the fixed IP as primary.
• Deleted servers: a deleted server is orphaned as expected.

[ekrichbaum]

Also used continuously in production with a large amount of vms.

I also detect internal but beyond having the cluster, the vms are detected in the correct one. 7 is probably due to trying to juggle all of these as separate projects while keeping a copy of all of them together for my actual use. I'll add these notes and see what I come up with.

[ekrichbaum]

I'll revamp and test tomorrow. I see really 5 adjustments that happen.

• AZ host mapping needs to prefer nova-compute membership although I don't have that sort of a setup. I'll build an instance for that edge case.
• Dupe VM names to retain the server UUID when power state changes. and a config option.
• Existing interface names retained when a NIC is matched by MAC (this one is interesting, see VMWare discussion post and PR from another user).
• Primary IP selection prefer first fixed address instead of last. This seems just as arbitrary as there isn't a way to truly guage intent, but it's as simple as the other.
• Update the docs and sync_tags guidance. Clean out the AWS branch references.

Once I test out these changes tomorrow I can update the 4 files again. Then I need to get back to the AWS and Digital Ocean sources. I've gotten an internal request for GCP as well but not sure I will go there. It's less than 1% of our 10k+ vms so may not be worth the effort unless there is more demand.